# AI agents that work inside your walls. URL: https://www.getdynamiq.ai/ > Build chat, voice and workflow agents on one governed platform, deployed in your cloud or data center with our engineers. SOC 2, HIPAA, GDPR, CASA Tier 2. ![An office floor lit at night, seen through the window of a dark building, with a team talking at their desks](https://www.getdynamiq.ai/_next/static/immutable/media/night-office.3808w5tf307od.jpg) The unified agentic AI platform for back-office automation and customer support by voice and chat, deployed with our engineers. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) _Illustration with sample data: an agent run in back office, customer support, compliance and AI Coworker._ Trusted by teams at Technology partners - SOC 2 · HIPAA · GDPR · CASA Tier 2 - AWS, Azure and IBM marketplaces - [Open-source SDK · 1.1k stars](https://github.com/dynamiq-ai/dynamiq) Dynamiq gives banks, governments and enterprises **sovereign AI**: agents on **your infrastructure**, with **the models you choose**, under **your rules**. ## Put agents on the work that fills your queues. Agents take the routine work; your people approve what matters. ### Back office Loan operations, reconciliations, payment exceptions and reporting, with approval gates before anything posts. What it moves - Reconciliation cycle time - Exception aging - Approval turnaround [How it works](https://www.getdynamiq.ai/use-cases/back-office) _Illustration of a back-office agent's approval queue: an invoice matched to a purchase order, a refund exception and a month-end reconciliation, each waiting for a person before posting._ ### Customer support Voice agents answer your phone lines and chat agents your website and apps; the rest goes to a person, with the transcript kept in the trace. What it moves - Handle time - First-contact resolution - Escalation rate In production: An Asian neo-bank automated about 85% of support inquiries and went live in 30 days. [How it works](https://www.getdynamiq.ai/use-cases/customer-service) _Illustration of an inbound voice call: the agent reschedules an appointment with the calendar system, with the response time for each turn._ ### Document review Read loan files, claims and contracts, check them against your rules, and flag exceptions for a reviewer. What it moves - Reviewer time per file - Exception rate - Cycle time [How it works](https://www.getdynamiq.ai/use-cases/document-review) _Illustration of a document review agent: fields extracted from a loan file with the page each came from, two flagged as exceptions against credit policy._ ### Compliance Triage alerts, track regulatory change and draft reports, with a record your auditors can follow. What it moves - Alert triage time - False-positive share - Regulatory change lag [How it works](https://www.getdynamiq.ai/use-cases/compliance-financial-crime) _Illustration of an AML alert triage agent: evidence gathered from transactions and the customer profile, a drafted disposition, and a required analyst sign-off._ ### Research and data Ask your documents and databases in plain language, and get cited answers, charts and HTML reports, on demand or on a schedule. What it moves - Time to answer - Answer coverage - Citation rate In production: A B2B accounting platform shipped document search in one week with one of our engineers. [How it works](https://www.getdynamiq.ai/use-cases/research-knowledge) _Illustration of a research assistant answering a policy question with numbered citations to two internal documents, each marked with the permissions that apply._ The platform ## One engine for chat, voice and workflow agents. Built in: inbound voice, an AI coworker with its own computer, real-time evals, and an agent that finds issues in your traces. [Explore the platform](https://www.getdynamiq.ai/product) - _Illustration of a loan-file review agent on the Dynamiq Agent Builder canvas: a PII guardrail, a document agent with credit-policy knowledge, an underwriter approval gate and an update to the loan system, with every step traced._ ### [Agent Builder](https://www.getdynamiq.ai/product/agents) Design agents on a visual canvas or in Python. Both run on the same engine. - _Illustration of AI Coworker completing a task: it queries the data warehouse, runs analysis code in its sandbox and delivers a board summary document._ ### [AI Coworker](https://www.getdynamiq.ai/product/chat) An agent with its own cloud computer, in the browser, Slack, Teams and Telegram. - _Illustration of an inbound voice call: the agent reschedules an appointment with the calendar system, with the response time for each turn._ ### [Voice Agents](https://www.getdynamiq.ai/product/voice-agents) Answer inbound calls, look up your systems and hand off to a person. - _Illustration of agentic memory: durable facts an agent saved about a customer across sessions, one updated in place instead of duplicated, and the agent recalling them in a new conversation._ ### [Agentic memory](https://www.getdynamiq.ai/product/agents#agentic-memory) Agents keep each person's conversations and save durable facts on their own, which people can see and correct. - _Illustration of a research assistant answering a policy question with numbered citations to two internal documents, each marked with the permissions that apply._ ### [Knowledge](https://www.getdynamiq.ai/product/knowledge-rag) Answers grounded in your documents, with each source's permissions enforced. - _Illustration of guardrail events: a social security number detected in a customer message and blocked, and a prompt injection detected in a retrieved web page and quarantined._ ### [Guardrails and approvals](https://www.getdynamiq.ai/product/guardrails) Detect PII and prompt injection, and pause tool steps for a reviewer. - _Illustration of an evaluation run comparing two versions of a support agent on cases drawn from production traces, with scores for answer quality, faithfulness, policy adherence and escalation accuracy._ ### [Evals](https://www.getdynamiq.ai/product/evaluations) Score agents on datasets, on live traffic in real time, and in voice simulations. - _Illustration of a run trace: each step of a loan-file review with its timing, a PII guardrail event, a knowledge lookup and the underwriter's approval._ ### [Observability](https://www.getdynamiq.ai/product/observability) Every run traced node by node, with cost and latency. An agent flags the issues. - _Illustration of an agent's actions across enterprise systems: updating a CRM opportunity, opening a service ticket, searching documents with the user's own permissions and posting an invoice that waits for approval._ ### [Integrations and AI Gateway](https://www.getdynamiq.ai/product/integrations) Actions across 2,100+ app integrations, and 29 model providers with one gateway for the most-used models. ## Business teams on the canvas. Engineers in Python. The visual builder and the open-source SDK run on the same engine, deploy the same way and leave the same audit trail. [Agent Builder](https://www.getdynamiq.ai/product/agents) · [Open-source SDK](https://www.getdynamiq.ai/open-source) · `pip install dynamiq` _Illustration of the same research agent on the Agent Builder canvas: an input, an agent with a model and a web search tool, and an output._ research\_agent.py ``` from dynamiq.connections import Anthropic as AnthropicConnection from dynamiq.connections import Tavily as TavilyConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import Anthropic from dynamiq.nodes.tools.tavily import TavilyTool agent = Agent( name="research-agent", llm=Anthropic( connection=AnthropicConnection(), model="claude-sonnet-5", ), tools=[TavilyTool(connection=TavilyConnection())], role="Research assistant that answers with current sources.", ) question = "What did central banks decide this week?" result = agent.run(input_data={"input": question}) print(result.output["content"]) ``` ![An office building with every floor lit against a deep blue evening sky](https://www.getdynamiq.ai/_next/static/immutable/media/building-blue-hour.33rm7-iwb-ney.jpg) Deployment ## Sovereign AI, running where your data lives. The same platform runs in Dynamiq Cloud, your own cloud account or on-prem, with open models you host. Isolated setups are delivered with our engineers. - Dynamiq Cloud Managed by us. Fastest way to start. - Your cloud AWS, Azure, GCP, IBM Cloud or OpenShift, in your account. - Your data center Any Kubernetes 1.32+ cluster you run. - Isolated Delivered with our engineers for environments with no outside access. [Deployment options](https://www.getdynamiq.ai/deployment/self-hosted) · Buy through AWS, Azure or IBM marketplaces. ## Our engineers ship the first agent with you. They work in your environment until it runs in production, then hand over the code, the evals and the runbook. [Plan a bootcamp](https://www.getdynamiq.ai/deployment) ![A team working on laptops in a glass meeting room late in the evening](https://www.getdynamiq.ai/_next/static/immutable/media/team-glass-room.2fcix80coz_9o.jpg) Forward-deployed engineers work inside your environment, on your data, next to your team. 1. Step 1 ### Bootcamp in your environment A working agent on your data, with success criteria agreed up front. - Success criteria agreed up front - An eval set built from your data - A working agent 2. Step 2 ### First production agent Built, tested and hardened with your team, evals first. - Production deployment - Monitoring and live evals - Support for your security review 3. Step 3 ### Handover Your team owns the agents, the code and the playbook. - Code, prompts and runbooks - Your team trained to extend it - Enterprise support Time to production 30 days · to put an Asian neo-bank's support agent into production 1 week · for one engineer to ship document search at a B2B accounting platform ## Controls your risk team will sign off on. Every run traced and replayable. Approvals on the steps that matter. PII and prompt injection caught before the model. - Traced and replayable Every run is recorded step by step, with inputs, outputs, cost and latency. - Approval gates Sensitive steps wait for a reviewer, who can edit fields before approving. - Guardrails Detect PII and prompt injection before data reaches a model. - SSO and roles Sign in with your identity provider on Enterprise, with projects, roles and retention controls. - Inside your perimeter Self-host in your cloud or data center, with the models you approve. - SOC 2 - HIPAA - GDPR - CASA Tier 2 [Visit the Trust Center](https://www.getdynamiq.ai/security) _Illustration of an execution approval: before the agent updates the loan system, a reviewer sees the exact call, edits the amount, rate and conditions it may change, and approves or rejects with a reason._ ## Built for banks, insurers, governments, telcos and hospitals. Where a wrong answer costs money, trust or a license. ![A marble banking hall with dark columns, a wooden door and three hanging lamps](https://www.getdynamiq.ai/_next/static/immutable/media/finance-hall.2fyz8aey0i1ra.jpg) ### Financial services Customer service, KYC and AML triage, and lending document review. [Explore financial services](https://www.getdynamiq.ai/industries/financial-services) ![A person reviewing and signing a spread of documents on a dark desk](https://www.getdynamiq.ai/_next/static/immutable/media/documents-desk.0qkvpd2c6szac.jpg) ### Insurance Policyholder service by phone and chat, and claims files prepared for adjusters. [Explore insurance](https://www.getdynamiq.ai/industries/insurance) ![The entrance of a civic building at night, its tall columns lit warm](https://www.getdynamiq.ai/_next/static/immutable/media/civic-night.42pwi16szw848.jpg) ### Government and public sector Citizen services and analysis with human oversight, on infrastructure you control. [Explore public sector](https://www.getdynamiq.ai/industries/public-sector) ![Telecommunication towers silhouetted against an amber dusk](https://www.getdynamiq.ai/_next/static/immutable/media/telecom-towers.1_in21eqoe1p6.jpg) ### Telecommunications Voice and chat care at scale, in your customers' languages. [Explore telecommunications](https://www.getdynamiq.ai/industries/telecommunications) ![A timber building facade with warm interior lighting at night](https://www.getdynamiq.ai/_next/static/immutable/media/timber-lights.3bcr5bnkr5ug7.jpg) ### Healthcare Patient intake and document workflows with PII detection and private endpoints. [Explore healthcare](https://www.getdynamiq.ai/industries/healthcare) ## Running in production today. Anonymized deployments you can ask us about, an open-source core you can read and partners you can call. - [An Asian neo-bank · Financial services](https://www.getdynamiq.ai/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) An Asian neo-bank automated about 85% of support inquiries and went live in 30 days. Read the story - [A B2B accounting platform · Financial services](https://www.getdynamiq.ai/case-studies/smart-document-search-for-a-b2b-accounting-platform-built-in-one-week-with-dynamiq) A B2B accounting platform shipped document search in one week with one of our engineers. Read the story [The core SDK is open source under Apache-2.0.](https://github.com/dynamiq-ai/dynamiq) dynamiq-ai/dynamiq · 1.1k stars [Security review kit](https://www.getdynamiq.ai/security) SOC 2 report, questionnaires and architecture ## Size your first agent. Estimate the value, score your readiness and read how teams put agents into production. [Read the blog](https://www.getdynamiq.ai/blog) ### [AI agent ROI calculator](https://www.getdynamiq.ai/resources/ai-agent-roi-calculator) Estimate hours and cost saved for your use case. Start now ### [Readiness assessment](https://www.getdynamiq.ai/resources/ai-agent-readiness-assessment) Score your data, integrations and governance in a few minutes. Start now - ### [How to automate loan file QC with rules your analysts can change](https://www.getdynamiq.ai/post/automate-loan-file-qc-rules-analysts-can-change) Automate loan file QC with rules your analysts can read and change, a calibrated judge for names, a routing table and test cases behind every release. October 5, 2026 · 15 min read - ### [How Dynamiq built a cost-aware legal research workflow with IBM watsonx](https://www.getdynamiq.ai/post/how-dynamiq-built-a-cost-aware-legal-research-workflow-with-ibm-watsonx) How a three-agent legal research workflow on IBM watsonx routes every question by cost: Granite triages, deep research runs only when needed, all traced. April 15, 2026 · 7 min read - ### [Best Dify alternatives in 2026: open-source and paid options](https://www.getdynamiq.ai/post/best-dify-alternatives-for-ai-workflow-automation) Dify is a source-available platform for agentic workflows and RAG. Compare the best Dify alternatives in 2026 on license, deployment, voice agents and evals. September 17, 2025 · 9 min read ## See an agent on your own workflow. Bring a process and its documents. Our engineers will show you how Dynamiq runs it, in your environment or ours. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # One governed platform for chat, voice and workflow agents. URL: https://www.getdynamiq.ai/product > Chat, voice and workflow agents on one governed platform: knowledge, integrations, models, guardrails, approvals, evals and observability, in your cloud or ours. ![A glass facade at night with one office floor lit behind vertical fins](https://www.getdynamiq.ai/_next/static/immutable/media/facade-night.0bwf8-rii8elu.jpg) Build, ground, govern and deploy agents without stitching tools together. Run it in Dynamiq Cloud or inside your own perimeter. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) In short Dynamiq gives enterprise teams one place to build agents (a visual canvas and a Python SDK on the same engine), ground them in company knowledge and 2,100+ app integrations, govern them with guardrails, approvals, evals and traces, and deploy them as APIs, chat widgets, voice lines or scheduled jobs in the cloud or on-prem. ## Agents An AI coworker for every team, voice agents on your phone lines, and a builder for the workflows only you run. - _Illustration of AI Coworker completing a task: it queries the data warehouse, runs analysis code in its sandbox and delivers a board summary document._ ### [AI Coworker](https://www.getdynamiq.ai/product/chat) Hand it a task in plain language. It plans the work, browses, runs code in its own cloud sandbox and uses the apps you connect. - _Illustration of an inbound voice call: the agent reschedules an appointment with the calendar system, with the response time for each turn._ ### [Voice Agents](https://www.getdynamiq.ai/product/voice-agents) Voice Agents answer calls on the phone, the web or your own backend, reason with an LLM, and speak back, with every turn measured. - _Illustration of a loan-file review agent on the Dynamiq Agent Builder canvas: a PII guardrail, a document agent with credit-policy knowledge, an underwriter approval gate and an update to the loan system, with every step traced._ ### [Agent Builder](https://www.getdynamiq.ai/product/agents) Design an agent on a visual canvas or in the open-source Python SDK. Both compile to the same engine, so what you build ships either way. - _Illustration of four Dynamiq environments: production in the customer's AWS account, a line of business in their Azure tenant, an on-premises install set up with Dynamiq's engineers and a Dynamiq Cloud sandbox._ ### [Apps and deployments](https://www.getdynamiq.ai/product/deployments) Deploy a saved workflow version as an App with its own hostname. Call it over HTTP, embed a chat widget, or trigger it on a schedule or an event. ## Knowledge and data Ground agents in your documents, act in the systems you already use, and choose the model behind every step. - _Illustration of a knowledge base ingestion pipeline in four lanes: files converted into documents, split into chunks, embedded into vectors and written to the store._ ### [Knowledge](https://www.getdynamiq.ai/product/knowledge-rag) A knowledge base converts your documents into searchable context: ingestion, chunking, embedding and storage, managed for you or pointed at your own vector store. - _Illustration of an agent's actions across enterprise systems: updating a CRM opportunity, opening a service ticket, searching documents with the user's own permissions and posting an invoice that waits for approval._ ### [Integrations](https://www.getdynamiq.ai/product/integrations) Connect a workflow to 2,100+ app integrations, or point it at your own databases, warehouses and MCP servers. - _Illustration of the AI Gateway routing requests: a primary hosted model, a fallback, a self-hosted open model for traffic that contains personal data, and a regional deployment._ ### [Models and AI Gateway](https://www.getdynamiq.ai/product/models) Use 29 model providers from the SDK, call the most-used models behind one endpoint and one credential, or host and fine-tune open models yourself. ## Govern Catch sensitive data before a model sees it, measure quality on real traffic, and trace every run. - _Illustration of guardrail events: a social security number detected in a customer message and blocked, and a prompt injection detected in a retrieved web page and quarantined._ ### [Guardrails and approvals](https://www.getdynamiq.ai/product/guardrails) Detect PII and prompt injection before a model sees them, enforce content policies, and pause tool steps for a person to approve. - _Illustration of an evaluation run comparing two versions of a support agent on cases drawn from production traces, with scores for answer quality, faithfulness, policy adherence and escalation accuracy._ ### [Evals](https://www.getdynamiq.ai/product/evaluations) Score an agent on a dataset before you ship, then keep scoring a sample of its live traffic after you deploy. - _Illustration of a run trace: each step of a loan-file review with its timing, a PII guardrail event, a knowledge lookup and the underwriter's approval._ ### [Observability](https://www.getdynamiq.ai/product/observability) Every run is traced and replayable, node by node, with the cost, latency and tokens each step used. ## Business teams on the canvas. Engineers in Python. The visual builder and the open-source SDK run on the same engine, deploy the same way and leave the same audit trail. [Agent Builder](https://www.getdynamiq.ai/product/agents) · [Open-source SDK](https://www.getdynamiq.ai/open-source) · `pip install dynamiq` _Illustration of the same research agent on the Agent Builder canvas: an input, an agent with a model and a web search tool, and an output._ research\_agent.py ``` from dynamiq.connections import Anthropic as AnthropicConnection from dynamiq.connections import Tavily as TavilyConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import Anthropic from dynamiq.nodes.tools.tavily import TavilyTool agent = Agent( name="research-agent", llm=Anthropic( connection=AnthropicConnection(), model="claude-sonnet-5", ), tools=[TavilyTool(connection=TavilyConnection())], role="Research assistant that answers with current sources.", ) question = "What did central banks decide this week?" result = agent.run(input_data={"input": question}) print(result.output["content"]) ``` ![An office building with every floor lit against a deep blue evening sky](https://www.getdynamiq.ai/_next/static/immutable/media/building-blue-hour.33rm7-iwb-ney.jpg) Deployment ## Sovereign AI, running where your data lives. The same platform runs in Dynamiq Cloud, your own cloud account or on-prem, with open models you host. Isolated setups are delivered with our engineers. - Dynamiq Cloud Managed by us. Fastest way to start. - Your cloud AWS, Azure, GCP, IBM Cloud or OpenShift, in your account. - Your data center Any Kubernetes 1.32+ cluster you run. - Isolated Delivered with our engineers for environments with no outside access. [Deployment options](https://www.getdynamiq.ai/deployment/self-hosted) · Buy through AWS, Azure or IBM marketplaces. ## Questions and answers ### What is the Dynamiq platform? Dynamiq is an enterprise platform for building, governing and deploying AI agents: chat agents, voice agents and workflow agents, with knowledge retrieval, integrations, guardrails, approvals, evals and observability in one place. ### Can we run Dynamiq in our own cloud or data center? Yes. Use Dynamiq Cloud, or self-host on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster. Fully isolated environments are delivered with our engineers. ### Which models does Dynamiq support? Agents can use any of 29 model providers through the SDK, the AI Gateway puts the most-used models behind one endpoint, and you can run and fine-tune open models you host. You choose the model for each agent. ### Is the platform open source? The core Python SDK is open source under Apache-2.0 and runs on the same engine as the platform, so agents built in code and on the canvas deploy the same way. ### How does Dynamiq handle security and compliance? Dynamiq has SOC 2, HIPAA, GDPR, CASA Tier 2. Every run is traced and replayable, sensitive steps can wait for approval, and guardrails detect PII and prompt injection. Enterprise adds SSO, roles and retention controls. ## See an agent on your own workflow. Bring a process and its documents. Our engineers will show you how Dynamiq runs it, in your environment or ours. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # Start free. Scale with Enterprise. URL: https://www.getdynamiq.ai/pricing > Start building AI agents on Dynamiq for $0. Enterprise adds self-hosting, SSO, forward-deployed engineers, compliance paperwork and support with an SLA. Build your first agent at no cost. Enterprise adds self-hosting, SSO, forward-deployed engineers and support. ## Free $0 to start For individuals and teams building their first agents. - A Dynamiq Cloud workspace - Build agents on the canvas or in Python - Core platform features within usage limits - The open-source SDK, Apache-2.0, free forever Usage limits apply. [Start free](https://app.getdynamiq.ai/signup) ## Enterprise For regulated teams in production Custom Everything in Free, deployed where you need it, with the engineers and paperwork a regulated rollout requires. ### Deployment - Dynamiq Cloud or your own cloud account - Your data center, on any Kubernetes 1.32+ cluster - Isolated environments with our engineers ### Security - SSO, organizations, projects and roles - Retention and deletion controls - SOC 2, HIPAA, GDPR, CASA Tier 2, with BAA and DPA ### Engineers - Forward-deployed engineers - A bootcamp in your environment - Onboarding and handover to your team ### Support - Support with an agreed SLA - Help with your security review - Upgrade guidance for self-hosted installs [Talk to the team](https://www.getdynamiq.ai/book-a-demo) · [How we deploy](https://www.getdynamiq.ai/deployment) Buy through your cloud marketplace. Procure Dynamiq through AWS, Microsoft Azure or IBM marketplaces, with AWS private offers that can count toward your commitment. [Marketplaces](https://www.getdynamiq.ai/partner-catalog) ## Questions and answers ### What counts as usage on the Free plan? Plans set limits on resources, such as apps and knowledge bases, and on monthly activity, such as chat messages, app runs and knowledge base ingestions. You can see where you stand under Settings, Usage in the app. ### Is the open-source SDK really free? Yes. The dynamiq Python SDK is licensed Apache-2.0, including for commercial use. You pay only for the platform features and services you choose. ### Can Dynamiq run fully isolated? Yes, on the Enterprise plan. Isolated environments with no outside access are delivered with our engineers. ### Do engineers come with the Enterprise plan? Enterprise engagements include forward-deployed engineers who run a bootcamp in your environment, build the first production agent with your team and hand it over. ### Can we buy through a cloud marketplace? Yes. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace, so you can buy through a cloud account you already have, and AWS private offers can count toward your AWS spend commitment. ### How is Enterprise priced? Enterprise pricing depends on the deployment model, usage and the engineering support you need. Talk to our team and we will scope it with you. ## See an agent on your own workflow. Bring a process and its documents. Our engineers will show you how Dynamiq runs it, in your environment or ours. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # Built to pass your security review. URL: https://www.getdynamiq.ai/security > SOC 2, HIPAA, GDPR, CASA Tier 2, SSO on Enterprise, roles, encrypted secrets, isolated code execution, traced runs and self-hosting. Request the security review kit. ![The round steel door of a bank vault with its radial locking bolts](https://www.getdynamiq.ai/_next/static/immutable/media/vault-door.0vbrzzay1ghli.jpg) Compliance, identity, data protection and agent governance, in our cloud or inside your perimeter. [Request the security kit](https://www.getdynamiq.ai/contact?topic=security) · [Sub-processors](https://www.getdynamiq.ai/sub-processors) In short Dynamiq holds SOC 2 and CASA Tier 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Credentials are encrypted in a Vault-based secrets manager, API keys are stored only as hashes, every agent run is traced and replayable, and the whole platform can run inside your own cloud account or data center. ## Compliance - ### SOC 2 An independent audit of our security controls. The report is available on request under NDA. - ### HIPAA Built to support HIPAA workloads. We sign a Business Associate Agreement (BAA) with covered entities. - ### GDPR A Data Processing Agreement (DPA) is available, and our sub-processors are listed publicly. - ### CASA Tier 2 Our cloud application passed the App Defense Alliance's Cloud Application Security Assessment at Tier 2, a lab-verified review against the OWASP Application Security Verification Standard. ## Identity and access Single sign-on On Enterprise, sign in through your identity provider over OpenID Connect, such as Okta or Microsoft Entra ID, limited to your verified email domains. Multi-factor authentication Each person can turn on a second factor for signing in to their account. Organizations, projects and roles Every request is checked against project membership, and management actions require an owner or admin role. Guests with project-only access Invite contractors and partners as guests who see only the projects they are added to. Private by default Chat conversations, their files and scheduled tasks are visible only to the person who owns them. ## Data protection Encrypted secrets Credentials for connected systems are encrypted at rest in a secrets manager built on HashiCorp Vault, and keys never leave it. Hashed API keys Access keys and tokens are stored only as SHA-512 hashes, can expire on a date you set, and can be revoked at once. Retention policies Set an organization-wide ceiling with project overrides, and remove prompt and response content from traces immediately or after a set number of hours. Deletion you can rely on Deleting a project cuts access at once and permanently removes its data after a retention window. ## Agent governance Traced and replayable Every run is recorded node by node, with inputs, outputs, cost and latency. Approvals Sensitive steps wait for a reviewer, who can edit fields before approving. Guardrails Detect PII and prompt injection before data reaches a model. Isolated code execution Agent code runs in isolated sandboxes, separated from the platform and from other customers. ## Deployment Inside your perimeter Self-host on AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster. Isolated environments Environments with no outside access are delivered with our engineers. Your models Choose from 29 model providers or run open models inside your own cluster. ## The security review kit Everything your security and procurement teams ask for, in one request. [Request the kit](https://www.getdynamiq.ai/contact?topic=security) - SOC 2 report - Data Processing Agreement (DPA) - Business Associate Agreement (BAA) - Architecture and self-hosting overview - Answers to your security questionnaire ## Policies The documents behind our commitments. - [Privacy policy](https://www.getdynamiq.ai/privacy-policy) - [Terms of service](https://www.getdynamiq.ai/terms-of-service) - [Sub-processors](https://www.getdynamiq.ai/sub-processors) - [Cookie declaration](https://www.getdynamiq.ai/cookie-declaration-dynamiq) - [Platform security documentation](https://docs.getdynamiq.ai/docs/platform/administration/security) ## Questions and answers ### Which compliance standards does Dynamiq meet? Dynamiq holds SOC 2 and CASA Tier 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. The SOC 2 report is available on request under NDA. ### Can we keep all data inside our own environment? Yes. Self-host the platform in your cloud account or data center. Outbound calls go only to the model and tool providers you configure, and isolated setups with our engineers need no outside access. ### How do you handle personal data in prompts? Guardrails detect PII and prompt injection before data reaches a model, so you can block or route those requests. Every run is traced so reviewers can see exactly what happened. ### How do we report a vulnerability? Email hello@getdynamiq.ai with the details, or contact your Dynamiq representative. We work with you on remediation and disclosure. ## Bring your security team to the call. We will walk through the architecture, the controls and the deployment model that fits your requirements. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # Engineers who put agents into production with you. URL: https://www.getdynamiq.ai/deployment > Our engineers run a bootcamp in your environment, ship your first production agent with your team, then hand over the code, prompts and runbooks. ![An engineer at a desk, blurred in motion behind ribbed glass](https://www.getdynamiq.ai/_next/static/immutable/media/engineer-motion.3ghiwm7gna79g.jpg) Our forward-deployed engineers work in your environment, from a bootcamp on your data to a production agent your team owns. [Plan a bootcamp](https://www.getdynamiq.ai/book-a-demo) · [Where Dynamiq runs](https://www.getdynamiq.ai/deployment/self-hosted) In short Dynamiq pairs the platform with forward-deployed engineers. They run a bootcamp in your environment, build the first production agent with your team against success criteria and evals agreed up front, then hand over the agents, code, prompts and runbooks so your team can run and extend them. ## From our engineers to your team, by design. They work in your environment until it runs in production, then hand over the code, the evals and the runbook. ### A working agent on your data, with success criteria agreed up front. Our engineers · Your team Our engineers lead, your team builds alongside them. - Success criteria agreed up front - An eval set built from your data - A working agent ### Built, tested and hardened with your team, evals first. Our engineers · Your team Your team takes over more of the build every week. - Production deployment - Monitoring and live evals - Support for your security review ### Your team owns the agents, the code and the playbook. Our engineers · Your team Your team runs and extends it; we stay on call. - Code, prompts and runbooks - Your team trained to extend it - Enterprise support ## How we work Four rules every engagement follows. ![A team working on laptops in a glass meeting room late in the evening](https://www.getdynamiq.ai/_next/static/immutable/media/team-glass-room.2fcix80coz_9o.jpg) Success criteria first Before we build, we agree what good looks like and turn it into an eval set from your own examples. Inside your environment The work happens where the agent will run: your cloud account, your data center or an isolated network. Your team builds alongside us Your engineers, process owners and security reviewers are part of the build, so nothing is a black box at handover. You own what we build The agents, prompts and code are yours, and the core SDK they run on is open source under Apache-2.0. ## Every deployment makes the platform better Reusable connectors and templates we build join the product. Your agents, prompts, code and data stay yours. 1. ### Build We solve your use case on the platform, with your data and your controls. 2. ### Prove The agent passes the evals agreed up front, in production, with monitoring on. 3. ### Generalize Connectors and components that other teams can reuse go back into the platform, so each deployment makes the next one faster to start. ## Deploy anywhere your data lives. The same platform runs in Dynamiq Cloud, your own cloud account or on-prem, with open models you host. Isolated setups are delivered with our engineers. - Dynamiq Cloud Managed by us. Fastest way to start. - Your cloud AWS, Azure, GCP, IBM Cloud or OpenShift, in your account. - Your data center Any Kubernetes 1.32+ cluster you run. - Isolated Delivered with our engineers for environments with no outside access. - [Where Dynamiq runs · Deployment models, what you provide and what runs in your cluster.](https://www.getdynamiq.ai/deployment/self-hosted) - [Security and Trust Center · Compliance, identity, data protection and the security review kit.](https://www.getdynamiq.ai/security) - [Partners and marketplaces · Cloud partners, and buying through AWS, Azure and IBM marketplaces.](https://www.getdynamiq.ai/partner-catalog) ## Questions and answers ### What is a forward-deployed engineer? An engineer who works inside your environment with your team to put an agent into production, instead of handing you software and leaving. Ours build on the Dynamiq platform and hand the result over to you. ### What do we own at the end? Everything built for you: the agents, prompts, evaluation sets, code and runbooks. Your team is trained to extend them, and the open-source SDK means the core engine is yours to run. ### Can your engineers work in isolated or air-gapped environments? Yes. Fully isolated environments, with no outside access, are delivered with our engineers, who set up the platform and models to run inside that boundary. ### How is success measured? We agree success criteria before building and turn them into an eval set built from your own examples. The agent goes to production when it passes them, and live evals keep measuring it afterwards. ### Can we buy through a cloud marketplace? Yes. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace, so the purchase can go through a cloud account you already have. ## Run a bootcamp in your environment. Tell us the workflow you want to automate. We will scope the bootcamp, the success criteria and the team. [Plan a bootcamp](https://www.getdynamiq.ai/book-a-demo) · [Start free](https://app.getdynamiq.ai/signup) --- # Run Dynamiq where your data has to stay. URL: https://www.getdynamiq.ai/deployment/self-hosted > Run Dynamiq in Dynamiq Cloud, in your AWS, Azure, GCP or IBM Cloud account, on OpenShift or any Kubernetes cluster, or fully isolated with our engineers. Dynamiq Cloud, your own cloud account, or your data center. The same platform in every case. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) In short Dynamiq runs as a managed service or inside your own infrastructure. The self-hosted edition installs from an official Helm chart on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, and fully isolated environments are delivered with our engineers. ## Four ways to run it 1. ### Dynamiq Cloud Teams that want to start today. Managed by us, with SOC 2, HIPAA, GDPR, CASA Tier 2, retention controls, and SSO on Enterprise. Nothing to install. 2. ### Your cloud Teams whose data must stay in their own cloud account. Install the platform with the official Helm chart on AWS, Azure, GCP, IBM Cloud or Red Hat OpenShift. Subscribe through AWS Marketplace if you prefer. 3. ### Your data center Teams that run their own infrastructure. Any Kubernetes 1.32+ cluster on your hardware, backed by your own database, messaging and object storage. 4. ### Isolated Environments with no outside access. Delivered with our engineers, who set up the platform and the models it uses to run entirely inside your boundary. _The Dynamiq platform running in four places: Dynamiq Cloud, your cloud, your data center and isolated environments_ ## What runs inside your perimeter Self-hosting runs the entire platform in your Kubernetes cluster. The whole platform The web app, the management API, the agent runtime and the services that host your deployed agents and knowledge bases. Your data stays put Documents, knowledge bases, traces and credentials live in your database and storage. Your secrets manager Pull credentials from your own store, such as AWS Secrets Manager, through the External Secrets Operator. Egress you control Outbound calls go only to the model and tool providers you configure. Isolated setups with our engineers need none. ## What you provide The chart deploys the Dynamiq services. You bring the backing services, so your data never sits in a system you do not control. Kubernetes A Kubernetes 1.32+ cluster with an ingress or Gateway API controller. PostgreSQL 16+ One server with three logical databases, managed or in-cluster. NATS with JetStream The message bus the platform services share. S3-compatible storage A bucket for platform artifacts: S3, MinIO, IBM Cloud Object Storage and others. DNS and TLS Records and wildcard certificates for the app, the API and deployed agents. A license and registry access Issued by Dynamiq with your agreement. ## Installation guides Every step, from requirements to upgrades and rollback, is in the documentation. - [Self-hosted overview](https://docs.getdynamiq.ai/docs/platform/self-hosted/overview) - [System requirements](https://docs.getdynamiq.ai/docs/platform/self-hosted/requirements) - [Install on Kubernetes with Helm](https://docs.getdynamiq.ai/docs/platform/self-hosted/install-kubernetes) - [Install on AWS (EKS)](https://docs.getdynamiq.ai/docs/platform/self-hosted/install-aws-eks) - [Install on IBM Cloud (IKS)](https://docs.getdynamiq.ai/docs/platform/self-hosted/install-ibm-cloud) - [Install on Red Hat OpenShift](https://docs.getdynamiq.ai/docs/platform/self-hosted/install-openshift) ## Questions and answers ### Is the self-hosted platform the same as Dynamiq Cloud? Yes. It is the same platform, installed from the official Helm chart into your cluster, with the same agents, knowledge, evals and observability. Some capabilities call outside services by default, such as voice infrastructure, some detectors and AI Coworker's web tools, so we confirm the ones you need during setup; fully isolated deployments are delivered with our engineers. ### Does any data leave our environment when we self-host? Your data, knowledge bases and traces stay in your database and storage. Outbound calls go only to the model and tool providers you configure, and isolated setups with our engineers need no outside access at all. ### Which clouds and platforms are supported? AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, and any Kubernetes 1.32+ cluster in your data center. The documentation has step-by-step guides for EKS, IBM Cloud and OpenShift. ### Can we buy the self-hosted edition through AWS Marketplace? Yes. Subscribe through the AWS Marketplace listing and install the chart from the Marketplace registry into your own account. ### Who operates the platform after installation? Your team does, with upgrade and rollback guides in the documentation and support from us. Our engineers can run the first install with you. ## Plan your deployment with our engineers. We will map your environment, your constraints and the fastest route to a running platform. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # The open-source engine behind Dynamiq. URL: https://www.getdynamiq.ai/open-source > dynamiq is an Apache-2.0 Python framework for agents, workflows and RAG, and the same engine that runs the Dynamiq platform. Build in code, deploy anywhere. The dynamiq Python SDK is Apache-2.0. Build agents, workflows and RAG in code, then deploy them to the platform when you need governance. [Star on GitHub · 1.1k](https://github.com/dynamiq-ai/dynamiq) · [Read the docs](https://docs.getdynamiq.ai/docs/sdk) terminal ``` pip install dynamiq ``` In short dynamiq is an open-source Python framework for building AI agents, multi-agent workflows and retrieval pipelines, licensed Apache-2.0. It is the same engine that runs the Dynamiq platform, so code you write locally deploys with traces, evals and guardrails. ## What is in the SDK Agents with tools Give an agent a model, a role and tools such as web search, code execution and your own functions. Graph orchestration Compose multi-step and multi-agent workflows with explicit state and conditional routing. Retrieval and RAG Ingest, split and embed documents with 8 embedding providers and 8 vector stores. Any model Swap between 29 model providers without rewriting your agent. Memory Keep conversation and long-term memory in one of 8 memory backends. Tracing to the platform Send runs to Dynamiq for traces, evals and deployment when you need governance. ## An agent in one Python file A research agent with a model and a web search tool. The same code runs locally, in your cloud or on the platform. research\_agent.py ``` from dynamiq.connections import Anthropic as AnthropicConnection from dynamiq.connections import Tavily as TavilyConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import Anthropic from dynamiq.nodes.tools.tavily import TavilyTool agent = Agent( name="research-agent", llm=Anthropic( connection=AnthropicConnection(), model="claude-sonnet-5", ), tools=[TavilyTool(connection=TavilyConnection())], role="Research assistant that answers with current sources.", ) question = "What did central banks decide this week?" result = agent.run(input_data={"input": question}) print(result.output["content"]) ``` ## Start building Guides, examples and the full reference. - [SDK quickstart](https://docs.getdynamiq.ai/docs/sdk) - [Agents](https://docs.getdynamiq.ai/docs/sdk/agents/agent) - [Model providers](https://docs.getdynamiq.ai/docs/sdk/llms/llm-providers) - [Examples on GitHub](https://github.com/dynamiq-ai/dynamiq/tree/main/examples) - [Release notes](https://github.com/dynamiq-ai/dynamiq/releases) ## Questions and answers ### What license is the dynamiq SDK under? Apache-2.0. You can use, modify and distribute it, including in commercial products. ### How is the SDK related to the Dynamiq platform? The platform runs on the same engine. Agents built in Python and agents built on the visual canvas deploy the same way, with the same traces, evals and guardrails. ### Do I need the platform to use the SDK? No. The SDK runs anywhere Python runs. Connect it to the platform when you want deployment, observability, evals and governance without building them yourself. ### How do I contribute? Open an issue or a pull request on GitHub. The repository has contribution guidelines, examples and release notes. ## Take it to production on the platform. Deploy your code-built agents with traces, evals, guardrails and approvals, in our cloud or yours. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # We put AI agents into production for regulated industries. URL: https://www.getdynamiq.ai/about > Dynamiq builds an enterprise platform for AI agents and deploys it with regulated organizations, in their cloud or data center, with forward-deployed engineers. ![A city skyline in silhouette against an amber dusk](https://www.getdynamiq.ai/_next/static/immutable/media/skyline-dusk.11mpz5w92op3-.jpg) Dynamiq builds an enterprise platform for AI agents and deploys it with the teams who use it, in their cloud or data center. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) In short Dynamiq Technologies, Inc. makes the Dynamiq platform for chat, voice and workflow agents, maintains the open-source dynamiq SDK, and runs a forward-deployed engineering practice that takes customers from a bootcamp in their environment to agents in production. ## What we believe Agents belong inside the perimeter Regulated data should not have to leave the environment it lives in for an agent to use it. Open by license, not by promise The core engine is open source under Apache-2.0, so customers can read, run and keep what they depend on. Production is the benchmark An agent counts when it runs in production against success criteria agreed up front, measured by evals every day. Your team owns the result Our engineers hand over the agents, code and runbooks. Nothing we build for you is a black box. ## What we build - [The Dynamiq platform · Chat, voice and workflow agents on one governed stack.](https://www.getdynamiq.ai/product) - [The open-source SDK · The Python engine behind the platform.](https://www.getdynamiq.ai/open-source) - [Forward-deployed engineering · From a bootcamp in your environment to production.](https://www.getdynamiq.ai/deployment) ## Who we work with - [Financial services · Customer service, KYC and AML triage, and lending document review.](https://www.getdynamiq.ai/industries/financial-services) - [Insurance · Policyholder service by phone and chat, and claims files prepared for adjusters.](https://www.getdynamiq.ai/industries/insurance) - [Government and public sector · Citizen services and analysis with human oversight, on infrastructure you control.](https://www.getdynamiq.ai/industries/public-sector) - [Telecommunications · Voice and chat care at scale, in your customers' languages.](https://www.getdynamiq.ai/industries/telecommunications) - [Healthcare · Patient intake and document workflows with PII detection and private endpoints.](https://www.getdynamiq.ai/industries/healthcare) ## Work with us Join the team, partner with us, or get in touch. - [Careers · Open roles in engineering and sales.](https://www.getdynamiq.ai/careers) - [Partner program · For services firms, resellers and technology companies.](https://www.getdynamiq.ai/become-a-partner) - [Contact · Questions, press and partnerships.](https://www.getdynamiq.ai/contact) ## See an agent on your own workflow. Bring a process and its documents. Our engineers will show you how Dynamiq runs it, in your environment or ours. [Start free](https://app.getdynamiq.ai/signup) · [Talk to the team](https://www.getdynamiq.ai/book-a-demo) --- # AI Coworker URL: https://www.getdynamiq.ai/product/chat AI Coworker is a ready-to-use agent that plans, searches, browses, writes code and produces files in its own cloud sandbox. It connects to your apps, runs on a schedule and works alongside your team in Slack, Microsoft Teams and Telegram, not just on the web. Teams use it for research, document work and routine automation without having to build a workflow first. ## What it does - **Own cloud computer:** Every conversation gets an isolated Linux sandbox where the agent runs shell commands, writes files and serves live web previews. - **Connect your apps:** Turn on Google Drive, Gmail, Notion, Slack and database connectors per conversation, personally or for the whole organization. - **A skills library:** Load packaged instructions for a class of task, from official skills or ones your team writes and shares. - **Scheduled tasks:** Have the agent run a prompt daily, weekly or monthly, with files attached, and land the result in a new conversation or a Microsoft Teams chat. - **Memory you can see:** It keeps durable notes about each person across conversations, and people can see and correct what it remembers on a Memory page. - **Subagents for big jobs:** The agent spawns isolated helper agents in parallel for deep research or fan-out work, then folds their results back in. - **Inside your chat apps:** The same agent answers in Slack, Microsoft Teams and Telegram, and in Slack it reads the thread for context and replies with files, not just text. ## A sandbox, not just a chat window Each conversation can get an isolated cloud sandbox, a real Linux machine the agent runs commands on. It installs packages, writes and reads files under its own working directory, generates images and speech, and can start a web server and hand you a live preview link. Close the conversation and come back later: the platform reattaches to the same sandbox, or provisions a fresh one if it expired, while your files stay downloadable from the chat. - Shell commands, file read and write, and code execution in one sandbox - Browser automation that can hand control back to you for a login - Web previews for sites the agent builds, viewable in desktop or mobile width - Pick the model for each conversation, or talk to your own deployed Apps as custom agents ## Connectors and skills Connect Google Drive, Gmail, Notion, Slack, Jira and dozens of other apps with one OAuth consent, then toggle each on or off per conversation. An owner or admin can also connect an app once for the whole organization, so a shared Slack workspace or production database is available to every teammate. Skills work alongside connectors: a Skill is a packaged SKILL.md procedure the agent reads when a task matches, built from your own instructions, imported from GitHub, or added from Dynamiq's official library. - Personal connections for your own accounts, organization connections for shared ones - SQL database connectors reach production systems on a private network through an SSH tunnel - Skills are the same format used by workflow agents, so they carry over - Register your organization's MCP servers once, and set organization-wide and personal instructions - Usage analytics show who uses AI Coworker and for which work ## Scheduled tasks, and a coworker in your chat apps A scheduled task runs a prompt on a timer, daily, weekly, monthly or once, and lands each run as its own conversation you can open later. The same agent also works as an AI coworker in Slack, Microsoft Teams and Telegram, wherever your team already talks. In Slack it reads the surrounding thread for context, confirms before anything risky, and returns reports, spreadsheets and decks as native file uploads instead of a wall of text. - Run history shows every past run with its status and full transcript - Run a task immediately with Run now, without waiting for its next slot - In channels the agent uses the organization's shared connections, not yours alone ## Built in, no setup Browser, Web search, Image generation, Text to speech (ElevenLabs v3) ## Questions and answers ### Is AI Coworker the same as a workflow agent? No. AI Coworker is ready to use: you talk to it directly, and its tools and cloud sandbox are built in. A workflow agent is one you design yourself in Agent Builder, with your own tools, knowledge and guardrails, then deploy as an App. The Custom Agents tab lets you talk to any deployed App from the same window. ### Can I use AI Coworker from Slack, Microsoft Teams or Telegram? Yes. The same agent that works on the web also works inside Slack, Microsoft Teams and Telegram, so a teammate can hand it work without leaving the channel they are already in. ### Does every conversation get its own computer? Yes. Each conversation can provision an isolated cloud sandbox where the agent runs shell commands, writes files and can serve a live web preview, separate from every other conversation. ### What is the difference between a connector and a skill? A connector is an OAuth authorization to an app like Gmail or Slack that the agent can call as a tool. A skill is a packaged set of instructions the agent reads when a task matches, such as a house style for a report or a checklist for an audit. ### Who can see a shared conversation? Sharing a conversation publishes a read-only snapshot up to that moment at a public link. Anyone with the link can view the transcript and download its files without a Dynamiq account, and nothing you send afterward is included until you update the snapshot. --- # Voice Agents URL: https://www.getdynamiq.ai/product/voice-agents Voice Agents are real-time phone and web agents that pick up inbound calls, reason with an LLM and speak back with text-to-speech, or run as one speech-to-speech model instead. They reach callers over WebRTC in the browser, SIP on real phone numbers, or a WebSocket for your own telephony backend. Every agent can call tools, hand off to a person, and gets tested against simulated callers scored by an LLM judge before it takes a real call. ## What it does - **Three ways in:** Answer inbound calls over SIP on numbers you bring, WebRTC from a browser or app, or a WebSocket from your own backend. Dynamiq has no outbound dialer. - **Knowledge during the call:** Search your knowledge bases mid-call, so answers come from your own policies and documents while the caller waits. - **A curated model catalog:** 15 voice model vendors for speech-to-text, language models, text-to-speech and speech-to-speech, each with its measured latency and cost. - **Per-turn latency:** Every call shows a waterfall of transcription, end-of-turn detection, LLM time to first token and speech synthesis, and turn-taking and key-term settings tune how the agent listens. - **Tools and transfers:** Give the agent HTTP tools, client tools your app answers, MCP servers and labeled transfer destinations to hand a caller to a person. Callers can key in digits, and the agent can send tones to work through another system's phone menu. - **Simulations, not guesswork:** Run scripted callers against the deployed agent and have an LLM judge score pass rate, accuracy and experience before real callers do. ## A curated model catalog, measured Every stage of a pipeline agent, speech-to-text, the language model and text-to-speech, and every realtime speech-to-speech model, comes from a catalog Dynamiq has wired up and measured, so you pick from combinations known to work rather than assembling one yourself. An estimate bar above the fields shows the configuration's cost per minute and latency before you deploy, so a choice is a comparison, not a guess. Across every stage, Dynamiq's catalog spans 15 voice model vendors. - The default combination, Cartesia, OpenAI and ElevenLabs, is chosen for latency, not for any single benchmark - Speech-to-text models with native end-of-turn detection remove a whole pipeline stage instead of speeding one up - Bring your own OpenAI-compatible LLM endpoint, or a custom voice ID from your own provider account ## Where the milliseconds go A caller judges a voice agent mostly on timing. On a pipeline agent, Dynamiq measures four stages on every turn: end-of-turn detection, transcription, LLM time to first token and text-to-speech time to first byte, and shows them as a proportional bar so the stage to fix is visible at a glance. Picking a speech-to-text model with native end-of-turn detection, such as Deepgram's flux models or Cartesia's ink-2, removes a whole stage instead of speeding one up, which is usually the biggest single latency win available. - Turn sensitivity and end-of-speech silence tune how eagerly the agent replies - Realtime, speech-to-speech agents report time to first token instead of a stage breakdown - Test calls and voice-mode simulations are the only way to get real numbers ## Tested before it takes a real call Simulations run a set of scripted callers, defined by a persona, a goal and an expected outcome, against the deployed agent and have an LLM judge decide whether it did what you said it should. Run them as text for fast, cheap iteration on reasoning and tool use, or as voice to also exercise speech recognition and synthesis and get real latency percentiles. Turn a call that went wrong into a scenario with one click from the Calls tab, so every regression becomes a permanent test. - A run scores scenarios passed, accuracy and experience, with a judge rationale per scenario - Generate starter scenarios from the agent's own configuration, or write them by hand - Simulations run against the deployed agent, so results reflect production behavior ## Voice model vendors Deepgram, AssemblyAI, Cartesia, ElevenLabs, Speechmatics, OpenAI, Groq, Baseten, Kokoro (via DeepInfra), Google, xAI, Anthropic, Cerebras, Fireworks AI, Together AI ## Questions and answers ### Can Dynamiq Voice Agents make outbound calls? No. Voice Agents answer inbound calls only, whether they arrive over the phone, the web or your own backend. There is no outbound dialing API. ### How do callers reach a voice agent? Over WebRTC from a browser or mobile app, over SIP from a phone number you connect through your own provider such as Twilio or Telnyx, or over a WebSocket from a backend that already holds the audio, such as a telephony stack. ### How is a voice agent tested before it goes live? You can talk to it yourself from the Test tab, and you can run simulations: scripted callers with a persona and an expected outcome, scored by an LLM judge for pass rate, accuracy and experience. Voice-mode simulations also produce real latency numbers. ### Do configuration changes need a redeploy? No. The worker reads the agent's current instructions, models, tools and settings at the start of every call, so an edit applies to the next call immediately. You only redeploy to pick up a new platform worker version. ### What happens when the agent cannot help a caller? If you configure transfer destinations, the agent can hand the caller to a person with a cold transfer, announcing the handoff first. Without any destinations configured, it cannot transfer and says so instead of failing silently. --- # Agent Builder URL: https://www.getdynamiq.ai/product/agents Agent Builder is Dynamiq's workflow engine: a drag-and-drop canvas for business teams and the open-source Python SDK for engineers, running on the same execution engine underneath. An agent is nodes on a graph, an LLM, tools, optional sub-agents and orchestration, wired between an Input and an Output node. Start from a blank canvas or from a prompt the generator turns into a first draft you refine. ## What it does - **One engine, two surfaces:** Build visually on the canvas or in the Python SDK; both produce the same workflow definition and deploy the same way. - **Generate from a prompt:** Describe the workflow in plain language and the generator drafts nodes and edges on the canvas for you to refine. - **Versions and rollback:** Every save is a new version. Deploy any saved version and roll back to an earlier one when you need to. - **Decision tables and rules:** Put eligibility checks, pricing grids and policy rules in tables and rule lists a domain expert can read and edit, with saved test cases that show what each rule covers. - **Judgment you can route on:** Ask yes or no, choice and score questions about a record and route on the answer, with an LLM, an agent or a calibrated judge service answering. - **Approvals on tool steps:** Pause a tool node for human sign-off before it runs, with fields a reviewer can edit and a message that shows exactly what will execute. - **Agentic memory:** Agents keep conversations by user and session, and long-term memory carries durable facts about each person into every new conversation. - **Agent tools:** Attach web search, databases, code execution, a knowledge base, MCP servers or sub-agents, and actions across 2,100+ app integrations, and the agent decides which to call. ## One engine, two ways to build Business teams build on the visual canvas: drag nodes from the palette, wire them with typed edges, and configure each one in an inspector panel. Engineers build the same graph in the open-source Python SDK, in git, with code review and their own CI. Both compile to the same workflow definition and the same execution engine, and a canvas workflow can be exported for engineers to run and extend with the SDK. - Every workflow starts as Input and Output nodes; everything else goes between them - Save creates a new version; only a saved version can be deployed or rolled back to - The SDK's to_yaml_file and from_yaml_file read and write the workflow definition, and Export downloads a canvas workflow as a ZIP that runs anywhere the SDK runs - Versioned prompt templates, tested in a playground, plug into LLM nodes and AI Coworker commands - The dynamiq CLI manages projects, workflows, versions, apps and triggers from a terminal ## A first draft from a prompt Describe the workflow in plain language, such as review a loan package against credit policy and route exceptions to an underwriter, and the generator drafts the nodes, edges and prompts in a side panel. Accept the draft onto the canvas, then refine each node in the inspector: swap the model, attach tools and knowledge, add an approval step. Nothing is saved until you click Save, so a draft costs nothing to try. - The generator drafts nodes, edges and prompts; you review each one before anything runs - Each save becomes a new version you can deploy or roll back to - Swap in your own Connections before testing, so the draft runs against your systems ## Decision nodes and approvals A Choice node routes a workflow on deterministic conditions, no model involved: the first matching branch wins, and a built-in default branch catches everything else. For process steps that need real judgment, attach a Graph Orchestrator instead, so a manager agent routes only where you let it. Either pattern can feed an approval gate: enable execution approval on a tool node and the run pauses with a rendered message until a reviewer approves, optionally editing the exact fields that are about to be sent, or rejects with feedback the agent reads and acts on. - Order Choice branches from most to least specific; the first true condition wins and skips the rest - Approval requests arrive over the same WebSocket or Runs API stream your App already uses - A rejected approval is not a dead end: its feedback flows back as an observation the agent can act on ## Business rules your experts can read Some decisions should never be left to a model. A Decision Table turns an eligibility check, a pricing grid or a scorecard into a step a credit or compliance lead can read and edit: typed inputs, a grid of rules, a hit policy for overlaps, and CSV import and export. A Rules node checks a record, such as a claim with its policy or an invoice with its purchase order and receipt, and reports one finding per rule with a severity and a message. Where a question needs judgment, the Judgement node answers yes or no, choice and score questions with a calibrated confidence the workflow routes on, and sends low-confidence cases to a person. - Saved test cases re-run after every change and show which rules they cover - Rules carry severities and effective dates, so a policy change has a start date - A judge can be an LLM, an agent that looks things up first, or a calibrated judge service ## Agentic memory Turn on memory for an Agent node and it keeps each conversation, scoped by user and session, and replays the most recent messages, the most relevant ones, or both, at the start of the next run. Long-term memory goes further: the agent gets tools to remember and recall durable facts about a user across every session, and decides during its own loop when a fact is worth keeping. When it hears the same thing said differently, it updates the existing fact instead of storing a duplicate. - Session memory scoped by user_id and session_id, with recent, relevant or both retrieval - Long-term facts the agent saves, recalls and updates itself; in AI Coworker, people see and correct them on a Memory page - Sub-agents keep their own history, so a delegate never reads another agent's memory ## Questions and answers ### What is the difference between the canvas and the Python SDK? They are two surfaces over the same execution engine. The canvas is for building visually and iterating with non-engineers in the room; the open-source Python SDK is for defining workflows in code, in git, with your own review process. A workflow moves between them through the same YAML definition. ### Can I start from a description instead of a blank canvas? Yes. Describe the workflow in plain language and the generator drafts nodes, edges and prompts for you to review on the canvas. Nothing is saved until you choose to save it. ### Can an agent call other agents? Yes. Attaching one Agent to another as a tool creates a sub-agent, so a manager can delegate a task and either combine the result or return the sub-agent's answer directly. A factory-built sub-agent gets a fresh instance per call, so a coordinator can fan out several in parallel. ### How do approvals fit into a workflow? Turn on execution approval for any tool node and the run pauses before that node executes, showing a reviewer a rendered message and, optionally, editable fields. Approve and the node runs with whatever the reviewer confirmed; reject with feedback and the agent reads that feedback as its next observation. ### Does a workflow need a data science team to build? No. The canvas is built for a business team to design an agent by dragging and connecting nodes, with a prompt-to-workflow generator to draft the first version. Engineers extend the same workflows in Python when the logic needs custom code. --- # Apps and deployments URL: https://www.getdynamiq.ai/product/deployments Apps are deployed workflows: each one gets a hostname, versioned releases and a full run history. Call an App over HTTP with sync, streaming or async requests, embed it as a chat widget, or fire it on a cron schedule or an event from a connected app. Every deployment is recorded, so rolling back to an earlier workflow version is a redeploy, not a rebuild. ## What it does - **Three ways to call it:** Get a single response, stream tokens over server-sent events, or fire an async run that posts its result to your callback URL. - **A chat widget in minutes:** Embed a ready-made chat widget or share a hosted chat link for a public App, no custom frontend required. - **Scheduled and triggered runs:** Run an App on a cron schedule or a one-off time, or fire it from an event in a connected app such as a new Slack message. - **Versions and rollback:** Every deployment pins a workflow version, a runtime and a config; rolling back is redeploying an earlier version to the same hostname. - **One account per end user:** Flag a node as a requirement instead of a shared connection, and each end user of your App authorizes their own account before it runs for them. - **A monitoring tab that reconciles:** Cost, tokens, requests and latency, charted by hour, day, week or month, next to the traces that explain any spike. ## Ship it three ways Call an App's hostname directly over HTTP for a synchronous response, or set stream to true for server-sent events that deliver output as the workflow produces it. For long runs you would rather not hold a connection open for, set execution_mode to async and pass one or more callback URLs; the App answers immediately with a request id and posts the result to each one when the run finishes. The same hostname also embeds as a chat widget or a hosted chat page once the endpoint is public, so a workflow becomes a product surface without a custom frontend. - Endpoint Authorization decides whether a Bearer Access Key is required on every request - A user_id and session_id pair groups runs into a remembered, multi-turn conversation - The Integration tab generates the exact Python and TypeScript snippets for your App - The Runs API creates synchronous, streaming or background runs, uploads files, and lists or cancels runs ## Versions, rollback and monitoring Each deployment freezes a workflow version, a runtime and a deployment config, and the App's hostname never changes underneath it, so callers never notice a redeploy. The History tab lists every deployment with who started it, its status and its runtime; rolling back is picking an earlier version from Existing deployment and deploying it again. The Monitoring tab charts cost, tokens, requests and latency over a date range you pick, and every run's full execution trace is one click away when a number looks wrong. - Variables, triggers and access settings are not versioned; they keep their current values through a rollback - Serverless is the default; server-based Apps add replica autoscaling with a target CPU percentage - Download a run's trace as JSON, or export every trace in a date range at once ## One credential per end user Not every credential should be shared. Flag a node's Connection, or an Action tool's linked account, as a requirement instead of a concrete value, and the platform resolves it per end user at run time instead of at build time. Before running the App for a user, check whether their requirements are satisfied; if not, mint a short-lived connect token and send them to a hosted setup page, or drive the same connect API from your own product. - A connection requirement takes credentials directly; an app-event requirement sends the user through OAuth consent - Connections created this way are end-user scoped and never appear in the project's shared Connections list - A single workflow can mix a shared connection for its LLM with a per-user requirement for one action ## Questions and answers ### What is the difference between an App and a workflow? A workflow is the graph you design in the editor. An App is a deployed, hosted instance of one saved workflow version, with its own hostname, run history, monitoring and traces. ### Can I roll back a bad deployment? Yes. The History tab lists every past deployment of an App; redeploying an earlier workflow version to the same App is the rollback, and the hostname, Access Keys and triggers stay exactly as they were. ### How do I embed an agent in my product? Call the App's hostname directly from your backend with an Access Key, or, for a public App, embed the ready-made chat widget or share the hosted chat link. Both talk to the same endpoint your own integration would use. ### How does a deployed App handle each end user's own accounts? Flag a node as a requirement instead of a shared connection, and each end user authorizes their own account once through a hosted setup page or your own UI before the App runs for them. ### What shows up on the Monitoring tab? Cost, token usage, request counts and latency for the App, charted over a date range you pick and bucketed by hour, day, week or month, with every run's full trace one click away. --- # Knowledge URL: https://www.getdynamiq.ai/product/knowledge-rag Knowledge bases turn files, websites and connected apps into grounded context agents can retrieve. Ingestion runs through configurable lanes, pre-processing, chunking, vectorization and storage, across 8 vector stores and 8 embedding providers. Turn on a knowledge graph alongside the vector store for GraphRAG, and every retrieval respects the permissions and filters you set at ingestion. ## What it does - **Three ways to fill it:** Upload files directly, crawl a website with depth and path filters, or sync Google Drive, Notion, Dropbox, SharePoint, Box or Confluence on an automatic schedule. - **An editable ingestion pipeline:** Pre-processing, chunking, vectorization and storage are real workflow nodes you can inspect, swap and extend in the same editor. - **Document parsing built in:** A multi-file converter routes PDFs, DOCX, PPTX, images and text to the right converter automatically, with AWS Textract, Google Document AI or an LLM for scans and unusual layouts. - **Permission-aware retrieval:** Filter search by the metadata you attach at ingestion, so a retriever only returns chunks a given user or department is allowed to see. - **Knowledge graphs:** Turn on GraphRAG to extract entities and relationships alongside the vector store, so an agent can follow connections, not just match text. - **Bring your own stack:** Use Dynamiq managed storage or point a knowledge base at your own store, across 8 vector stores and 8 embedding providers, or, from the SDK, search Bedrock Knowledge Bases and Vertex AI RAG Engine directly. ## Parsing and extraction Every file first goes through a multi-file converter that routes it by type: PDF, DOCX, PPTX, CSV and text converters each handle their own format, an LLM image converter reads scans and photos, and an unstructured-file converter catches anything else. The same LLM-based OCR pipeline is available directly through the AI Gateway's document endpoints: parse a PDF or image to clean Markdown, or extract structured JSON that matches a schema you define, such as an invoice's line items and totals. - Swap in a dedicated converter, such as one just for CSV exports, without touching the rest of the pipeline - A document splitter cuts by character, word, sentence, page, passage or title, with size and overlap you tune - Structure-aware splitters (recursive, Markdown header and semantic) keep a section together where fixed-size chunks would split it ## Retrieval that respects permissions Every chunk carries the metadata you attached when it was ingested, department, source, access tier, whatever your documents need, and a retriever's Filters restrict search to chunks matching conditions built from that metadata. The same retriever can run per request rather than per workflow: your integration passes a caller's tenant or department as a parameter at invoke time, and the platform merges it into the retriever's filter before the model ever sees it, so one deployed agent can serve many users without leaking across a boundary. - Queries embed with the same model used at ingestion; there is no separate query embedder to keep in sync - A similarity threshold drops weak matches; hybrid search blends vector similarity with keyword matching - Every retriever call is traced with its query and the chunks it returned, so a wrong answer is easy to audit - With permission syncing on, Google Drive and Confluence sources apply each person's own access at retrieval ## Knowledge graphs for multi-step questions A knowledge base stores chunks and finds them by similarity; a knowledge graph stores the same documents as facts, such as a person working at an organization, and finds them by connection. Turn on Build knowledge graph when you create a knowledge base, choose an extraction model and a graph database, and define the entity types, relationship types and allowed triples the model may extract. Every document you add from then on populates the vector store and the graph together, and an agent with that knowledge base attached can follow relationships an ordinary vector search would miss. - The extraction schema is an ontology: anything outside it is discarded, not written - Edit the ontology later from the knowledge base's Ontology tab; changes apply to future extractions only - The Knowledge Graph Writer resolves duplicate entities across ingestion runs instead of creating copies ## Vector stores Pinecone, Weaviate, Qdrant, Milvus, Chroma, Elasticsearch, OpenSearch, pgvector ## Embedding providers OpenAI, Cohere, AWS Bedrock, Mistral, Gemini, Hugging Face, IBM watsonx, Vertex AI ## Questions and answers ### How many vector stores and embedding providers does Dynamiq support? 8 vector stores and 8 embedding providers are available for a knowledge base, or you can use Dynamiq's managed storage and skip the choice entirely. ### Does retrieval respect who is allowed to see what? Retrieval filters on the metadata attached at ingestion, so you can restrict a search to a department, a tenant or any other boundary your documents carry. Per-request filters can also be injected at invoke time, outside what the model itself can see or change. ### What is a knowledge graph and when do I need one? A knowledge graph extracts entities and relationships from your documents alongside the normal vector store, so an agent can answer questions that depend on connections between things, not just similar wording. Turn it on when you create a knowledge base and define which entity and relationship types matter to your documents. ### Can I change the embedding provider later? You can, but queries are always embedded with the same model used at ingestion, so switching providers makes existing vectors unsearchable until you reprocess the affected items. ### Do I have to use Dynamiq's managed vector storage? No. A knowledge base uses Dynamiq managed storage by default, backed by Weaviate, but you can point it at your own vector store connection instead and keep the same managed ingestion pipeline. --- # Integrations URL: https://www.getdynamiq.ai/product/integrations Integrations give a workflow or agent access to 2,100+ app integrations, 11,800+ actions and 3,300+ triggers, plus databases, data warehouses and any Model Context Protocol server. A credential is either an organization-owned Connection shared across a project, or a per-user account each end user of a deployed App links themselves. Every action call is traced like any other node, so what an agent did in a third-party app is auditable after the fact. ## What it does - **A single action catalog:** Send a Gmail message, post to Slack, update a CRM record and more, across 2,100+ app integrations with 11,800+ actions. - **Triggers from any connected app:** Fire a deployed App from 3,300+ triggers across the same catalog, a new Slack message or an incoming email among them. - **MCP servers as tools:** Point an agent at any Model Context Protocol server and every tool it exposes becomes an individual agent tool automatically. - **Databases and warehouses:** Query PostgreSQL, MySQL, Snowflake and Redshift directly (Databricks SQL from the SDK), or reach a private database through an SSH tunnel. - **Org connections or per-user:** Share one connection across a whole project, or require each end user of a deployed App to link their own account. - **Every credential encrypted:** Secrets are stored server-side and never embedded in a workflow definition; a workflow references a connection only by its id. ## One catalog, three ways to use it An Action tool runs one operation of one app: search the catalog, pick an action such as sending a Gmail message or updating a CRM record, and the tool's schema is generated for you, with any field you fill in at build time made optional for the model to override. The same catalog powers triggers, so an App event trigger can start a run when a new message lands in a channel or a form is submitted, and AI Coworker's connectors give it the same apps for one conversation at a time. Across all three, Dynamiq connects to 2,100+ app integrations with 11,800+ actions and 3,300+ triggers. - Add all actions on an app row to hand the agent its whole surface at once - Some fields load their options live from the linked account, such as a Slack channel list - A failed action with a 400, 401, 402 or 422 status comes back as an observation the agent can react to ## MCP servers Model Context Protocol is the open standard for exposing a tool server to an agent. Add an MCP Streamable HTTP or MCP SSE connection with the server's URL and any auth headers it needs, attach an MCP Server tool to an agent, and at initialization the node asks the server which tools it offers and turns each one into a typed agent tool automatically. Include or exclude specific tool names to keep a large server's surface from crowding out everything else in the agent's prompt. - Schemas are discovered at run time, so a server's new tools appear without a workflow change - Servers that only support interactive OAuth need a local stdio bridge; token auth works directly - A tool call that fails comes back to the agent as a recoverable observation, not a run failure ## Organization connections and per-user connections Most integrations belong to a project and are shared by whatever runs there, an LLM provider key, an internal database, a service account nobody wants to reauthorize per user. Some actions have to run as the person using the product instead, reading their inbox or posting to their own workspace. Flag that node as a requirement and each end user of a deployed App authorizes their own account once, through a hosted page or your own UI, kept completely separate from every other user's connection. - A single workflow can mix a shared Connection for its LLM with a per-user requirement for one action - AI Coworker's own connectors are a third, separate system: personal or organization-wide, used only inside a conversation - Deleting a source or a connection removes what it authorized; nothing lingers for another user to inherit - Mark an integration optional, so end users connect only the apps they choose through scoped connect links ## Catalog categories Productivity, Communication, Developer tools, Cloud, Monitoring, Analytics, Security, Database ## Questions and answers ### How many apps can a Dynamiq agent act in? 2,100+ app integrations are in the catalog, with 11,800+ actions and 3,300+ triggers across them, covering productivity, communication, developer tools, cloud, monitoring, analytics, security and database apps. ### What is an MCP server? Model Context Protocol is an open standard for a tool server an agent can call. Connect one over Streamable HTTP or SSE, and every tool it exposes becomes an individual agent tool, discovered automatically rather than defined by hand. ### What is the difference between an organization connection and a per-user connection? An organization connection is shared credentials every run in a project uses, set up once by a builder. A per-user requirement instead has each end user of a deployed App authorize their own account, kept separate from every other user's. ### Can an agent reach my own database? Yes. PostgreSQL, MySQL, Snowflake and Redshift connections are available directly, Databricks SQL from the SDK, and a database on a private network can be reached through an SSH tunnel to a bastion you expose. ### Are integration credentials stored safely? Secrets are encrypted server-side and never embedded in a workflow definition. A workflow references a connection by its id only, and rotating a credential updates every node that uses it without a redeploy. --- # Models and AI Gateway URL: https://www.getdynamiq.ai/product/models The AI Gateway is one OpenAI-compatible endpoint for the most-used models, resolving credentials server-side so your code holds one Dynamiq key instead of one per provider. Swap models by changing a model slug, with usage metered per organization. The SDK supports 29 model providers through one LLM node interface. For models you want to run yourself, deploy open models such as Llama, Gemma, Mistral and Qwen on managed inference, and fine-tune them with LoRA adapters served dynamically over a shared base model. ## What it does - **One endpoint, many providers:** The router speaks the OpenAI chat completions protocol; change the model slug and the request goes to a different provider, unchanged otherwise. - **One credential:** A single Dynamiq Access Key replaces a provider key per service, and revoking it is one action in Settings. - **29 model providers:** From OpenAI and Anthropic to self-hosted Ollama, every provider node shares the same parameters, streaming and tool-calling shape. - **Prompt caching by default:** Calls to Anthropic models and Bedrock reuse cached prompt prefixes by default, so long system prompts and documents cost less and answer faster. - **Open models you host:** Deploy Llama, Gemma, Mistral, Qwen and other open models on managed inference behind an OpenAI-compatible endpoint, without managing GPUs yourself. - **Fine-tuning with LoRA adapters:** Serve fine-tuned adapters dynamically over one shared base model instead of a separate deployment per adapter. ## One endpoint, one credential Point any OpenAI SDK client at the gateway's base URL with a Dynamiq Access Key instead of a provider key, and everything else about your integration stays the same: same request shape, same streaming format, same response. The model field is a Dynamiq router slug; the gateway resolves which provider serves it, authenticates with a Dynamiq-managed connection on the server side, and relays the request, so your client never holds a provider credential. Usage is metered against your organization's plan in one place, and switching models is changing one string. - Streaming returns server-sent events of chat.completion.chunk objects, exactly what an OpenAI client expects - The router covers chat completions only; call a provider directly for embeddings, audio or image endpoints - Extra request fields pass through to the upstream provider, so provider-specific parameters still work ## Providers behind one node Every LLM is a node that subclasses the same base class, so swapping OpenAI for Anthropic, Gemini or a self-hosted Ollama model changes the class, the model name and the connection; the prompt, parameters, streaming, tools and output shape stay identical. Dynamiq supports 29 model providers, from the major hosted labs to inference platforms such as Groq, Fireworks, Together and Cerebras, plus a generic OpenAI-compatible connection for anything else. - A fallback config runs a secondary LLM automatically on a rate limit or connection error - Vision and PDF input are reported per model, with an override for a model the registry does not know - The same node classes power the AI Gateway underneath, so platform and SDK stay in sync ## Open models and fine-tuning Deploy an open model such as Llama 3, Gemma, Mistral or Qwen on a vLLM runtime behind its own hostname and an OpenAI-compatible endpoint, with replica autoscaling between a minimum and maximum you set. Fine-tune LoRA adapters on top of a base model and serve them from one shared deployment that loads an adapter by name, instead of standing up a separate deployment for every fine-tune. Point a request's model field at an adapter alias to route to that specific adapter, or leave it on the base model. - Embedding and speech-to-text models deploy the same way, behind the matching OpenAI-compatible API - Creating an inference deployment also creates a system-managed connection, so workflow nodes call it with no extra setup - Pods and their live logs are visible from the deployment page while a large model downloads and loads ## Model providers OpenAI, Anthropic, Gemini, AWS Bedrock, Azure AI, AI21, Anyscale, Cerebras, Cohere, Databricks, DeepInfra, DeepSeek, Fireworks AI, Groq, Hugging Face, Mistral, Nvidia NIM, Ollama, OpenRouter, Perplexity, Replicate, SambaNova, Together AI, Vertex AI, watsonx, xAI, Custom (OpenAI-compatible) ## Questions and answers ### How many model providers does Dynamiq support? 29 model providers are available through the same LLM node interface in the SDK, and the most-used ones are routable through the AI Gateway behind one endpoint and one credential. ### Do I need a separate API key for every provider? No. Calling a model through the AI Gateway needs only a Dynamiq Access Key; the gateway resolves the upstream provider's credentials server-side and your client never holds them. ### Can I run an open-source model instead of a hosted provider? Yes. Deploy an open model such as Llama, Gemma, Mistral or Qwen on managed inference and call it behind its own OpenAI-compatible endpoint, for cost, privacy or fine-tuning reasons. ### How does fine-tuning work on Dynamiq? Train LoRA adapters on top of a base model, then serve them over one shared deployment that loads an adapter by name, instead of standing up a new one per fine-tune. ### What does the AI Gateway not do? It proxies chat completions only. Embeddings, audio and image endpoints, and anything outside the chat completions shape, call the provider directly rather than through the router. --- # Guardrails and approvals URL: https://www.getdynamiq.ai/product/guardrails Guardrails are workflow nodes that detect PII, prompt injection and policy violations before an agent acts on them, and validators that check an output's shape afterward. Approvals pause a tool call or an agent plan for a person to review, with fields they can edit before it runs. Detectors flag; they do not rewrite or redact what they find, so every decision about a flagged message is yours to make. ## What it does - **Detect PII:** Flag names, emails, addresses and account numbers in a message before it reaches a model or leaves your boundary in a prompt. - **Detect prompt injection:** Classify whether a message is trying to override your system prompt, so you can branch it away from the agent. - **Content policy checks:** Screen a message against content-safety policies and see exactly which category it violated, so you can route it deliberately. - **Approvals with editable fields:** Pause a tool node until a reviewer signs off, and let them edit the exact fields that are about to run, not just approve blindly. - **Execution approval on any tool:** Gate the one step that touches the outside world, sending an email, writing to a database, calling a payment API, without slowing down everything else. - **Every verdict traced:** Every detector result and every approval decision is recorded in the run's trace, so a flagged input is reviewable after the fact. ## Detect before a model ever sees it Three detectors screen free-form text before it reaches an agent: a PII detector flags personal information such as names, emails and account numbers; a prompt injection detector classifies whether a message is trying to override your system prompt; and a policy detector checks a message against content-safety categories and reports which ones it violated. Each one returns a flag you branch on with a Choice node, routing a clean message to the agent and a flagged one to a refusal, a rewrite step you own, or a person. - Detectors are classifiers, not redactors: they tell you what was found, they do not rewrite the message - Chain a prompt injection check and a PII check before the same agent for defense in depth - Every detector verdict is recorded in the run's trace alongside the message it scored ## Approvals a person can edit Execution approval pauses a tool node every time it is about to run, not just when the agent decides to ask. Turn it on for the one step that matters, an email send, a database write, a payment call, and the run emits an approval request with a rendered message showing exactly what is about to happen. A reviewer can approve as is, edit any field you marked mutable before approving, such as the body of an email, or reject with feedback the agent reads as its next observation and can act on. - Orchestrator nodes have the same gate for a task plan, so a person approves the plan before any task runs - A paused run is checkpointed, so a reply minutes or days later resumes it from where it stopped - The Human Feedback tool covers the other direction: the agent asks a person a clarifying question mid-run ## Guard the step that matters Guardrails compose. Put a detector in front of an agent, a validator behind it checking the output is well-formed JSON or matches a required pattern, and an approval gate on the one tool that actually changes something outside the workflow. A validator can either return a valid flag you branch on, or fail the node outright and trigger its retry policy, giving a model one more attempt to produce well-formed output before the run fails. - Valid JSON, Valid Python, Valid Choices and Regex Match cover the common output contracts - A rejected approval is not a dead end; its feedback flows back into the agent's next step - Route a flagged branch to a person instead of a static refusal for genuinely ambiguous cases ## Detectors and validators PII detector, Prompt injection detector, Policy detector, Valid JSON, Valid Python, Valid choices, Regex match ## Questions and answers ### Does Dynamiq redact PII automatically? No. The PII detector tells you that personal information was found and what kind, so you can route the message away from a model or to a person; it does not rewrite or redact the message itself. ### What is the difference between a detector and an approval? A detector classifies a message, PII, prompt injection or a policy violation, so you can branch on the result. An approval pauses a specific step every time it is about to run and waits for a person to sign off, independent of what any detector found. ### Can a reviewer change what an approved step actually does? Yes, for fields you mark as mutable on that node. A reviewer can edit those values before approving, so what runs is what they confirmed, not just what the agent drafted. ### What happens if an approval is rejected? The node does not run, and the reviewer's feedback flows back into the workflow. An agent sees it as its next observation and can revise its plan instead of the run simply failing. ### Can I require approval on every tool call? You can, but it is usually more useful to gate the one step that touches the outside world, an email, a database write, a payment call, and let the agent plan and draft freely up to that point. --- # Evals URL: https://www.getdynamiq.ai/product/evaluations Evaluations score agent outputs against metrics: an LLM-as-a-judge rubric, a predefined RAG evaluator, or your own Python code. Datasets are versioned test data you build by hand, upload, or capture straight from production traces, and a run can either re-execute a workflow first or score what was already recorded. Beyond batch runs, a real-time eval attaches to a live deployment and samples its traffic continuously, and voice agents get their own simulation runs, scripted callers scored by an LLM judge before they take a real call. ## What it does - **Three metric types:** Write an LLM-as-a-judge rubric, use a predefined RAG evaluator such as faithfulness, or score with your own Python function. - **Datasets from real traces:** Capture a production trace into a dataset with one click, so a real failure becomes a permanent regression test. - **Offline runs before you ship:** Re-run a candidate workflow version against a released dataset and compare its scores with what is in production today. - **Real-time evals on live traffic:** Attach a metric to a deployed App and score a sampled share of its traces automatically, with no re-run and no extra latency for callers. - **Voice simulations:** Run scripted callers against a deployed voice agent and have an LLM judge score pass rate, accuracy and experience. - **Every score traced back:** A run's results table links each score to the trace it came from, so a low score is one click from the evidence. ## Datasets built from real traffic A dataset is a versioned table of test items you build by typing rows in, uploading a JSON file, or capturing production traces directly from an App's Traces tab. Draft versions are editable; releasing one freezes it so evaluation runs stay reproducible against exactly the data they scored. Turning a bad trace into a dataset item is the highest-value path: when a real call goes wrong, capture it, and it becomes a regression test that catches the same failure again. - Released versions are immutable; fork one when you need to change frozen data - A trace-captured item carries input, output and status; With workflow mode scores it as recorded, not re-run - Reshape captured items into flat rows to replay real inputs through a new workflow version instead ## Real-time evals on a live deployment An online evaluation attaches a saved metric to a deployed App from its Evaluations tab: pick a metric and version, a sample rate for what share of traffic to score, and an input path that pulls the metric's input out of each trace. Every finished trace becomes eligible as it arrives, sampled deterministically rather than by a random draw, so two evaluators at different rates never just score the same half of your traffic. Nothing is re-executed and callers see no difference; the score lands in the evaluator's own run list, one click from the trace that produced it. - Start with a high sample rate while validating a new evaluator, then lower it on high-traffic Apps - Failed and canceled runs are eligible too, not only successful ones - An LLM-as-a-judge evaluator's reasoning, not just its score, is there for every run ## Voice simulations A voice agent gets a form of evaluation the rest of the platform does not: scenario sets, scripted callers defined by a persona, a goal and an expected outcome, run against the deployed agent and scored by an LLM judge. Run a set as text for a fast, cheap check of reasoning and tool use while iterating, or as voice to also exercise speech recognition and synthesis and get real latency percentiles before you ship a change. - Generate starter scenarios from the agent's own configuration, then tighten the ones that matter - Turn a real call that went wrong into a scenario directly from the Calls tab - A run scores pass rate, accuracy and experience, with the judge's rationale for every scenario ## Predefined evaluators AnswerCorrectness, ContextPrecision, ContextRecall, FactualCorrectness, Faithfulness ## Questions and answers ### What is the difference between an evaluation run and a real-time eval? An evaluation run scores a fixed, released dataset on demand, useful for comparing a candidate workflow version against production before you ship. A real-time eval instead attaches to a live deployment and scores a sample of its actual traffic continuously, after you ship. ### Can I turn a real production failure into a test case? Yes. Open the trace from an App's Traces tab and add it to a dataset in one click; the input, output and status become a dataset item you can score or use as a regression test. ### What is an LLM-as-a-judge metric? A metric where an LLM follows a rubric you write and returns a score, suited to subjective qualities such as hallucination, relevance or tone that a fixed rule cannot capture. ### How are voice agents evaluated differently? Through simulations: scripted callers with a persona and an expected outcome run against the deployed agent, in text for fast iteration or in voice to also exercise speech recognition and synthesis, and an LLM judge scores the result. ### Does a real-time eval slow down my App? No. Traces are scored after the fact from what was already recorded; nothing re-executes and callers see no added latency. --- # Observability URL: https://www.getdynamiq.ai/product/observability Observability records every run as a trace: a tree of every node's input, output, status, timing and cost, replayable exactly as it happened. The Monitoring tab charts requests, latency, tokens and cost over time for a deployment, and the same trace explorer covers SDK workflows that ship their runs in with a callback handler. An AI agent reviews production traces on top of that record and surfaces issues, regressions and failures for you to check. ## What it does - **Every run traced:** Each run records its full execution tree: every node's input, output, status and timing, replayable exactly as it happened. - **Cost, latency and tokens:** See prompt and completion tokens, dollar cost and duration for every node, not just a total for the run. - **Monitoring over time:** Chart requests, latency, tokens and cost by hour, day, week or month, and drill from a spike straight into the traces behind it. - **Traces from SDK workflows:** Attach a tracing callback handler to a workflow running in your own infrastructure and its runs show up next to deployed App traces. - **An AI agent reviews traces:** An AI agent works through production traces and surfaces issues, regressions and failures for a person to check. - **Downloadable for offline analysis:** Export a single trace or every trace in a date range as JSON, with node-level runs included when you need them. ## A run tree with cost, latency and tokens on every node Open a trace and the execution structure renders as a graph, a tree or a timeline, with every node's resolved input, output, status and timing one click away. LLM and Agent nodes additionally show the rendered prompt after templating, so a wrong answer is traceable to the exact text the model saw, not just the node that produced it. The same view works whether the run came from a deployed App, a workflow test in the editor, or an open-source SDK workflow that shipped its trace in with a tracing callback handler. - Filter by status and start time, and download a single trace or a whole date range as JSON - A failed run or a failed node shows an error banner with the exact message - A voice call's transcript, recording and per-turn latency waterfall are the same idea, applied to a call ## Monitoring over time The Monitoring tab of an App charts four series over a date range you pick, cost split into prompt and completion tokens, total tokens, request counts split into successes and failures, and average latency, bucketed by hour, day, week or month. The same numbers back a metrics API you can pull into your own dashboards. A rising failure count or a latency spike is a starting point, not an answer; every chart is one click from the traces that explain it. - Traces from SDK workflows need a project-scoped Access Key; an org-scoped key is rejected - Delivery failures in a tracing handler are logged, never raised, so tracing cannot break a production run ## An AI agent that reviews your traces Beyond the charts and the trace tree, an AI agent works through your production traces on its own and surfaces issues, regressions and failures for a person to check, rather than waiting for someone to notice a chart moved. It gives a team a starting point for what changed and where, on top of the same trace record every run already produces, so nothing about how you trace or monitor a deployment changes to get it. - Built on the same traces the Traces and Monitoring tabs already record - A finding points at the run and the node it came from, not just a metric ## Questions and answers ### What does a trace actually record? Every node a run touched, with its resolved input, output, status, timing and, for LLM and Agent nodes, the rendered prompt. A run's overall input, output, total duration and token usage sit alongside it. ### Can I see traces from workflows that are not deployed on Dynamiq? Yes. Attach a tracing callback handler to a workflow running in your own infrastructure with the open-source SDK, and its runs report into the same trace explorer as deployed Apps. ### What is on the Monitoring tab? Cost, token usage, request counts and latency for a deployment, charted over a date range you choose and bucketed by hour, day, week or month, with every chart one click from the traces behind it. ### What does the AI agent that reviews traces actually do? It works through your production traces and surfaces issues, regressions and failures for a person to check, on top of the same trace record every run already produces. ### Is every run replayable later? Yes. Every run is traced and replayable, so you can open a run from last week exactly as it executed, node by node, to see what a model saw and what each step returned. --- # Financial services URL: https://www.getdynamiq.ai/industries/financial-services Dynamiq gives banks, lenders and capital markets firms one platform for back-office operations, customer service and compliance agents, deployed on Dynamiq Cloud or inside infrastructure you control. Back-office agents reconcile accounts, clear exceptions and prepare onboarding files, voice and chat agents answer customers on your own numbers and channels, and every consequential step pauses for a person. Every run is traced and replayable, so you can show a regulator exactly how an agent reached its answer. ## Challenges - **Back-office queues grow with the business:** Reconciliations, payment exceptions, disputes and onboarding files grow with every new product and customer, and each one still passes through several systems by hand. - **Support volume outpaces headcount:** Customers expect an answer at any hour, on the phone and in chat, without the inconsistency of a large, distributed support team. - **Every decision needs a paper trail:** Examiners and auditors expect to reconstruct exactly why an account was approved, an alert was closed, or a report went out. ## Agents we deploy - **Back-office operations:** Reconciles accounts and suspense items, investigates payment exceptions and prepares dispute cases, with approval before anything posts. - **KYC and customer onboarding:** Reads onboarding documents, gathers screening results and applies your risk rules, so an analyst approves a prepared file. - **Customer service by phone and chat:** Answers balance, card and password requests on inbound calls and chat, and transfers to a person when needed, with the transcript saved in the trace. - **AML alert triage:** Reviews flagged transactions and screening hits, drafts the writeup, and routes it to an analyst for sign-off. - **Loan file review:** Reads loan files and supporting documents, checks them against your credit policy, and flags exceptions to an underwriter. - **Market and credit research:** Answers analyst questions from research, filings and internal notes, with a citation back to the source for every claim. ## Controls - **Per-analyst credentials:** Database access resolves to each analyst's own login at run time, so the database's own grants decide what comes back. - **Approval before anything leaves the firm:** Payments, postings, reports and filings pause for a person to approve the exact output before delivery. - **SSO and scoped access:** SSO on Enterprise, organization roles, private projects and scoped access keys control who can build and run agents, and keys can expire on a date you set. - **Full audit trail:** Every run is traced and replayable, exportable in bulk for exam and audit requests. - **PII and prompt injection detection:** Customer messages and documents are screened for PII and prompt injection before a model sees them. ## Deployment Run on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster. Dynamiq holds SOC 2 and signs a DPA under GDPR. Fully isolated or air-gapped environments are delivered with our engineers, for workloads that cannot reach the public internet. ## Questions and answers ### Which banking back-office processes fit agents first? Repetitive, rules-heavy work with a clear approval point: reconciliations, payment exceptions and repairs, dispute preparation, KYC onboarding files and recurring regulatory reports. ### Can the same platform answer customer calls? Yes. Voice Agents answer inbound calls on the numbers you already own and chat agents answer in your app, using the same tools, knowledge and controls as your back-office agents. Voice is inbound only. ### Can Dynamiq run inside our own cloud account? Yes. Deploy on Dynamiq Cloud or self-host on AWS, Azure, GCP, IBM Cloud, OpenShift or your own Kubernetes cluster, including fully isolated environments delivered with our engineers. ### How does the platform handle database access for analysts? Each analyst's queries can resolve to their own database credentials at run time, so your database's own grants and row-level rules decide what comes back, not the model. ### Can an agent approve a transaction or report on its own? No. Anything that leaves the firm, a payment, a report or a filing, pauses for a person to approve the exact output first. ### Does every run leave an audit trail? Yes. Every run is traced and replayable, node by node, and exportable in bulk for exam and audit requests. ### What compliance does Dynamiq hold? Dynamiq holds SOC 2 and supports GDPR with a DPA. Compliance documentation is available on request. --- # Insurance URL: https://www.getdynamiq.ai/industries/insurance Dynamiq gives insurers one governed platform for policyholder service, claims intake and review, underwriting file preparation and back-office work, deployed on Dynamiq Cloud or inside infrastructure you control. Voice and chat agents answer inbound calls and messages, document agents read claims and submissions against your guidelines, and an adjuster or underwriter approves every coverage, pricing and payment decision. Every run is traced and replayable, so you can show an auditor or regulator how an agent reached its recommendation. ## Challenges - **Claims surges swamp every channel:** A storm, a flood or a busy travel season sends calls, emails and documents in at once, and adjusters end up answering status questions instead of working claims. - **Files arrive as documents, not data:** Submissions, estimates, invoices and medical reports come in as PDFs and scans that someone still has to read, re-key and check against the policy. - **Every decision is regulated:** Coverage, pricing and payment decisions need a named person, guidelines applied the same way every time, and a record an examiner can follow. ## Agents we deploy - **Claims intake and review:** Takes first notice of loss by phone or chat, opens the claim, reads the documents that follow and prepares the file for an adjuster. - **Policyholder service by phone and chat:** Answers policy, billing and claim-status questions on inbound calls and chat, and transfers to a person when needed, with the transcript saved in the trace. - **Underwriting submission review:** Reads submissions and supporting documents, extracts the risk details and flags anything outside appetite or guidelines to an underwriter. - **Policy and guideline assistant:** Answers adjuster and underwriter questions from policy wordings and guidelines, with a citation to the clause it used. - **Premium, commission and payment operations:** Reconciles premiums, commissions and claim payments across systems and drafts the report for approval before it is filed. - **Governance across every agent:** One set of guardrails, approvals, traces and evals across claims, underwriting and service agents, instead of separate controls for each. ## Controls - **A person decides coverage and payment:** Coverage decisions, denials, pricing and payments pause for an adjuster or underwriter to approve the exact output. - **Detectors before the model:** Policyholder messages and claim documents are screened for PII and prompt injection before a model reads them, and the verdict is recorded. - **Guidelines applied the same way:** Claims and underwriting checks run against your own guidelines in a knowledge base, with the clause cited on every recommendation. - **SSO and scoped access:** SSO on Enterprise, organization roles and scoped access keys decide who can build and run agents, with private projects for claims and underwriting data. - **Full audit trail:** Every run is traced and replayable, and traces export in bulk for internal audit and regulatory exams. ## Deployment Run on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster next to your policy and claims systems. Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Phone lines connect over SIP from your existing carrier, and fully isolated environments are delivered with our engineers. ## Questions and answers ### Can an AI agent take first notice of loss? Yes. A voice or chat agent takes the report, verifies the policy, captures what happened and opens the claim in your claims system, then prepares the file for an adjuster. ### Does the agent decide coverage or pay claims? No. It prepares the file and recommends; coverage decisions, denials and payments are always approved by an adjuster first. ### Can it help underwriters with submissions? Yes. It reads submissions and supporting documents, extracts the risk details into your schema, checks them against your appetite and guidelines, and flags exceptions to an underwriter. ### Can Dynamiq run inside our own cloud or data center? Yes. Self-host on AWS, Azure, GCP, IBM Cloud, OpenShift or your own Kubernetes cluster, or run fully isolated with our engineers. ### How is policyholder data protected? Detectors flag PII and prompt injection before a model reads a message or document, access follows your roles, and every run is traced. Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Detectors flag; they do not redact. --- # Government and public sector URL: https://www.getdynamiq.ai/industries/public-sector Dynamiq gives government and public sector teams agents for citizen inquiries, benefits and permit review, and budget or grant reporting, deployed on infrastructure your agency controls. Every run is traced and replayable, and a person approves anything that leaves the agency or affects a constituent's case. Built for teams that need to show exactly how an agent reached an answer, including on fully isolated infrastructure delivered with our engineers. ## Challenges - **Constituent volume without added headcount:** Citizens expect an answer to routine questions at any hour, across phone and web, without a budget for a proportionally larger team. - **Casework buried in paperwork:** Benefits applications, permits and case files arrive as scans and PDFs that someone still has to read and check against policy. - **Infrastructure and oversight requirements:** Agencies often cannot send data to a shared public cloud, and every automated decision needs a person accountable for it. ## Agents we deploy - **Citizen and constituent inquiries:** Answers routine questions on inbound calls and chat, and transfers to a caseworker for anything that needs judgment, with the transcript saved in the trace. - **Benefits and permit review:** Reads applications and supporting documents, checks them against program rules, and flags exceptions to a caseworker. - **Policy and program research:** Answers staff questions from agency policy, regulations and prior case notes, with a citation back to the source. - **Budget and grant reporting:** Reconciles spend against a budget or grant and drafts the recurring report, approved by a program manager before it is filed. - **Supplier invoices and payments:** Matches supplier invoices to purchase orders and receipts and routes exceptions to finance staff, with approval before any payment. - **Casework correspondence drafting:** Drafts case notes and constituent correspondence for a caseworker to review and send, rather than sending on its own. ## Controls - **Human oversight on every decision:** Anything that affects a constituent's case, benefit or filing pauses for a caseworker or program manager to approve. - **Infrastructure your agency controls:** Runs on your own cloud account, on-premises, or fully isolated with our engineers, for workloads that cannot reach the public internet. - **SSO and role-based access:** SSO on Enterprise and organization roles decide who can build and run agents, with private projects for sensitive casework. - **PII and prompt injection detection:** Constituent submissions are screened for PII and prompt injection before a model sees them. - **Full audit trail:** Every run is traced and replayable, so a decision can be reconstructed months later. ## Deployment Self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, on-premises, or fully isolated with our engineers for workloads that cannot reach the public internet. SOC 2 is in place, and GDPR applies wherever an agency handles EU resident data. ## Questions and answers ### Can Dynamiq run fully isolated, with no internet access? Yes. Our engineers deliver fully isolated or air-gapped deployments for workloads that cannot reach the public internet. ### Who approves what an agent decides? A person. Anything that affects a constituent's case, benefit or filing pauses for a caseworker or program manager to approve before it goes out. ### Can models and data stay in our country? Yes, when the platform runs on infrastructure in your country. Self-host it in your own cloud account or data center, serve open models on vLLM inside that environment, and limit the AI Gateway to those self-hosted models. ### Can this run on infrastructure we already operate? Yes. Self-host on AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, in your own account or data center. ### How do you keep casework private? Casework, connections and knowledge bases live in private projects, open only to the caseworkers and administrators you add. ### Can we show how an agent reached a decision? Yes. Every run is traced and replayable, including every document it read and every system it queried. --- # Telecommunications URL: https://www.getdynamiq.ai/industries/telecommunications Dynamiq gives telecom operators voice and chat agents that answer subscribers in their own language, back-office agents that reconcile billing and usage, and knowledge assistants for support and network teams. Calls can be recorded and transcribed, every chat is traced, and a person handles anything the agent cannot resolve. Built for operators running support and operations at a scale no team can staff by hand. ## Challenges - **Support spans languages and channels:** Subscribers call and message in several languages across a footprint no single support team can staff around the clock. - **Billing and usage reconciliation at volume:** Usage, billing and plan changes flow through several systems that rarely agree without someone checking them by hand. - **Operational knowledge scattered across teams:** Runbooks, outage history and network documentation live in enough different places that finding the right answer takes longer than fixing the issue. ## Agents we deploy - **Inbound voice and chat care:** Answers billing, plan and outage questions over phone and chat, and hands off to a person for anything it cannot resolve. - **Network and support knowledge assistant:** Answers questions from runbooks, outage history and network documentation, with a citation back to the source. - **Billing reconciliation and reporting:** Reconciles usage and billing across systems and drafts the recurring report, approved before it is sent. - **Multilingual voice deployment:** The same agent answers in the languages your subscriber base speaks, configured per market rather than rebuilt from scratch. - **Outage and incident updates:** Drafts subscriber-facing status updates for known incidents from the network operations feed, for a person to approve and post. ## Controls - **Escalation as a gate:** Disputes, account changes and anything the agent is unsure of are routed to a person rather than resolved automatically. - **Full call and chat trail:** Every call and chat is traced and replayable, with the transcript, recording and every lookup attached. - **Scoped, expiring access:** Access keys can be scoped to one project and set to expire, so a leaked key exposes only that deployment. - **PII and prompt injection detection:** Subscriber messages are screened for PII and prompt injection before a model sees them. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, close to your telephony and billing systems. Phone numbers connect over SIP from Twilio, Telnyx or your own carrier, with SOC 2 and GDPR in place. ## Questions and answers ### Can the same agent answer in several languages? Yes. Speech-to-text, the language model and the voice are configured per deployment, so the same agent design answers in the languages your subscribers use. ### Does Dynamiq support outbound calling or dialers? No. Voice Agents handle inbound calls only. ### How is a voice agent tested before it takes subscriber calls? Scripted callers with a persona, a goal and an expected outcome run against the deployed agent in text or voice, and an LLM judge scores pass rate, accuracy and experience before real subscribers reach it. ### How does billing reconciliation stay accurate across systems? The agent queries each system directly over API or SQL and reconciles the figures against your own rules, with a person approving the report before it goes out. ### Can support and network teams use the same knowledge assistant? Each team can have its own assistant, scoped to its own documents and permissions, or share a knowledge base where that makes sense. --- # Healthcare URL: https://www.getdynamiq.ai/industries/healthcare Dynamiq gives healthcare organizations agents for patient intake, referral document processing and clinical knowledge questions, deployed behind a private endpoint. Detectors classify PII and prompt injection before any reasoning model sees the text, and a clinician or intake coordinator reviews anything a detector flags or that falls outside the expected schema. Built for organizations that need a BAA for HIPAA workloads, a DPA under GDPR and a full trace of every run. ## Challenges - **Intake buried in scanned paperwork:** Referral letters, medication lists and lab summaries arrive as scans that someone still has to retype into the record system. - **Patient data needs screening before any model sees it:** Every document contains PII by nature, and it has to be classified before a reasoning model reads it, not after. - **Clinical knowledge scattered across teams:** Intake protocols, coding guidance and referral rules live in enough different places that staff spend real time just finding the current version. ## Agents we deploy - **Patient intake and scheduling:** Answers scheduling and intake questions over phone and chat, and hands off to staff for anything clinical or urgent. - **Referral and document intake:** Converts referral packets to structured intake data, screened for PII and prompt injection before a model reads them. - **Clinical guideline assistant:** Answers staff questions from intake protocols and coding guidance, with a citation back to the source passage. - **Claims and prior-authorization support:** Gathers the supporting documentation for a claim or authorization request and drafts it for staff to submit. - **Supplier invoices and payments:** Reads supplier invoices, matches them to purchase orders and receipts, and routes exceptions, with approval before anything posts or pays. - **Care team handoff:** Routes anything outside protocol to a clinician or coordinator, with the full conversation and any extracted fields attached. ## Controls - **Detectors before the model:** PII and prompt injection are classified on incoming text before any reasoning model sees it, and the verdict is recorded on every run. - **Private endpoint, scoped access:** The deployed agent sits behind a private endpoint, reachable only with a scoped, expiring access key. - **Dedicated project per pipeline:** Intake, connections and knowledge bases live in a project scoped to only what that pipeline needs, nothing shared. - **Structured output only:** Extracted data is forced into a fixed schema, so a downstream system receives structured fields, never free text. - **Full audit trail:** Every run is traced and replayable, including each detector's verdict, inspectable in the UI or exported as JSON. ## Deployment Self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or run on Dynamiq Cloud behind a private endpoint. Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. ## Questions and answers ### Does Dynamiq support HIPAA workloads? Yes. Dynamiq supports HIPAA workloads and signs a business associate agreement (BAA) with healthcare customers. ### Does the platform redact patient data automatically? No. Detectors classify and flag PII and prompt injection; they do not redact or rewrite the text. Routing a flagged document to a person is the control. ### Where does patient text travel? Only to the model and detector providers you connect, under the connections you configure, and it is recorded in the run's trace like any other input and output. ### Can this run behind our own private endpoint? Yes. Deploy behind a private endpoint with a scoped access key, or self-host the whole platform inside your own environment. ### Can a voice agent answer patient calls? Yes. Voice Agents answer inbound calls on your existing numbers for scheduling and intake questions, and transfer to staff for anything clinical or urgent. Dynamiq does not place outbound calls. ### Who reviews what the agent extracts? An intake coordinator or clinician, for anything flagged by a detector or falling outside the expected schema. Extraction does not skip review on its own. --- # Customer service URL: https://www.getdynamiq.ai/use-cases/customer-service A customer service agent answers inbound phone calls and chat messages, looks up the customer's account, order or policy in your systems, and resolves routine requests inside the rules you set. Phone lines are answered by Voice Agents on the numbers you already own, with every call recorded, transcribed and measured turn by turn. When a request needs judgment, chats hand off to a person with the transcript and calls transfer to your team, and every version is tested against simulated conversations before it ships. ## The problem Support and care teams answer the same requests all day, on the phone and in chat: balance checks, order status, password resets, plan changes, claim updates. Phone queues spike at peak hours, menu-based IVR sends callers in circles, answers vary from one person to the next, and a customer who gets transferred often has to repeat everything. A rules engine alone cannot hold a natural conversation, and a model alone cannot be trusted to act on an account without a rule to check first. ## How the agent works 1. **Call or chat opens**: A customer calls a number you already own and a voice agent answers over SIP, or opens chat on your site or app. The agent starts from its instructions and decision tree. 2. **Understand and verify**: The agent works out what the customer needs from each turn, transcribed live on calls, asks a clarifying question when the request is unclear, and verifies the customer against your records where your policy requires it. 3. **Look up and act**: It looks up the account, order or policy through HTTP tools or MCP servers and answers from the live record, or performs an allowed action, such as resetting a password or updating a plan. 4. **Hand off with context** (human review): Disputes, complaints, anything outside its rules, or a customer who asks for a person go to a person: a phone transfer on calls, a handoff to your support team in chat. The transcript stays in the trace. 5. **Record and measure**: Calls can be recorded with their transcript and per-turn latency, and every chat is traced, so a supervisor can review exactly what the agent said and did. 6. **Test before it ships**: Each new version runs against simulated callers and a dataset of real conversations, scored by an LLM judge, before it reaches customers. ## Controls - **Guardrails on chat input:** Chat messages are screened for PII and prompt injection before they reach the model, and flagged messages route to a refusal or a person. - **Escalation as a gate:** Disputes, account changes and anything the agent is unsure of are routed to a person rather than resolved automatically. - **Inbound only:** Voice Agents answer inbound calls on numbers you own. There is no outbound dialer, so every conversation starts with the customer. - **Scoped access:** Access keys can be scoped to one project and set to expire, so a leaked key exposes only that deployment. - **Full audit trail:** Every run is traced and replayable, with the recording or transcript, every lookup and every tool call attached. ## Systems it connects to Phone numbers you own, over SIP (Twilio, Telnyx or your own carrier), Web and app chat, and browser voice over WebRTC, CRM (Salesforce, HubSpot), Helpdesk and ticketing (Zendesk, Jira Service Management), Slack and Microsoft Teams for escalation, Account and billing systems, over API, Policy and product knowledge (Confluence, Notion, Google Drive, SharePoint), MCP servers for internal tools ## What it moves - **Average handle time:** How long a routine request takes from open to resolution. - **First-contact resolution:** The share of requests the agent resolves without a transfer or callback. - **Escalation rate:** How often a conversation needs a person, and why. - **Customer satisfaction:** Post-conversation ratings, tracked by channel and request type. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment. Phone numbers stay with your carrier and connect over SIP from Twilio, Telnyx or your own trunk, and chat deploys as a widget or through the Runs API. Our engineers configure the decision tree, tools and transfer destinations with your team and run simulations before the agent takes real conversations. ## Questions and answers ### Can an AI agent answer our customer service calls and chats? Yes. Voice Agents answer inbound calls on the numbers you already own, and a deployed App or AI Coworker answers chat. Both share the same knowledge base and backend tools. ### What happens when the agent cannot help? It hands the conversation to a person through a phone transfer on calls, or a paused run a support engineer answers in chat, and the transcript stays in the trace. ### Can it call our own systems? Yes, through HTTP tools, MCP servers or direct API calls, so it can look up an account or perform an action rather than just answer questions. ### Does Dynamiq support outbound calling? No. Voice Agents handle inbound calls only. ### How is the agent tested before customers use it? Voice agents run against simulated callers scored by an LLM judge, and every version is scored against a dataset of real conversations before it ships, so a change is measured rather than guessed. ### How fast does the voice agent respond? Every call shows per-turn latency by stage, transcription, end-of-turn detection, model time to first token and speech synthesis, so you can tune the slowest stage. Speech-to-text models with native end-of-turn detection remove a stage entirely. --- # Contact center voice agents URL: https://www.getdynamiq.ai/use-cases/contact-center A contact center voice agent answers inbound calls on the numbers you already own, looks up the caller's account in your systems, and resolves routine requests inside the rules you set. When a call needs a person, it transfers with the handoff announced, and calls can be recorded with their transcript and per-turn latency. Voice Agents answer inbound calls only, and each agent is tested against simulated callers before it takes a real one. ## The problem Contact centers answer the same calls all day: a declined card, a higher bill, a rescheduled appointment, the status of a claim. Menu-based IVR makes callers press through options before they reach anyone, queues spike at peak hours, and a caller who gets transferred often repeats the whole story. Staffing to the peak is expensive, and a voice bot that cannot look anything up or act on an account only moves the queue. The work needs an agent that holds a natural conversation, uses your systems within your rules, and knows when to hand the caller to a person. ## How the agent works 1. **Call connects**: A caller dials a number you already own and your carrier routes the call over SIP to the voice agent. Web and app callers connect over WebRTC, and your own telephony stack can stream audio over a WebSocket. 2. **Understand the request**: The agent transcribes each turn as the caller speaks, detects when they have finished, and works out what they need, asking a clarifying question when the request is unclear. 3. **Verify and look up**: Where your policy requires it, the agent verifies the caller against your records, then looks up the account, booking or case through HTTP tools or MCP servers. 4. **Resolve within your rules**: It answers from the live record or completes an action your rules allow, such as moving an appointment or resending a statement, and confirms it with the caller. 5. **Transfer to a person** (human review): Disputes, complaints, anything outside its rules, or a caller who asks for a person go to a labeled transfer destination, with the handoff announced to the caller. 6. **Review and test** (human review): Supervisors review recordings, transcripts and per-turn latency, turn a call that went wrong into a simulation scenario, and check simulation scores before a change reaches callers. ## Controls - **Inbound only:** Voice Agents answer calls that reach them on numbers, web and app sessions you own. There is no outbound dialer to govern. - **Transfers as a gate:** Disputes, account changes outside policy and anything the agent is unsure of go to a person, set up as labeled transfer destinations. - **Simulations before go-live:** Scripted callers with a persona, a goal and an expected outcome run against the deployed agent, scored by an LLM judge, before real callers reach it. - **Full call record:** Every call keeps its recording, transcript, tool calls and per-turn latency, so a disputed answer can be checked against what the caller heard. - **Scoped, expiring access:** Access keys can be scoped to one project and set to expire, so a leaked key exposes only that deployment. ## Systems it connects to Phone numbers you own, over SIP (Twilio, Telnyx or your own carrier), Web and mobile callers over WebRTC, Your telephony backend over WebSocket, CRM (Salesforce, HubSpot, Microsoft Dynamics 365), Helpdesk and case systems (Zendesk, ServiceNow, Jira Service Management), Billing, booking and account systems, over API, MCP servers for internal tools, Policy and product knowledge bases ## What it moves - **Calls resolved without a transfer:** The share of calls the agent resolves end to end, by call reason. - **Time to answer:** How long callers wait to be answered, especially at peak hours. - **Transfer quality:** Whether transferred calls reach the right person with enough context that the caller does not repeat themselves. - **Response latency:** How quickly the agent replies on each turn, broken down by stage, so you know which stage to tune. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment, close to your telephony and account systems. Your numbers stay with your carrier; you point its SIP trunk at the agent. Our engineers set up the call flows, tools and transfer destinations with your contact center team and run simulations before the agent answers real calls. ## Questions and answers ### Can an AI agent answer our contact center calls? Yes. Point your carrier's SIP trunk at a deployed voice agent and it answers calls on the numbers you already own, looks up the caller's account through your systems, and transfers to a person when a call needs one. ### Does Dynamiq make outbound calls? No. Voice Agents answer inbound calls only, over the phone, the web or your own backend. There is no outbound dialer. ### Do we have to move our phone numbers? No. Dynamiq does not sell phone numbers. Your numbers stay with your carrier, such as Twilio, Telnyx or your own, and its trunk points at the agent's SIP address. ### Can it replace our IVR menu? It can answer in place of a menu: callers say what they need in their own words, and the agent resolves it or transfers to the right destination. A common first step is a handful of high-volume call reasons, with everything else routed to your existing queues. ### How is it tested before it takes real calls? With simulations: scripted callers with a persona, a goal and an expected outcome run against the deployed agent, and an LLM judge scores pass rate, accuracy and experience. Voice-mode runs also report real latency percentiles. ### Which languages can it handle? Speech-to-text, the language model and the voice are chosen per agent from 15 voice model vendors, so the languages an agent handles depend on the models you pick for each deployment. --- # Document review and underwriting URL: https://www.getdynamiq.ai/use-cases/document-review A document review agent reads loan files, insurance underwriting submissions and contracts, extracts the fields your policy needs, and checks them against your rules. Clean files move to the next step on their own; anything ambiguous, out of range or incomplete goes to a reviewer with the extracted fields and the source page attached. Built for lending, underwriting and contract operations teams that process high volumes of similar documents. ## The problem Lending, underwriting and contract teams read the same kinds of documents by hand every week: application packages, bank statements, broker submissions, signed contracts. The work is high volume and rules-heavy, which makes it exactly the kind of task where a person's attention drifts and a rule gets applied inconsistently from file to file. Outsourcing the reading does not remove the risk. What is needed is software that applies every rule the same way, every time, and still asks a person before anything gets approved, declined or priced. ## How the agent works 1. **Document arrives**: A file enters through an upload, an inbox integration or a case management system, as a PDF, scan or image. 2. **Parse to text**: Dynamiq converts the document to text with a vision-capable model, handling scans and multi-page files the same way as native PDFs. 3. **Screen before extraction**: PII and prompt injection detectors run on the extracted text before any reasoning model sees it, and flagged documents route straight to manual review. 4. **Extract structured fields**: The clean text is extracted into a fixed schema (applicant, amounts, dates, clauses) so downstream systems receive structured data, never prose. 5. **Check against policy**: The agent checks the extracted fields against your credit, underwriting or contract policy, grounded in a knowledge base of your own rules. 6. **Exceptions to a reviewer** (human review): Anything ambiguous, out of policy range or incomplete is flagged to a reviewer with the extracted fields and the source page, instead of being approved automatically. ## Controls - **Rules your experts own:** Credit policy checks run as decision tables and rules your underwriters can read and edit, with saved test cases that show what each rule covers. - **Detectors before the model:** PII and prompt injection are classified on the extracted text before it reaches the reasoning model, and the verdict is recorded on every run. - **Structured output only:** The agent's answer is forced into a fixed JSON schema, so a downstream system never has to parse free text. - **Reviewer sign-off on exceptions:** Files outside policy range or missing required fields pause for a person, not an automatic approval or decline. - **Project-level scoping:** The workflow, connections and knowledge base live in a private project, open only to the people who build or operate it. - **Full audit trail:** Every run is traced and replayable, including which page a field was extracted from and which policy passage the agent cited. ## Systems it connects to Loan origination and underwriting systems, over API, Contract and document storage (SharePoint, Box, Google Drive, Dropbox), E-signature (DocuSign), Case and workflow management (Jira, ServiceNow), CRM (Salesforce), Credit, underwriting and contract policy knowledge base, Document parse and extract through the AI Gateway, Core record systems, over API ## What it moves - **Reviewer time per file:** How long a person spends on a file that reaches manual review. - **Exception rate:** The share of files that need a person, by reason. - **Cycle time:** Time from intake to a decision, across the full file volume. - **Policy consistency:** Whether the same rule was applied the same way across reviewers and shifts. ## Deployment Deploy on Dynamiq Cloud or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment, so extracted documents never have to leave infrastructure you control. Our engineers build the extraction schema and policy checks with your lending, underwriting or legal team before any file skips manual review. ## Questions and answers ### Does the agent approve or decline files on its own? No. It extracts and checks fields against your rules, and anything ambiguous or out of range goes to a reviewer. Approval and decline stay a human decision. ### What document types does it handle? PDFs, scanned images, Word files and other common formats, converted to text with a vision-capable model before extraction. ### Can it review insurance underwriting submissions? Yes. It extracts the risk details from a submission and its supporting documents into your schema, checks them against your appetite and guidelines, and flags exceptions to an underwriter. ### Can it screen for PII before a model reads the document? Yes. PII and prompt injection detectors run on the extracted text before any reasoning model sees it, and the platform detects and flags rather than rewriting the text. ### How do you handle documents from different states or jurisdictions? The extraction schema and policy knowledge base are yours to define, so rules can vary by product, region or jurisdiction within the same workflow. ### What about insurance claims? Claims run as their own flow, from first notice of loss by phone or chat through document review to adjuster sign-off, described on the insurance claims processing page. --- # Insurance claims processing URL: https://www.getdynamiq.ai/use-cases/claims-processing A claims agent takes first notice of loss on an inbound call or in chat, opens the claim in your system, and reads the estimates, invoices and reports that follow. It extracts the fields your process needs, checks them against the policy and your claims guidelines, flags gaps and inconsistencies, and prepares the file for an adjuster, who decides coverage and approves every payment. Built for auto, property and specialty claims teams handling high volumes of similar claims. ## The problem Claims teams handle a flood of routine work around every decision: taking the first report, chasing documents, re-keying estimates and invoices, checking the policy for limits and exclusions, and answering the same status question from the claimant again and again. Adjusters spend more time assembling files than judging them, claimants wait on hold for updates, and handling that varies from one adjuster to the next shows up in complaints and leakage. Automation that pays or denies claims on its own is not acceptable to regulators or customers. Preparing the claim so an adjuster can decide faster is. ## How the agent works 1. **First notice of loss**: A policyholder reports the loss on an inbound call answered by a voice agent, or in chat. The agent verifies the policy, captures what happened and opens the claim in your claims system. 2. **Documents arrive**: Estimates, invoices, police and medical reports arrive by upload or email and are converted to text with a vision-capable model, scans and multi-page files included. 3. **Screen and extract**: PII and prompt injection detectors flag the text before any reasoning model reads it, then dates, amounts, parties and line items are extracted into a fixed schema with the source page for each field. 4. **Check coverage and guidelines**: The agent checks the claim against the policy's limits, deductibles and exclusions and your claims guidelines, grounded in a knowledge base, and flags gaps, duplicates and inconsistencies. 5. **Adjuster decision** (human review): The prepared file, with extracted fields, source pages and flags, goes to an adjuster, who decides coverage and approves, edits or rejects any payment before it is issued. 6. **Status updates**: Claimants ask for status by phone or chat and the agent answers from the live claim record, handing off to the adjuster when a question needs one. ## Controls - **Rules your experts own:** Coverage, limits and fraud checks run as rules your claims leads can read and edit, with one finding per rule and saved test cases. - **An adjuster decides every claim:** The agent prepares and recommends. Coverage decisions, denials and payments always wait for a person's approval. - **Detectors before the model:** PII and prompt injection are flagged on claim documents before a reasoning model reads them, and the verdict is recorded on every run. - **Source page on every field:** Each extracted amount and date links to the page it came from, so an adjuster can check a figure in one step. - **Guidelines applied the same way:** Coverage checks run against your own guidelines in a knowledge base, so the same rule is applied the same way on every claim. - **Full audit trail:** Every run is traced and replayable, including the documents read, the guideline cited and the adjuster's decision. ## Systems it connects to Claims administration and policy systems, over API, Phone lines over SIP (Twilio, Telnyx or your own carrier), Web and app chat, Email inboxes and uploads (Gmail, Microsoft Outlook, SharePoint), CRM (Salesforce, Microsoft Dynamics 365), Payment systems, over API, behind an approval gate, Claims guidelines and policy wordings as a knowledge base, Slack and Microsoft Teams for adjuster handoffs ## What it moves - **Claim cycle time:** Time from first notice to a decision, by line of business. - **Adjuster time per claim:** How long an adjuster spends on a prepared file compared with assembling it. - **Status contacts handled:** How many status calls and chats are answered without pulling in an adjuster. - **Handling consistency:** Whether the same guideline is applied the same way across adjusters and offices. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment next to your claims system. For claims that carry medical or personal data, Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Our engineers build the intake flow, extraction schema and coverage checks with your claims team before the agent touches a live claim. ## Questions and answers ### Does the agent approve or deny claims? No. It takes the first notice, prepares the file and flags issues. Coverage decisions, denials and payments always need an adjuster's approval. ### Can it take first notice of loss by phone? Yes. A voice agent answers inbound claim calls on your existing numbers, captures the loss and opens the claim, and transfers to a person when the caller needs one. Dynamiq does not place outbound calls. ### Can it assess damage from photos? Damage appraisal stays with your adjusters and the estimating tools you already use. The agent reads the documents around a claim, such as estimates, invoices and reports, and brings them into the file. ### Which lines of business fit first? High-volume lines with repeatable documents, such as auto, property and travel claims. Complex or litigated claims stay with adjusters, with the agent preparing the documents for them. ### How do you handle medical and personal data in claims? Detectors flag PII and prompt injection before a model reads claim text, the deployment can stay inside your own environment, and Dynamiq signs a BAA for HIPAA workloads and a DPA under GDPR. Detectors flag; they do not redact. --- # Compliance and financial crime URL: https://www.getdynamiq.ai/use-cases/compliance-financial-crime A compliance agent triages KYC and AML alerts against your policies, tracks regulatory changes across the sources you follow, and drafts the writeup an analyst reviews. Nothing is closed or filed without a person's sign-off, and every step, from the data it queried to the policy it cited, is recorded in a trace your auditors can follow. Built for financial crime, compliance and risk teams working under strict recordkeeping requirements. ## The problem Financial crime and compliance teams triage a steady stream of alerts from transaction monitoring and screening systems, most of which turn out to be false positives, and a smaller number of which need real investigation. Analysts spend more time gathering the same supporting data for every alert than they spend actually deciding on it, and regulatory bulletins that should change a policy can sit unread for weeks. A team also has to reconstruct, months later, exactly why an alert was closed the way it was. ## How the agent works 1. **Alert arrives**: A flagged transaction or screening hit enters from your monitoring system, on a schedule or as an event. 2. **Gather supporting data**: The agent queries transaction history and case systems with read-only, per-analyst credentials, and retrieves the relevant policy passages from your compliance knowledge base. 3. **Draft the assessment**: The agent drafts a writeup: what was flagged, what it found, and which policy section applies, citing its sources. 4. **Analyst sign-off** (human review): The draft pauses for an analyst to approve, edit or reject before anything is closed, filed or reported, with reasons captured if it is sent back. 5. **File and report**: Approved assessments are filed to your case system, and recurring reporting is compiled from the same traced runs. ## Controls - **Per-analyst credentials:** Each analyst's queries run under their own database role, so access follows the same grants and row-level rules your database already enforces. - **Runtime scoping the model never sees:** Values such as a desk or entity identifier come from the caller, not the model, so a query cannot be steered by anything in the alert text. - **Sign-off before anything closes:** No alert is closed, filed or reported without a person's approval, and a rejection sends the draft back with feedback attached. - **Builder and operator separation:** The workflow and connections live in a private project; analysts and approvers only ever touch the deployed app, never the underlying workflow. - **Full audit trail:** Every run is traced and replayable, exportable in bulk for periodic compliance archiving. ## Systems it connects to Transaction monitoring and screening systems, over API, Case management (Jira, ServiceNow-style ticketing), Sanctions and watchlist data feeds, over API, Regulatory bulletins and policy sources (websites, SharePoint, Confluence), Slack and Microsoft Teams for analyst sign-off, Reporting databases (PostgreSQL, ClickHouse), over SQL ## What it moves - **Alert triage time:** How long it takes to move an alert from flagged to a drafted assessment. - **Analyst review time:** How long an analyst spends reviewing a drafted assessment before sign-off. - **False-positive share:** How much of the queue turns out not to need escalation. - **Regulatory change lag:** Time from a bulletin publishing to your policy knowledge base reflecting it. ## Deployment Self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment, or run fully isolated with our engineers when the workload cannot reach the public internet. Every credential is scoped to a project and every run is traced and replayable for your auditors. ## Questions and answers ### Does the agent close alerts on its own? No. It drafts the assessment and gathers supporting data; an analyst approves, edits or rejects before anything is closed or filed. ### How does it avoid widening database access? Each analyst's queries run under their own database credentials at run time, so the database's own grants and row-level rules decide what comes back, not the model. ### Can we reconstruct why an alert was closed? Yes. Every run is traced and replayable, including the data queried, the policy cited, and the analyst's decision, exportable in bulk for compliance archiving. ### How does regulatory change monitoring work? The agent watches the sources you point it at, such as regulator bulletins and internal policy pages, and flags changes for a person to fold into the policy knowledge base. ### Can this run fully isolated from the internet? Yes, our engineers deliver fully isolated or air-gapped deployments for environments that require them. --- # KYC and customer onboarding URL: https://www.getdynamiq.ai/use-cases/kyc-onboarding A KYC onboarding agent reads the documents a new customer submits, extracts the fields your policy needs, pulls results from the screening, registry and identity verification services you already use, and applies your risk rules to propose a rating. An analyst reviews the prepared file and approves, edits or rejects before any account is opened, and every step is traced for examiners. Built for retail, business and private banking onboarding teams, and for periodic KYC reviews that run on a schedule. ## The problem Onboarding teams spend most of each case collecting and re-keying: identity documents, proof of address, company registry extracts, ownership details, source-of-funds letters. Screening hits have to be looked up and cleared one by one, periodic reviews pile up on the calendar, and every file has to show an examiner why the customer was accepted at the rating they got. Customers wait for an account while analysts do work a system could prepare, and the judgment calls that actually need an analyst get the least time. ## How the agent works 1. **Application arrives**: A new customer's application and documents arrive from your onboarding portal, a relationship manager's upload or an email inbox, as PDFs, scans or images. 2. **Read and extract**: Documents are converted to text with a vision-capable model and extracted into a fixed schema, such as names, dates of birth, addresses, registration details, directors and owners, with the source page for every field. 3. **Gather screening results**: Prompt injection detection runs on document text before any model reasons over it, and the agent calls your sanctions screening, company registry and identity verification providers over API for their results. 4. **Apply your risk rules**: Rules nodes apply your policy deterministically, by jurisdiction, product, ownership and screening outcome, to propose a risk rating and list what is missing or inconsistent. 5. **Analyst review** (human review): The prepared file, with extracted fields, screening results, the proposed rating and any gaps, pauses for an analyst to approve, edit or reject, or to request missing documents. 6. **Open and schedule**: Once approved, the agent sends the account opening request to your core banking or CRM system over API, files the case with its trace, and records the next review date for the scheduled periodic review to pick up. ## Controls - **Rules your experts own:** Risk scoring and document requirements run as decision tables your compliance team can read and edit, with effective dates for each policy change. - **An analyst decides every account:** The agent prepares the file and proposes a rating. Opening, declining or escalating an account always needs a person's approval. - **Rules you can read:** Risk rules run as deterministic nodes, not a model's judgment, so the same file gets the same outcome and a reviewer can see which rule fired. - **Source page on every field:** Each extracted field links to the document and page it came from, so an analyst checks the evidence, not a summary. - **No wider access than the analyst:** Onboarding runs in a private project, and database access can resolve to each analyst's own credentials, so the agent never sees more than the analyst could. - **Examiner-ready trail:** Every run is traced and replayable, from the documents read to the screening results, the rule that set the rating and the analyst's decision. ## Systems it connects to Onboarding portals and document inboxes, Sanctions, PEP and adverse media screening, over API, Company registries and identity verification providers, over API, Core banking and CRM (Salesforce, Microsoft Dynamics 365), over API, Case management (Jira, ServiceNow), Document storage (SharePoint, Box, Google Drive), Slack and Microsoft Teams for analyst review, Your KYC policy and risk methodology as a knowledge base ## What it moves - **Time to onboard:** Time from a complete application to an opened account, by customer segment. - **Analyst time per file:** How long an analyst spends on a prepared file compared with assembling it by hand. - **Rework rate:** How often files come back for missing documents or corrections. - **Periodic review backlog:** How many reviews are overdue, and whether scheduled reviews start on time. ## Deployment Self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment, or run on Dynamiq Cloud, so customer documents stay in the infrastructure you choose. Our engineers build the extraction schema, risk rules and review step with your onboarding and compliance teams, and test them on past files before a live application reaches the agent. ## Questions and answers ### Does the agent approve or open accounts on its own? No. It prepares the file and proposes a risk rating; an analyst approves, edits or rejects before any account is opened. ### Does Dynamiq verify identities or run sanctions screening itself? No. The agent calls the identity verification and screening providers you already use over API and brings their results into the file, with the source attached. ### Can it onboard business customers and beneficial owners? Yes. The extraction schema is yours to define, so it can pull directors, shareholders and ownership details from registry extracts and corporate documents, and your rules decide what needs a closer look. ### How does it handle periodic KYC reviews? A schedule trigger starts a review when it falls due, and the agent prepares the refreshed file for an analyst the same way it does for a new customer. ### Can examiners see why a customer was accepted? Yes. Every run is traced and replayable, including each document read, each screening result, the rule that set the rating and the analyst's decision. --- # Research and knowledge assistants URL: https://www.getdynamiq.ai/use-cases/research-knowledge A research and knowledge assistant answers questions from your own documents and systems, retrieving only what the asking person is allowed to see and citing the source for every claim. Knowledge graphs let it follow relationships across documents, not just match keywords, for questions that need several steps to answer. Built for analysts, researchers and internal teams who need answers they can check, not just a summary. ## The problem Analysts and internal teams lose hours a week hunting across five systems and a colleague's inbox for something the organization already wrote down. A model answering from memory alone will guess, and guessing is worse than not answering in a regulated business. Teams need an assistant that answers only from retrieved sources, cites what it used, follows relationships between documents when a question needs more than one hop, and never shows a person a document their own permissions would not let them open. ## How the agent works 1. **Documents sync in**: A knowledge base syncs from the systems of record you point it at, kept current on a schedule rather than copied once and forgotten. 2. **Question arrives**: An analyst or an internal user asks a question in plain language, through a chat surface or your own application. 3. **Grounded retrieval**: The agent retrieves the passages and, where a knowledge graph is enabled, the related entities that answer the question, filtered to what that user can see. 4. **Cited answer**: The agent answers from what it retrieved, with an inline citation back to the source document for every claim. 5. **Analyst checks the source** (human review): For decisions that matter, the analyst opens the cited source before acting on the answer, and can flag a wrong or thin answer back into the review queue. ## Controls - **Permission-aware retrieval:** Retrieval is scoped to what the asking person's own account can see, either through a shared connection limited to public content or a per-user authorization. - **Per-team isolation:** Each team's knowledge base, connections and workflow live in their own private project, not one shared pool. - **Citations on every answer:** Every answer links back to the retrieved passage, so a person can verify a claim instead of trusting it. - **Full audit trail:** Every retrieval and every answer is recorded in a trace, so what the assistant looked up to produce an answer is always answerable. ## Systems it connects to Google Drive, SharePoint, OneDrive, Box and Dropbox, Notion and Confluence, Website and public filings crawls, Neo4j for knowledge graphs, Slack and Microsoft Teams for answers in the flow of work, Internal wikis and policy repositories, Databases (PostgreSQL, MySQL), over SQL ## What it moves - **Time to answer:** How long it takes a person to get a sourced answer instead of searching several systems themselves. - **Retrieval precision:** Whether the passages the assistant retrieves are the ones that actually answer the question. - **Answer coverage:** The share of real questions the knowledge base can answer at all. - **Citation rate:** How often an answer includes a source a person can check. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment so source documents never leave infrastructure you control. Our engineers scope retrieval permissions and the knowledge graph schema with your team before rollout. ## Questions and answers ### Does the assistant answer from the model's general knowledge? No. It answers only from what it retrieves and cites the source, and says so when the documents do not cover a question. ### How does it respect document permissions? Retrieval can run under a shared connection scoped to content everyone may see, or under each user's own authorization when access needs to match their individual permissions. ### What is a knowledge graph used for here? It stores documents as entities and relationships, not just chunks, so the assistant can follow a connection across documents for questions that need more than one retrieval step. ### Does content stay current? Yes. Connected sources sync in the background rather than being ingested once, so the assistant answers from current documents. ### Can different teams have different assistants? Yes. Each team's documents, connections and assistant can live in their own project, isolated from every other team's. --- # Back-office and finance operations URL: https://www.getdynamiq.ai/use-cases/back-office A back-office agent works the queues behind every bank and finance team: reconciling accounts and suspense items, investigating payment exceptions, preparing dispute and chargeback cases, and drafting management and regulatory reports. It runs on a schedule or on demand, queries each system under scoped or per-user credentials, and nothing that leaves the building, a posting, a payment, a filing or a report, goes out without a person approving the exact output. Built for operations, finance and shared-services teams that run the same cycle every day or period. ## The problem Bank operations and finance teams run the same cycle every day and every period: pull data from the core system, the card and payment platforms and the general ledger, reconcile it, investigate what does not match, and prepare the correction or the report. Every system has its own login and export format, breaks live in spreadsheets, and a payment repair or a dispute waits while someone gathers the evidence by hand. Analysts spend more time moving data between systems than deciding what it means, and a recurring report is only as reliable as whoever remembered to run it. ## How the agent works 1. **Run starts**: A schedule fires for the daily or period-end cycle, an event such as a failed payment or a new dispute arrives, or an analyst starts a run through an API call or a chat message. 2. **Gather and reconcile**: The agent queries the systems it needs, core banking, card and payment platforms, the general ledger or your warehouse, and reconciles the figures against your rules or the prior period. 3. **Investigate exceptions**: For each break, failed payment or dispute, it gathers the transaction history, messages and account notes, and drafts the likely cause and the proposed correction or case. 4. **Approval before it commits** (human review): Nothing posts, pays, files or sends until a reviewer approves the exact output, with the ability to edit allowed fields before approving. 5. **Deliver and file**: Approved corrections, cases and reports go to their destination systems, and the run is filed as a trace for the record. ## Controls - **Rules your experts own:** Matching tolerances, approval limits and posting rules run as decision tables and rule lists your finance team can read and edit, with saved test cases. - **Approval gate on every output:** The step that posts, pays, files or sends carries an approval gate, so a run pauses until a person signs off on the exact payload. - **Scoped credentials:** Each connected system uses its own credential, and rotating one updates every step that uses it, so no single key spans every system the agent touches. - **Scheduled and on-demand paths deploy separately:** Recurring runs use a shared service credential; analyst-triggered runs can resolve to that analyst's own credentials instead, so access never widens. - **Full audit trail:** Every run is traced and replayable, including every figure queried and every approval decision. ## Systems it connects to Core banking systems, over API or SQL, Card, payment and dispute platforms, over API, General ledger and ERP, over API, Databases and warehouses (PostgreSQL, MySQL, ClickHouse, Trino), Slack and Microsoft Teams for approvals, Spreadsheets and reporting tools (Google Sheets), Ticketing and case systems (Jira, ServiceNow), Scheduled and event triggers across 2,100+ app integrations ## What it moves - **Reconciliation cycle time:** How long a daily or period-end reconciliation takes end to end. - **Exception aging:** How long breaks, payment exceptions and disputes wait before they are resolved. - **Approval turnaround:** How long an output waits for a reviewer's sign-off. - **On-time reporting:** Whether a recurring report or filing goes out on schedule without manual intervention. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment alongside the systems it reconciles. Our engineers wire the schedule, the approval gate and the first set of connections with your operations or finance team before handover. ## Questions and answers ### Which back-office processes fit an agent first? Repetitive, rules-heavy queues with a clear approval point: account and suspense reconciliations, payment exceptions and repairs, dispute and chargeback case preparation, and recurring management or regulatory reports. ### Can this run on a schedule? Yes. A schedule trigger runs the workflow on a cron expression, an event can start it, and an analyst can also start the same workflow on demand. ### Does anything ship without approval? No. The step that posts, pays, files or sends carries an approval gate, so it pauses until a reviewer signs off on the exact output. ### Can it prepare disputes and chargebacks? It gathers the transaction history and evidence, drafts the case for your disputes team and files it once approved. The decision stays with your team. ### How many systems can it reach? Well beyond your core stack: 2,100+ app integrations, plus MCP servers and direct API calls for anything not in the catalog. ### What happens if a run needs an analyst's own access? Deploy a second version of the same workflow where the connection is flagged as a per-user requirement, so an analyst's run resolves to their own credentials instead of a shared one. --- # Accounts payable and invoices URL: https://www.getdynamiq.ai/use-cases/accounts-payable An accounts payable agent reads incoming invoices, extracts the header and line items, matches them to purchase orders and goods receipts in your ERP, and checks for duplicates and changes to supplier bank details. Matched invoices go to the approver your rules name with the evidence attached, exceptions go to the AP team with the reason, and nothing posts or pays without a person's approval. It also answers supplier status questions and prepares the payment run on your calendar. ## The problem AP teams still key invoices by hand, chase approvers by email, and lose hours to matches that fail for small reasons: a changed unit price, a partial delivery, a missing PO number. Suppliers call to ask when they will be paid. Duplicate invoices and fraudulent bank-detail changes slip through when a payment run is rushed, and month-end accruals depend on knowing which invoices are still stuck. Most of the volume is routine, but it all arrives through the same inbox, so the exceptions that need judgment wait behind everything else. ## How the agent works 1. **Invoice arrives**: Invoices arrive in an AP inbox, a supplier portal or an upload, as PDFs, scans or images, and are converted to text with a vision-capable model. 2. **Extract to your schema**: The supplier, invoice number, dates, currency, amounts, tax and line items are extracted into a fixed schema, with the source page for each value. 3. **Match and check**: The agent queries your ERP for the purchase order and goods receipt, applies the match and tolerances your policy sets, and checks for duplicates and bank-detail changes against your supplier records. 4. **Route exceptions** (human review): Price or quantity variances, missing purchase orders and suspected duplicates go to the AP team with the reason and the evidence, instead of sitting in a queue. 5. **Approve before posting** (human review): Matched invoices pause for the approver your rules name, who can approve, edit allowed fields such as the cost center, or reject before the invoice posts to the ERP. 6. **Prepare the payment run** (human review): On your payment calendar, the agent prepares the proposed payment run for treasury to approve and drafts the list of open invoices for month-end accruals. ## Controls - **Rules your experts own:** Three-way match tolerances, vendor checks and approval limits run as rules your AP team can read and edit, each change checked against saved test cases. - **Approval before anything posts or pays:** Posting and payment steps carry approval gates, so nothing reaches the ERP or the bank without a person signing off on the exact values. - **Bank-detail changes go to a person:** A change to a supplier's bank details never flows straight into a payment; it routes to a person to verify through a separate channel. - **Separation of duties:** A private project limits who can change the workflow, and approval gates decide who signs off on invoices and payments. - **Evidence on every line:** Each extracted value links to its source page, and each match result shows the purchase order and receipt it used. - **Full audit trail:** Every run is traced and replayable for auditors, including each approval and each field a reviewer edited. ## Systems it connects to ERP and finance systems, over API, Accounting (QuickBooks, Xero), AP inboxes (Gmail, Microsoft Outlook), Document storage (SharePoint, Google Drive, Box), Payment and banking platforms, over API, behind an approval gate, Slack and Microsoft Teams for approvals, Supplier and PO data in databases (PostgreSQL, MySQL, ClickHouse), Chat and email for supplier status questions ## What it moves - **Touchless rate:** The share of invoices that match and reach approval without anyone keying data. - **Invoice cycle time:** Time from receipt to approval, and from approval to payment. - **Exception aging:** How long exceptions wait, by reason. - **Duplicates and bank-detail changes caught:** Issues stopped before a payment run instead of found after it. ## Deployment Deploy on Dynamiq Cloud, or self-host in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes environment next to your ERP. Our engineers build the extraction schema, matching rules and approval steps with your AP and treasury teams, and run them on past invoices before live ones. ## Questions and answers ### Does the agent pay invoices on its own? No. Posting and payment steps carry approval gates, so an approver signs off on the exact values before anything posts to the ERP or goes to the bank. ### Can it do two-way and three-way matching? Yes. The agent looks up the purchase order and goods receipt in your ERP and applies the match rules and tolerances your policy sets, with variances routed to your AP team. ### Which ERPs does it work with? It calls your ERP over API, or reads it over SQL where your team allows it, and connects to accounting tools such as QuickBooks and Xero from a catalog of 2,100+ app integrations. ### How does it handle suspected fraud? It flags suspected duplicates and any change to a supplier's bank details and routes them to a person, rather than deciding on its own. ### Can suppliers ask about payment status? Yes. A chat or email agent answers status questions from the live invoice record and hands anything unusual to the AP team. --- # AI agent governance URL: https://www.getdynamiq.ai/use-cases/ai-agent-governance AI agent governance decides what an agent may see, what it may do, who signs off, and how you prove what happened. On Dynamiq those controls are part of the platform: detectors screen input for PII and prompt injection, validators check output, approval gates pause consequential steps for a person, and every run is traced and replayable. Real-time evals score live traffic and an AI agent surfaces issues from production traces, so oversight continues after launch. ## The problem Most organizations can build an agent. Fewer can answer the questions that come before it goes live: what data can it reach, what can it change, who approved the risky step, and how will anyone know if its answers get worse next month. A policy document does not stop a model from ignoring an instruction, and controls each team builds on its own end up different for every agent. Risk, compliance and security teams need the same enforceable controls on every agent, and evidence they can pull without asking engineers to reconstruct a run. ## How the agent works 1. **Access decided up front**: SSO on Enterprise, organization roles, private projects and scoped access keys decide who can build, deploy and run agents, and credentials can resolve per end user where access must follow each person's own permissions. 2. **Input screened**: Detectors flag PII, prompt injection and content policy violations before the model sees a message, and a Choice node routes flagged input to a refusal or a person. 3. **Output checked**: Validators confirm the output is valid JSON, an allowed value or a matching pattern before anything downstream uses it, and fail the step or retry when it is not. 4. **Approval on consequential steps** (human review): A payment, a database write, an email or a filing pauses for a person to approve the exact payload, edit the fields you allow, or reject with feedback the agent acts on. 5. **Evaluated before release** (human review): A new version is scored against datasets built from real traces and compared with the version in production, and a reviewer decides whether to deploy it or roll back. 6. **Monitored after launch**: Real-time evals score a sample of live traffic, an AI agent reviews production traces for issues and regressions, and every run stays replayable for audit. ## Controls - **Enforced, not prompted:** Detectors, validators and approval gates are workflow nodes the platform runs every time, not instructions a model can talk its way around. - **Detect, never rewrite:** Detectors classify and flag; they do not redact or rewrite text, so what happens to a flagged message is a routing rule you own. - **Least-privilege access:** Private projects, scoped access keys that can expire, and per-user credentials keep an agent's reach no wider than the person or service it acts for. - **Change control:** Every save is a new version. Deploy a specific version, compare it with the last one and roll back in one step. - **Evidence on demand:** Every run is traced and replayable, including detector verdicts and approval decisions, and traces export in bulk as JSON. - **Continuous evaluation:** Real-time evals and AI-found issues keep scoring an agent after it ships, so drift shows up in a chart before it shows up in a complaint. ## Systems it connects to Identity providers over OIDC (Okta, Microsoft Entra ID), Slack and Microsoft Teams for approvals, Jira and ServiceNow for review queues, Models from 29 model providers, the most-used through the AI Gateway, Databases and warehouses, under per-user credentials, Trace exports as JSON for your archive or SIEM, Evaluation datasets built from production traces ## What it moves - **Time to approve a new agent:** How long risk and security review takes when the controls and evidence already exist. - **Flag rate:** How often detectors flag PII, prompt injection or policy violations, by agent and channel. - **Approval turnaround:** How long consequential steps wait for a person, and how often reviewers edit or reject. - **Quality drift:** Changes in real-time eval scores and AI-found issues after a model, prompt or data change. ## Deployment Run governance where the agents run: on Dynamiq Cloud, or self-hosted in your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster. Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Our engineers define the success criteria, evals and approval points with your risk and compliance teams before the first agent ships. ## Questions and answers ### What is AI agent governance? The controls that decide what an agent may see and do, who approves its consequential actions, and how you prove what happened afterward: access control, input and output checks, approvals, audit trails and ongoing evaluation. ### How is governing an agent different from governing a model? A model produces text; an agent calls tools that change records, send messages and move money. Governance has to cover the actions, so approval gates and per-user credentials matter as much as content checks. ### Does Dynamiq redact PII? No. Detectors flag PII and prompt injection so you can route a message to a refusal or a person; they do not redact or rewrite it. ### Can we show an auditor what an agent did? Yes. Every run is traced and replayable node by node, including each detector verdict and approval decision, and traces export as JSON in bulk. ### Does governance stop at launch? No. Real-time evals score a sample of live traffic, and an AI agent reviews production traces and surfaces issues and regressions for a person to check. ### Does this make us compliant with the EU AI Act? It supports record-keeping and human oversight with traces and approval gates, but Dynamiq does not certify compliance. Whether a system is high-risk, and what that requires, depends on your own assessment. --- # Sovereign AI URL: https://www.getdynamiq.ai/use-cases/sovereign-ai A sovereign AI deployment keeps agents, models, data and the record of every decision under your control, on infrastructure you choose and operate. Dynamiq runs the full platform (agents, knowledge, guardrails, evals and traces) inside your own cloud account or data center, serves open models such as Llama, Gemma, Mistral and Qwen inside your environment, so agents can answer without calling an outside model provider. The SDK is open source under Apache-2.0, so the agents you build stay yours. ## The problem Ministries, central banks, defense organizations and regulated firms are being asked to show that their AI meets the bar of any other critical system: data stays in a known jurisdiction, models run where the organization can inspect them, and a record of every decision is kept on infrastructure it controls. Shared AI services make that hard to prove. Prompts and documents travel to a provider's endpoint, traces sit in a vendor's tenant, and agent logic is locked into a proprietary builder that cannot move. Teams end up choosing between useful agents and control over them. ## How the agent works 1. **A signed-in request**: A person signs in through your identity provider with SSO on Enterprise, or a customer channel or scheduled job starts a run, against an agent deployed in your own cloud account or data center. 2. **Screened before the model**: PII and prompt injection detectors screen the input, and a flagged request routes to a refusal or to a person instead of the model. 3. **A self-hosted model answers**: The agent calls an open model served on vLLM inside your environment, so the prompt and the answer stay on your infrastructure. 4. **Tools reach only your systems**: The agent queries your databases, document stores and internal APIs through connections you configure, under each person's own credentials where access must follow their permissions. 5. **Approval before anything leaves** (human review): A payment, a filing or a message to a citizen or customer pauses for a person to approve the exact output, with fields they are allowed to edit. 6. **The record stays with you**: The run is traced inside your deployment, replayable node by node for auditors, and kept according to the retention controls you set. ## Controls - **The whole platform in your perimeter:** Agents, knowledge bases, guardrails, evals and traces run in your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes 1.32+ cluster, or on-premises. - **Models you host:** Open models such as Llama, Gemma, Mistral and Qwen run on vLLM inside your environment, with LoRA adapters fine-tuned on your data, so agents need no outside model provider. - **Your identity and your policies:** SSO through your identity provider on Enterprise, organization roles and retention controls follow your own policies. - **Evidence stays with you:** Every run is traced and replayable inside your deployment, exportable for your auditors without a vendor in the loop. - **No lock-in:** The SDK is open source under Apache-2.0 and runs on the same engine as the platform, and workflows export as YAML, so the agents you build can move with you. - **Isolated when required:** Fully isolated or air-gapped environments are delivered with our engineers, for networks that cannot reach the public internet. ## Systems it connects to Cloud accounts you own on AWS, Azure, GCP and IBM Cloud, Red Hat OpenShift and any Kubernetes 1.32+ cluster, on-premises, Open models on vLLM, with LoRA adapters, Identity providers over OIDC (Okta, Microsoft Entra ID), Internal databases and warehouses, over SQL, Document repositories you host, such as SharePoint, Internal APIs and MCP servers, Your own carrier over SIP for inbound voice ## What it moves - **Calls to outside providers:** Which model and tool calls, if any, leave your environment, visible in every run's trace. - **Time to a first production agent:** How long it takes to go from an empty cluster to an agent your team runs in production. - **Audit response time:** How quickly you can show an auditor or regulator exactly what an agent did on a given run. - **Portability:** Whether an agent can move between clusters, clouds or data centers without a rebuild. ## Deployment Install from the official Helm chart into your own AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-premises. Dynamiq holds SOC 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. Our forward-deployed engineers work inside your environment to ship the first production agent with your team and hand it over. Fully isolated or air-gapped environments are delivered with our engineers. ## Questions and answers ### What is sovereign AI? It is control over the whole AI stack: the infrastructure it runs on, the models it uses, where data and logs are stored, and who operates it. For agents, that means the platform, the models and every trace stay on infrastructure your organization controls, with no dependency you cannot replace. ### Can Dynamiq run with no calls to outside model providers? Yes. Serve open models on vLLM inside your environment and point your agents at them, so prompts and documents stay on your infrastructure. ### Which models can we run ourselves? Open-weight models such as Llama, Gemma, Mistral and Qwen, served on vLLM, with LoRA adapters fine-tuned on your own data. Hosted providers stay optional. ### Can it run fully air-gapped? Yes. Fully isolated or air-gapped environments are delivered with our engineers, on top of the standard self-hosted install. ### Are we locked in to Dynamiq? No. The SDK is open source under Apache-2.0 and runs on the same engine as the platform, and workflows export as YAML, so what you build stays yours. ### Do your engineers work inside our environment? Yes. Forward-deployed engineers work inside your environment with your team: they run a bootcamp, ship the first production agent with you, then hand it over. ### Can we buy through our cloud marketplace? Yes. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace, so it can be procured through a cloud account you already have. --- # How an Asian neo-bank automated customer support URL: https://www.getdynamiq.ai/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai A digital bank in Asia needed to answer a fast-growing volume of support requests without losing accuracy or control. It built a support agent on Dynamiq that answers from its own knowledge base, acts on backend systems, and hands off to a person when a request needs one. ## The challenge The bank serves more than 4 million customers across Asia. It is a digital-first lender whose support queue was growing faster than its team could hire. Every day brought thousands of the same kinds of requests: balance checks, card issues, password resets and policy questions, mixed in with harder cases that genuinely needed a person's judgment. A basic chatbot could not carry that load. The team needed an agent that understood a wide range of requests, stayed grounded in the bank's own policies and product details, could safely act on backend systems such as checking a balance or resetting a password, and left a record a compliance team could review after the fact. ## What we built One of our engineers worked inside the bank's own environment, alongside its team, to design a support agent on Dynamiq: a single agent with tools, memory and a knowledge base, rather than a maze of separate bots handling separate intents. Success criteria and evaluation checks were agreed before the first version was built, so "working" had a definition everyone shared from day one. The agent answers from the bank's internal knowledge bases, calls the bank's own APIs to perform account actions, and escalates to a person with the full conversation attached whenever a request needs a human decision. Because the workflow made it easy to test and compare models from Anthropic and OpenAI side by side, the team picked whichever option answered accurately at the lowest cost, and kept the option open to switch again later without rebuilding the agent. ## How it runs A customer's question arrives from the bank's app or website. The agent checks its knowledge base for the relevant policy or product answer, calls a backend API when the request needs account data or an action such as a password reset, and hands off to a support agent when the request is a dispute, an edge case, or anything outside what it is allowed to resolve alone, carrying the full transcript into that handoff so nothing gets repeated. Every run is traced and replayable, so support leads can see exactly what the agent looked up and said. An evaluation suite scores every candidate change against real conversations before it reaches customers, so the team can adjust prompts or swap models without risking a regression in production. ## Results - An Asian neo-bank automated about 85% of support inquiries and went live in 30 days. - The support agent now saves an estimated $1.5M a year in support costs. - The bank serves more than 4 million customers across Asia. The team is now working to extend automation to a larger share of its inbound support volume, with the same evaluation suite gating every change before it ships. --- # Document search for a B2B accounting platform URL: https://www.getdynamiq.ai/case-studies/smart-document-search-for-a-b2b-accounting-platform-built-in-one-week-with-dynamiq A B2B accounting platform needed natural-language search across contracts, invoices and policies, where keyword search kept coming up empty. One of our engineers built a retrieval pipeline on Dynamiq that answers in plain language with the source documents attached. ## The challenge The client runs a B2B accounting platform that its own customers use to manage contracts, invoices and other financial paperwork, and users kept hitting the same wall: finding one specific document inside a large, unstructured pile of PDFs, Word files, slide decks and images. The platform's existing search was keyword-only, so a question like "which contracts expire next month" or "find invoices sent in March with overdue payments" returned nothing useful. The team wanted search that understood plain language, handled every common document type, and scaled without a pile of new infrastructure to operate. Building that in-house would have meant standing up a machine-learning team and a vector database from scratch, at a cost the roadmap could not absorb. ## What we built One of our engineers worked directly with the client's team to design a retrieval-augmented search pipeline on Dynamiq: documents are converted and embedded on ingestion, stored in Amazon OpenSearch as the vector store, and queried through Anthropic Claude for the answer itself. A user can now ask "show me all NDAs signed last month" or "which contracts have not been renewed" in plain language and get back the right documents with an answer synthesized from them, directly inside the product they already use. One-click connections to Claude and Titan Text Embeddings over Amazon Bedrock meant the team never had to stand up a separate inference pipeline or manage a model deployment by hand. ## How it runs A document enters the pipeline as a PDF, a Word file, a slide deck or an image. Dynamiq converts it to text, embeds it, and writes it into Amazon OpenSearch, which serves both vector similarity and keyword search from the same index. When a user asks a question, Dynamiq's retrieval pipeline pulls the most relevant chunks, hands them to Claude, and returns an answer with the source documents attached, so a user can always check where the answer came from. Model access runs through Amazon Bedrock, so the client's existing AWS relationship and controls extend to the new feature without a separate vendor to manage. Because the pipeline is a workflow rather than custom code, the client's own team can see and adjust every step themselves. ## Results A B2B accounting platform shipped document search in one week with one of our engineers. The feature shipped with no additional infrastructure to operate and no custom machine-learning model to maintain, so the roadmap moved straight to the next problem instead of a long build. The client is now extending the same pipeline to document summarization and structured extraction, such as pulling contract terms and payment details automatically. --- # Dynamiq vs Amazon Bedrock AgentCore URL: https://www.getdynamiq.ai/compare/amazon-bedrock-agentcore Amazon Bedrock AgentCore is AWS's modular, consumption-priced platform for building, deploying and operating agents: Runtime, Gateway, Memory, Identity, Policy, Evaluations and Observability services that work with any framework and any foundation model, all operated by AWS. Dynamiq is a complete platform for chat, voice and workflow agents, with guardrails, evals and observability, that runs on Dynamiq Cloud, any major cloud or on-prem, with forward-deployed engineers. Teams standardized on AWS that assemble their own agents on a framework such as Strands or LangGraph fit AgentCore; teams that need to run outside AWS, want packaged chat and voice agents, or want a deployment team fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Amazon Bedrock AgentCore | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Teams already building on AWS that want managed, serverless services for agents written in any framework, such as Strands, LangGraph, CrewAI or LlamaIndex, with any foundation model. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | AWS only. Runtime runs agents on serverless microVMs, or on Instances: AWS-managed Amazon EC2 capacity in your own AWS account for multi-day sessions and GPU workloads. Available in AWS commercial regions and AWS GovCloud (US-West). | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | No. AgentCore is a set of AWS-managed services, and every compute option it documents runs on AWS. Its Agent Registry, in preview, can catalog agents, MCP servers and skills deployed on AWS, on-prem or another cloud. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | The managed service is not open source. AWS publishes the AgentCore CLI, a Python SDK and an MCP server on GitHub under Apache 2.0; the older starter toolkit is now marked legacy. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Any foundation model in or outside Amazon Bedrock, including OpenAI, Google Gemini, Anthropic Claude, Amazon Nova, Meta Llama and Mistral models, per AWS's documentation; Gateway can also route inference requests across model providers. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | A developer platform rather than packaged agents. Runtime hosts agent code from any framework, including voice agents that stream over WebSocket or WebRTC, and a managed harness runs agents declared as configuration. AWS documents no ready-made chat app or phone-line (PSTN or SIP) product. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Code-first: the AgentCore CLI and SDK, framework support for Strands, LangGraph, CrewAI and AutoGen, protocols including MCP, A2A and AG-UI, and a harness that can export an agent to Strands code. | | Memory | Session memory scoped by user and session, plus long-term memory, where the agent decides when to save and recall durable, user-scoped facts that people can see and correct. Knowledge adds ingestion pipelines, permission-aware retrieval filters and knowledge graphs. | AgentCore Memory, a managed service: short-term memory keeps turn-by-turn context within a session, and long-term memory automatically extracts user preferences, facts and session summaries across sessions. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Policy for Gateway tool calls, written in Cedar or natural language, plus Dogwood for session-aware rules such as requiring that an approval step was recorded earlier in the session; every decision is logged in CloudWatch. Identity works with Amazon Cognito, Okta, Microsoft Entra ID, Auth0 and other providers. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | AgentCore Evaluations, with built-in and custom evaluators run online, on demand, in batch or against datasets, and Observability dashboards in Amazon CloudWatch with OpenTelemetry-compatible traces. An Optimization feature suggests prompt and tool description changes from traces; its insights are in public preview. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | Gateway turns APIs, AWS Lambda functions and existing MCP servers into tools, with one-click integrations for tools such as Salesforce, Slack, Jira, Asana and Zendesk; no total connector count is published. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | HIPAA eligible, and in scope for SOC, PCI DSS and FedRAMP (Class C, formerly Moderate, and Class D, formerly High, in GovCloud) per AWS's compliance pages; AWS's FAQ also lists ISO 27001, HITRUST, CSA STAR, C5 and other programs. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Not publicly documented; AgentCore's product page and FAQ describe no professional services or deployment team. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Consumption-based, with no upfront commitment or minimum fees: per-unit rates for Runtime compute and memory, Gateway invocations, Identity, Memory, Policy, Evaluations and Registry records, and observability billed as Amazon CloudWatch usage. A committed Runtime option is launching by October 2026. | ## Choose Dynamiq when - Runs outside AWS as well as on it: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - Packaged agents on one engine: AI Coworker with its own cloud sandbox, in the browser, Slack, Microsoft Teams and Telegram, Voice Agents for inbound calls, and Agent Builder, instead of services you assemble into a product. - Guardrails, approvals, evals (including real-time evaluations on live deployments) and observability with AI issue discovery are part of the same platform, not separate services to wire together. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Amazon Bedrock AgentCore when - Already standardized on AWS and want agent infrastructure billed, secured and operated like the rest of an AWS estate, including in AWS GovCloud (US-West). - Consumption pricing with no upfront commitment or minimum fees, so cost tracks the usage of each component, such as Runtime, Gateway or Memory. - Freedom to bring any open-source framework and any model to a serverless runtime AWS operates, with multi-day sessions and GPUs on Runtime Instances. ## Migration Agents written in Strands, LangGraph or CrewAI for AgentCore Runtime are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK, starting from the same prompts, tool definitions and model choices. Tools exposed through AgentCore Gateway, including Lambda functions and existing MCP servers, reconnect as Dynamiq MCP servers or integrations, and knowledge sources are re-ingested into Dynamiq Knowledge. Cedar policies attached to Gateway become Dynamiq guardrails and approval steps, and AgentCore evaluators move to Dynamiq evals. ## Questions and answers ### Can Dynamiq run outside AWS? Yes. Dynamiq runs on Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Every compute option Amazon Bedrock AgentCore documents, serverless microVMs or Instances, runs on AWS. ### Does Amazon Bedrock AgentCore include a chat or voice product? Not as a packaged product. AgentCore Runtime can host voice agents that stream audio over WebSocket or WebRTC, but AWS documents no ready-made chat app or phone-line product; teams build the interface themselves. Dynamiq includes AI Coworker for chat and Voice Agents for inbound calls. ### Is Amazon Bedrock AgentCore open source? The managed service is not. AWS publishes the AgentCore CLI, a Python SDK and an MCP server on GitHub under Apache 2.0 for building and deploying agents onto it. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### How is Dynamiq priced compared with AgentCore? AgentCore bills each component, such as Runtime compute, Gateway calls and Memory events, separately by consumption, with a committed Runtime option launching by October 2026. Dynamiq offers a free plan to start and custom Enterprise pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. ## Sources - [Amazon Bedrock AgentCore overview (AWS documentation)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html) - [Amazon Bedrock AgentCore product page](https://aws.amazon.com/bedrock/agentcore/) - [Amazon Bedrock AgentCore pricing](https://aws.amazon.com/bedrock/agentcore/pricing/) - [Amazon Bedrock AgentCore FAQs](https://aws.amazon.com/bedrock/agentcore/faqs/) - [AgentCore Runtime Instances](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-instances.html) - [AgentCore supported regions](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-regions.html) - [AgentCore Runtime bidirectional streaming](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-bidirectional-streaming.html) - [AgentCore harness](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness.html) - [AgentCore Memory](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html) - [AgentCore Policy](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy.html) - [AgentCore Evaluations](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/evaluations.html) - [AgentCore Optimization](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/optimization.html) - [AgentCore Gateway](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html) - [AgentCore Agent Registry](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/registry.html) - [AgentCore CLI on GitHub (Apache 2.0)](https://github.com/aws/agentcore-cli) - [AWS services in scope: SOC](https://aws.amazon.com/compliance/services-in-scope/SOC/) - [AWS services in scope: FedRAMP](https://aws.amazon.com/compliance/services-in-scope/FedRAMP/) - [AWS services in scope: PCI](https://aws.amazon.com/compliance/services-in-scope/PCI/) - [AWS HIPAA eligible services reference](https://aws.amazon.com/compliance/hipaa-eligible-services-reference/) --- # Dynamiq vs Microsoft Copilot Studio URL: https://www.getdynamiq.ai/compare/microsoft-copilot-studio Microsoft Copilot Studio is a low-code studio for building agents and workflows that publish to Teams, Microsoft 365 Copilot, websites and other channels, with OpenAI and Anthropic models, a Visual Studio Code extension for developers, and billing in Copilot Credits. It runs only as a Microsoft-hosted service, including US government clouds. Dynamiq is a governed platform for chat, voice and workflow agents with an open-source SDK that runs on Dynamiq Cloud, any major cloud or on-prem. Teams standardized on Microsoft 365 and Power Platform fit Copilot Studio; teams that need agents inside their own environment, outside Microsoft's cloud, fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Microsoft Copilot Studio | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Organizations standardized on Microsoft 365 and Power Platform that want business users to build agents and workflows for Teams, Microsoft 365 Copilot, websites and other channels, for internal teams or external customers. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Microsoft-hosted only, as a Microsoft online service on Power Platform and Azure, with US government versions (GCC and GCC High) and EU Data Boundary support; no self-hosted or other-cloud deployment is documented. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | No. Copilot Studio runs only as a Microsoft cloud service. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | The service is proprietary. Microsoft publishes MIT-licensed tooling around it on GitHub, including the Copilot Studio extension for Visual Studio Code and the Microsoft 365 Agents SDK. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | OpenAI GPT and Anthropic Claude models are generally available for agents, with Mistral and xAI models as experimental options hosted outside Microsoft, and Azure AI Foundry models for prompts, per Microsoft Learn. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Chat agents, autonomous agents with triggers, workflows and agent flows, and computer use (generally available since May 2026); voice agents over PSTN and SIP through Dynamics 365 Contact Center, with real-time voice agents in preview. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Low-code authoring for business users, plus a Visual Studio Code extension (generally available since January 2026) that edits the full agent definition in YAML with Git, the Microsoft 365 Agents SDK, MCP servers as tools, and agent-to-agent (A2A) connections. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Administrators inventory and manage agents with role-based access and cost management; each new agent gets a Microsoft Entra Agent ID; workflows and agent flows include human review steps; Microsoft Purview audit logging is on by default. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Agent evaluations (generally available since March 2026) with test sets, a shared grader library, multi-turn tests and safety evaluators, run before and after publishing and from CI/CD through REST APIs, plus built-in analytics. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | Power Platform standard, premium and custom connectors, including non-Microsoft services such as Salesforce and Google, plus MCP servers; no total connector count is published on the pages checked. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | Covered under Microsoft's HIPAA business associate agreement, and listed for HITRUST CSF, FedRAMP, SOC, ISO, PCI DSS, CSA STAR and regional frameworks such as ENS High, K-ISMS and MTCS, per Microsoft Learn. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Not publicly documented on the Copilot Studio pages checked. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Copilot Credits: tenant-wide packs of 25,000 credits at $200 a month, pay-as-you-go at $0.01 per credit, and prepaid commit units that save up to 20 percent. A free trial is available, and unused credits do not carry over. | ## Choose Dynamiq when - Runs inside your environment, not only in Microsoft's cloud: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - The agent engine is an open-source Python SDK under Apache-2.0, with the visual canvas on the same engine, so developers can read, audit and fork what their agents run on. - 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM inside your environment and LoRA fine-tuning. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Microsoft Copilot Studio when - Already standardized on Microsoft 365 and Power Platform, with agents that should live in Teams and Microsoft 365 Copilot and reuse existing Power Platform connectors and Entra identities. - Business users build and edit agents and workflows with little technical background, while developers work on the same agents in Visual Studio Code with Git. - Government and regulated deployments inside Microsoft's cloud: US government clouds (GCC and GCC High), EU Data Boundary support and a HIPAA business associate agreement. ## Migration Agents built with Copilot Studio's standard harness, including topics, workflows and agent flows, are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK; the YAML agent definition available through the Visual Studio Code extension is a practical starting point for instructions and structure. Power Platform connectors reconnect as Dynamiq integrations or MCP servers, knowledge sources are re-ingested into Dynamiq Knowledge, and Azure Bot Service skills become direct API or MCP tool calls. Agents on the GitHub Copilot harness move over from their instructions, skills and MCP tools. ## Questions and answers ### Does Dynamiq require a Microsoft 365 subscription? No. Dynamiq runs independently of Microsoft 365, on Dynamiq Cloud or self-hosted on your own AWS, Azure, GCP, IBM Cloud, OpenShift or Kubernetes 1.32+ cluster, or on-prem. ### Can Copilot Studio run outside Microsoft's cloud? No. Copilot Studio is a Microsoft online service on Power Platform and Azure, with US government versions on Azure Government, per Microsoft Learn; no self-hosted option is documented. ### Does Copilot Studio support voice agents? Yes, over PSTN and SIP through Dynamics 365 Contact Center, and real-time voice agents were in preview when we checked, per Microsoft Learn. Dynamiq Voice Agents answer inbound calls over WebRTC, SIP and WebSocket. ### Which models can Copilot Studio agents use? OpenAI GPT and Anthropic Claude models are generally available, with Mistral and xAI models as experimental options hosted outside Microsoft, per Microsoft Learn. Dynamiq supports 29 model providers in its SDK, puts the most-used models behind one OpenAI-compatible AI Gateway endpoint, and can serve open models on vLLM inside your environment. ### Is Copilot Studio open source? No. The service is proprietary; Microsoft publishes MIT-licensed tooling around it, such as its Visual Studio Code extension and the Microsoft 365 Agents SDK. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ## Sources - [Copilot Studio overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio) - [Copilot Studio harnesses](https://learn.microsoft.com/en-us/microsoft-copilot-studio/harnesses-overview) - [Select a model for an agent](https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-select-agent-model) - [Copilot Studio pricing](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio) - [Power Platform pay-as-you-go meters](https://learn.microsoft.com/en-us/power-platform/admin/pay-as-you-go-meters) - [Copilot Studio licensing and subscriptions](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-licensing-subscriptions) - [Copilot Studio billing and licensing](https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing) - [Copilot Studio for US Government customers](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-licensing-gcc) - [Copilot Studio data locations](https://learn.microsoft.com/en-us/microsoft-copilot-studio/geo-data-residency) - [Copilot Studio compliance offerings](https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-certification) - [Real-time voice agents](https://learn.microsoft.com/en-us/microsoft-copilot-studio/voice-realtime-voice-agents) - [What's new in Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/whats-new) - [Agent evaluation](https://learn.microsoft.com/en-us/microsoft-copilot-studio/analytics-agent-evaluation-intro) - [Microsoft Entra Agent ID for Copilot Studio agents](https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-use-entra-agent-identities) - [Copilot Studio audit logging](https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio) - [Copilot Studio extension for Visual Studio Code](https://learn.microsoft.com/en-us/microsoft-copilot-studio/visual-studio-code-extension-overview) - [Connectors in Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-connectors) - [vscode-copilotstudio on GitHub (MIT)](https://github.com/microsoft/vscode-copilotstudio) - [Microsoft 365 Agents SDK on GitHub (MIT)](https://github.com/microsoft/Agents) --- # Dynamiq vs LangGraph and LangSmith URL: https://www.getdynamiq.ai/compare/langgraph LangGraph is an MIT licensed, code-first framework and runtime for stateful agents, and LangSmith is LangChain's agent engineering platform around it: observability, evaluation, deployment, the no-code Fleet agent builder, sandboxes, an LLM gateway in beta, and Engine, which finds agent failures and proposes fixes, billed by seats, traces and compute units. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 SDK, built-in evals and observability, and forward-deployed engineers. Developers who want maximum low-level control and framework-agnostic observability fit LangGraph and LangSmith; teams that want one platform including voice agents, deployed inside their environment with a delivery team, fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | LangGraph and LangSmith | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Python and JavaScript developers who want low-level control over stateful agents (single, multi-agent or hierarchical), with LangSmith for observability, evals and deployment across any framework. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | LangGraph runs wherever your code runs. LangSmith is a managed cloud with US, EU and APAC regions, with hybrid, bring-your-own-cloud and self-hosted options on its Enterprise plan; a standalone Agent Server needs a license key. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes for the LangGraph library. Self-hosted LangSmith is an add-on to the Enterprise plan, per LangChain's docs. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | LangGraph, LangChain and the LangSmith SDKs are MIT licensed. The langgraph-api server package is under the Elastic License 2.0, and the LangSmith platform is a commercial product that self-hosts under a license key. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Any provider LangChain integrates with (1,000+ integrations), called in code; LangSmith adds an LLM Gateway, in beta, for one key across providers with fallbacks, cost controls and PII redaction. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | A framework for custom agents, plus LangSmith Fleet, a no-code builder for agents used in chat or in tools such as Slack. No packaged voice agent product, though LangChain's docs include a voice agent tutorial. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Code-first primitives for agent graphs in Python and JavaScript with native token streaming, LangSmith SDKs in Python, TypeScript, Go and Java, the Studio agent IDE, a CLI, and MCP and A2A endpoints on Agent Server. | | Memory | Session memory scoped by user and session, plus long-term memory, where the agent decides when to save and recall durable, user-scoped facts that people can see and correct. Knowledge adds ingestion pipelines, permission-aware retrieval filters and knowledge graphs. | LangGraph's built-in memory stores conversation histories and keeps context across sessions, per LangChain's site. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Human-in-the-loop checks to steer and approve agent actions, and PII middleware in the framework; LangSmith Enterprise adds SAML SSO, SCIM, RBAC, ABAC and audit logs. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | LangSmith tracing, offline and online evals, dashboards and alerts, OpenTelemetry support, and Engine, which finds agent failures, diagnoses root causes and proposes fixes with eval coverage. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2, GDPR and HIPAA, with annual third-party audits and penetration testing, per LangSmith's shared responsibility model. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Enterprise support includes team trainings, architectural guidance and access to deployed engineers, per LangChain's pricing page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | LangSmith Developer is free (1 seat, up to 5,000 base traces a month), Plus is $39 a seat a month (up to 10,000 base traces), and Enterprise is custom; extra traces and storage bill in storage units ($1.00 each), and deployments, Fleet, Engine and sandboxes in compute units ($1.50 each). LangGraph itself is free. | ## Choose Dynamiq when - One governed platform for chat, voice and workflow agents, instead of a framework plus a separate platform billed by seats, traces and compute units. - Voice Agents answer inbound calls on the same engine as chat and workflow agents, with per-turn latency measurement and LLM-judged simulations. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - A defined engagement: forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. ## Choose LangGraph and LangSmith when - Maximum low-level control over agent orchestration for developers who want to write custom graphs in Python or JavaScript. - A very large open-source ecosystem, with 350M+ monthly open-source downloads reported on LangChain's homepage, and observability that works across frameworks, not only LangGraph. - A free LangSmith tier to start, with Engine to find failures, Fleet for no-code agents, and hybrid or self-hosted deployment on Enterprise. ## Migration Graphs, nodes and edges built with LangGraph are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK, keeping the same prompts, tools and control flow as the blueprint. Tools called from graph nodes reconnect as Dynamiq integrations or MCP servers, LangSmith datasets and evaluators move to Dynamiq evals, and tracing moves to Dynamiq observability. Memory maps to Dynamiq session and long-term memory, and knowledge sources are re-ingested into Dynamiq Knowledge. ## Questions and answers ### Does Dynamiq replace both LangGraph and LangSmith? It covers the core of both: Dynamiq's Agent Builder and Python SDK handle agent orchestration, and its built-in evals and observability handle tracing and evaluation, on one platform. LangSmith also sells products such as Fleet, sandboxes and an LLM gateway, so compare the pieces you use. ### Is LangGraph open source? Yes. LangGraph is open source under the MIT license, per its GitHub repository. The langgraph-api server package is under the Elastic License 2.0, and the LangSmith platform is a commercial product whose self-hosted version needs a license key. ### Does LangGraph or LangSmith include a voice agent product? No. LangChain's docs include a tutorial for building a voice agent, but neither LangGraph nor LangSmith ships a voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### How is LangSmith priced compared with Dynamiq? LangSmith bills by seat with base traces included per plan, plus storage units for extra traces and compute units for deployments, Fleet, Engine and sandboxes, per LangChain's pricing page. Dynamiq offers a free plan to start and custom Enterprise pricing. ## Sources - [LangChain homepage](https://www.langchain.com/) - [LangGraph overview](https://www.langchain.com/langgraph) - [LangSmith observability](https://www.langchain.com/langsmith) - [LangChain and LangSmith pricing](https://www.langchain.com/pricing) - [LangGraph overview (docs)](https://docs.langchain.com/oss/python/langgraph/overview) - [LangSmith deployment options (docs)](https://docs.langchain.com/langsmith/deployment) - [Self-hosted LangSmith (docs)](https://docs.langchain.com/langsmith/self-hosted) - [LangSmith Cloud regions (docs)](https://docs.langchain.com/langsmith/cloud) - [LangSmith Fleet (docs)](https://docs.langchain.com/langsmith/fleet) - [LangSmith LLM Gateway (docs)](https://docs.langchain.com/langsmith/llm-gateway) - [LangSmith user management and SSO (docs)](https://docs.langchain.com/langsmith/user-management) - [LangSmith shared responsibility model (docs)](https://docs.langchain.com/langsmith/shared-responsibility-model) - [Build a voice agent with LangChain (docs)](https://docs.langchain.com/oss/python/langchain/voice-agent) - [LangGraph GitHub license (MIT)](https://github.com/langchain-ai/langgraph/blob/main/LICENSE) - [langgraph-api on PyPI (Elastic License 2.0)](https://pypi.org/project/langgraph-api/) --- # Dynamiq vs CrewAI URL: https://www.getdynamiq.ai/compare/crewai CrewAI is an MIT licensed framework for multi-agent crews and flows, plus CrewAI AMP, a commercial platform with a no-code Studio, tracing, governance and deployment on CrewAI's cloud, your VPC or your own infrastructure through CrewAI Factory; forward-deployed engineering is sold a la carte. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 SDK, built-in evals including real-time evaluations, and forward-deployed engineers who ship the first production agent with your team. Developers who want a popular multi-agent framework with a managed layer fit CrewAI; teams that need voice agents, built-in evals or a defined deployment engagement fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | CrewAI | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Business and technical teams that build multi-agent crews and flows, with platform teams governing them in CrewAI AMP. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | The open-source framework runs wherever Python runs. CrewAI AMP deploys on CrewAI's cloud, your VPC or your own infrastructure through CrewAI Factory, on Kubernetes such as EKS, AKS, GKE or Red Hat OpenShift. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. The MIT framework runs in your environment, and CrewAI Factory self-hosts the AMP platform in a data center or VPC. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | The crewAI framework is open source under the MIT license, per its GitHub repository; CrewAI AMP is a commercial platform with a free Basic tier. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Providers documented on its LLMs page include OpenAI, Anthropic, Google, Azure, AWS Bedrock, Mistral, NVIDIA, Groq, IBM watsonx.ai and Ollama, through each provider's native SDK. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Multi-agent crews and event-driven flows, conversational flows for multi-turn chat, and workflow chat in a web UI, Slack or Teams; no voice agent product is described. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A code-first Python framework (crews, agents, tasks and flows, in code or YAML) with a CLI, plus Crew Studio, a no-code visual editor on the free plan that exports to Python, and exports as MCP servers or UI components. | | Memory | Session memory scoped by user and session, plus long-term memory, where the agent decides when to save and recall durable, user-scoped facts that people can see and correct. Knowledge adds ingestion pipelines, permission-aware retrieval filters and knowledge graphs. | Memory and knowledge are built into the framework, alongside guardrails and observability, per CrewAI's docs. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Guardrails and human-in-the-loop approval gates in the framework; AMP Enterprise adds SSO with Microsoft Entra ID, Okta and other providers over SAML or OIDC, RBAC with custom roles, immutable audit trails, PII redaction and a hallucination guardrail. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Real-time tracing of LLM calls, tool calls and memory reads with cost accounting, OpenTelemetry export, crewai test and eval commands, and evaluation through partners such as Arize, Galileo, Datadog and Patronus. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | Crew Studio connects to more than 1,000 applications, and the framework ships hundreds of open-source tools with MCP support. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | CrewAI cites a SOC 2 report and a HIPAA Security Rule gap assessment, per a May 2025 post on its blog. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Forward-deployed engineering and training are available a la carte, and Enterprise includes a 45-day onboarding with support for deployment and development, per CrewAI's pricing page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | A free Basic plan with 50 workflow executions a month and the Studio editor, and custom Enterprise pricing sized to your workflows, per CrewAI's pricing page. | ## Choose Dynamiq when - Voice Agents answer inbound calls on the same engine as chat and workflow agents, with per-turn latency measurement and LLM-judged simulations. - Built-in evals rather than partner add-ons: datasets from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments. - AI Coworker works with its own cloud sandbox in the browser, Slack, Microsoft Teams and Telegram, next to the workflows you build. - Forward-deployed engineers are part of the engagement: a bootcamp in your environment, the first production agent shipped with your team, then a handover. ## Choose CrewAI when - A widely adopted open-source multi-agent framework, with 450M+ agentic workflows a month reported on CrewAI's homepage. - A simple role-based abstraction, crews of agents with tasks, plus a free no-code Studio that exports to Python. - Self-hosting through CrewAI Factory on Kubernetes, including Red Hat OpenShift, with forward-deployed engineering bought a la carte when you want it. ## Migration Crews, agents, tasks and flows defined in CrewAI's Python framework, in code or YAML, are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK, keeping each agent's role, goal and instructions. Tools attached to CrewAI agents reconnect as Dynamiq integrations or MCP servers, knowledge sources move to Dynamiq Knowledge, and crew memory maps to Dynamiq session and long-term memory. ## Questions and answers ### Is CrewAI open source? The crewAI framework is open source under the MIT license, per its GitHub repository. CrewAI AMP, the platform layered on top, is commercial, with a free Basic tier. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Can CrewAI be self-hosted? Yes. CrewAI Factory deploys the AMP platform on your own infrastructure or VPC, per CrewAI's pricing page and docs. Dynamiq also self-hosts on your cloud or on-prem, with isolated setups delivered with our engineers. ### Does CrewAI include forward-deployed engineers? CrewAI offers forward-deployed engineering and training a la carte, and its Enterprise plan includes a 45-day onboarding, per its pricing page. Dynamiq's forward-deployed engineers run a bootcamp in your environment and ship your first production agent with your team as part of the engagement. ### Does CrewAI support voice agents? Its site and documentation describe multi-agent crews, flows and chat, without a packaged voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### How is Dynamiq priced compared with CrewAI? CrewAI offers a free Basic plan with 50 workflow executions a month and custom Enterprise pricing, per its pricing page. Dynamiq offers a free plan to start and custom Enterprise pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. ## Sources - [CrewAI homepage](https://www.crewai.com/) - [CrewAI pricing](https://www.crewai.com/pricing) - [CrewAI open source](https://www.crewai.com/open-source) - [CrewAI Agent Management Platform](https://www.crewai.com/agent-management-platform) - [CrewAI docs](https://docs.crewai.com/) - [CrewAI LLMs (docs)](https://docs.crewai.com/en/concepts/llms) - [CrewAI platform SSO (docs)](https://docs-platform.crewai.com/platform/en/features/sso) - [crewAI GitHub license (MIT)](https://github.com/crewAIInc/crewAI/blob/main/LICENSE) - [CrewAI Factory for self-hosting (May 2025)](https://www.crewai.com/blog/unlocking-agent-native-transformation-with-crewai-factory-and-nvidia) - [Crew Studio, the automated agent builder (July 2026)](https://www.crewai.com/blog/crew-studio-automated-agent-builder) --- # Dynamiq vs Dify URL: https://www.getdynamiq.ai/compare/dify Dify is a visual platform for agentic workflows, chat apps and RAG pipelines with a large plugin marketplace, available as Dify Cloud, as a self-hosted Community Edition under a modified Apache License that bars multi-tenant hosting without permission, or as Dify Enterprise in a VPC, on-prem or air-gapped. Dynamiq is a governed platform for chat, voice and workflow agents with an unrestricted Apache-2.0 SDK, built-in evals and observability, and forward-deployed engineers. Teams that want a fast, community-backed builder for LLM apps and RAG fit Dify; teams that need voice agents, built-in evals, an unrestricted open-source engine or a deployment team fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Dify | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Teams that want a visual platform for agentic workflows, chat apps and RAG pipelines, from a free cloud Sandbox or self-hosted Community Edition up to Enterprise deployments in a VPC or on-prem. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Dify Cloud (Sandbox, Professional and Team), a self-hosted Community Edition with Docker, Dify Enterprise in your VPC, on-prem or air-gapped with Helm and Terraform, and Dify Premium as an AWS machine image. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. The Community Edition self-hosts with Docker, and Dify Enterprise deploys in your VPC, on-prem or in an air-gapped network. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Source-available under a modified Apache License 2.0: without Dify's written authorization you may not operate a multi-tenant environment (one tenant per workspace), and you may not remove Dify's logo or copyright from the console or apps, per its GitHub license file. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Hundreds of proprietary and open-source models from dozens of providers and self-hosted solutions, including OpenAI, Anthropic, Gemini and any OpenAI-compatible model, installed as plugins from its marketplace. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Agentic workflows, chat apps and agents, including agents you build by chatting and agents with their own sandbox. Speech-to-text and text-to-speech are app features, but no voice agent product is described. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A visual builder with Python or JavaScript code nodes, apps exported as a YAML DSL, the difyctl CLI, and publishing as hosted apps, API endpoints, embeds or MCP servers; Dify also connects to MCP servers as a client. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | A Human Input node pauses workflows for a person's decision, and a moderation extension reviews inputs and outputs; Dify Enterprise adds SSO over SAML or OIDC, SCIM, workflow-level RBAC, audit logs streamed to a SIEM and PII-redacted prompt history. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Logs, feedback, annotations, latency and token usage in its dashboard, plus tracing through integrations such as LangSmith, Langfuse, Opik, W&B Weave, Arize and Phoenix; no built-in evaluation suite is described on the pages checked. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | A marketplace of model providers, tools, data sources, agent strategies and MCP integrations, with 928 listed plugins in a July 2026 snapshot, per Dify's blog; Knowledge Pipelines turn files, websites and drives into knowledge bases. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2 and ISO 27001, with a GDPR badge, per Dify's homepage; HIPAA is not stated on the pages checked. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Dify Enterprise includes a named customer success manager, a shared Slack channel and a one-hour response on severity-one issues; no engineering deployment engagement is described. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Cloud: a free Sandbox (200 message credits), Professional at $590 per workspace a year, Team at $1,590 per workspace a year, and custom Enterprise; the Community Edition is free to self-host, per Dify's pricing page. | ## Choose Dynamiq when - The SDK is open source under Apache-2.0, with no added conditions on how you run it; the platform around it is commercial. - Voice Agents answer inbound calls on the same engine as chat and workflow agents, with per-turn latency measurement and LLM-judged simulations. - Built-in evals: datasets from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments, next to traces with cost and latency per node. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Dify when - A free, self-serve start: a cloud Sandbox and a self-hosted Community Edition for prototyping LLM apps and RAG pipelines quickly. - A large marketplace of plugins for models, tools, data sources and MCP, and an active open-source community around one product. - Low-cost cloud tiers for small teams, with a path to Enterprise in a VPC, on-prem or air-gapped. ## Migration Dify apps and workflows export as a YAML DSL with their orchestration, node settings, prompts and model parameters, which gives a clear blueprint for rebuilding them as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK. Tools and MCP integrations reconnect as Dynamiq integrations or MCP servers, and knowledge bases are re-ingested into Dynamiq Knowledge. Third-party tool keys, which the DSL does not export, are reconnected once. ## Questions and answers ### Is Dify open source like Dynamiq's SDK? Dify's Community Edition is source-available under a modified Apache License 2.0 that bars operating a multi-tenant environment without Dify's written authorization and requires keeping its logo and copyright, per its GitHub license file. Dynamiq's Python SDK is open source under Apache-2.0 with no added conditions; the Dynamiq platform is not open source. ### Does Dify support voice agents? Not as a product. Dify offers speech-to-text and text-to-speech as app features, but its site describes no voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### Does Dify have built-in evaluations? Dify records logs, feedback, annotations and token usage, and connects to tracing tools such as LangSmith and Langfuse, but its pages describe no built-in evaluation suite. Dynamiq includes evals with LLM-as-a-judge and real-time evaluations on live deployments. ### Is Dify free to use? Yes, with a free cloud Sandbox and a free, self-hosted Community Edition, alongside paid Professional, Team and Enterprise plans, per Dify's pricing page. Dynamiq also offers a free plan to start, plus custom Enterprise pricing. ### Does Dify support MCP? Yes, in both directions: Dify apps publish as MCP servers, and Dify connects to external MCP servers as a client, per its docs. Dynamiq also supports MCP servers alongside 2,100+ app integrations. ## Sources - [Dify homepage](https://dify.ai/) - [Dify pricing](https://dify.ai/pricing) - [Dify Enterprise](https://dify.ai/enterprise) - [Dify GitHub license](https://github.com/langgenius/dify/blob/main/LICENSE) - [Dify on GitHub (README)](https://github.com/langgenius/dify) - [Dify Human Input node (docs)](https://docs.dify.ai/en/cloud/use-dify/nodes/human-input) - [Publish Dify apps as MCP servers (docs)](https://docs.dify.ai/en/cloud/use-dify/publish/publish-mcp) - [Dify Marketplace](https://marketplace.dify.ai/) - [Dify on governing its plugin ecosystem (September 2026)](https://dify.ai/blog/trust-is-a-feature-how-dify-is-governing-a-growing-plugin-ecosystem) - [Dify's official CLI, difyctl (July 2026)](https://dify.ai/blog/dify-s-official-cli-difyctl-ships) - [Introducing the new Dify Agent (August 2026)](https://dify.ai/blog/introducing-new-dify-agent) --- # Dynamiq vs n8n URL: https://www.getdynamiq.ai/compare/n8n n8n is a workflow automation platform, source-available under its fair-code Sustainable Use License, with a large integration library, AI Agent nodes, a Guardrails node, human approvals, evaluations and, since September 2026, standalone n8n Agents in preview. Dynamiq is a governed platform built for agents, chat, voice and workflow, with an Apache-2.0 SDK, real-time evaluations on live deployments, inbound voice agents and forward-deployed engineers. Teams automating business processes across many apps, with AI as one part of the work, fit n8n; teams building governed AI agents, including voice, fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | n8n | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Teams that want visual workflow automation across many apps and APIs, with AI agents, code and human steps in the same workflows, self-hosted or on n8n's cloud. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | n8n Cloud (Starter, Pro and Enterprise), or self-hosted: a free Community Edition, a self-hosted Business plan, and Enterprise hosted by n8n or self-hosted, including air-gapped on a private network, from a Helm chart or Compose file. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. The Community Edition self-hosts for free, and the paid Business and Enterprise plans add SSO, LDAP, Git version control and air-gapped deployment. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Source-available under the Sustainable Use License (fair-code): use and modification are limited to internal business, non-commercial or personal purposes, and distribution is allowed only free of charge for non-commercial purposes; enterprise files are under a separate n8n Enterprise License. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Model nodes for many providers with your own keys, plus Gateway credits (September 2026) on Cloud Starter and Pro for models such as OpenAI, Anthropic and Google Gemini from one prepaid balance; no total provider count is published. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Workflows with AI Agent nodes, sub-agents and human steps, plus n8n Agents (in preview since September 2026) that people reach through chat, channels such as Slack, Telegram and Linear, or schedules, and Chat Hub for chat across models and agents. No voice agent product is described. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A node-based canvas with JavaScript and Python code, an AI Workflow Builder that drafts workflows from a description, a REST API and CLI, JSON export, MCP as client and server, and Git version control on Business and Enterprise. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | A Guardrails node that checks text for jailbreaks, PII, secret keys and off-topic content, human approval before an AI agent runs a chosen tool, RBAC on paid plans, and SSO over SAML or OIDC plus LDAP on Business and Enterprise. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Quick evaluations during development and metric-based evaluations (such as correctness, helpfulness and tools used) in production, Insights dashboards, OpenTelemetry tracing with agent spans, and log streaming on Enterprise. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | Its integrations page lists more than 2,200 integrations, plus an HTTP Request node for any API, vector store nodes, data tables and a library of more than 10,000 workflow templates. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2 audited, with a public SOC 3 report and GDPR listed on its trust center; HIPAA and ISO 27001 are not stated on its security page. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Dedicated support with an SLA on Enterprise, and an Experts partner program for agencies and integrators; no n8n-staffed deployment engagement is described. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Cloud Starter at 20 euro a month (2,500 executions) and Pro at 50 euro a month (10,000), self-hosted Business at 667 euro a month (40,000), all billed annually, and custom Enterprise; the Community Edition is free, and a 14-day trial is available, per n8n's pricing page. | ## Choose Dynamiq when - The SDK is open source under Apache-2.0, with no limits on commercial use or hosting; the platform around it is commercial. - Built for agents first: AI Coworker with its own cloud sandbox, Voice Agents for inbound calls, and Agent Builder on one engine. - Real-time evaluations on live deployments and an AI agent that surfaces issues from production traces, next to traces with cost and latency per node. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose n8n when - A mature workflow automation platform with a very large integration and template library, useful well beyond AI use cases. - A free, self-hosted Community Edition and low-cost cloud plans for smaller teams. - Guardrails, human approvals and evaluations inside the same workflows your team already automates, with Git version control on Business and Enterprise. ## Migration n8n workflows export as JSON, which gives a clear map for rebuilding them as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK. App nodes and webhooks reconnect as Dynamiq integrations or MCP servers, JavaScript or Python code nodes move to SDK functions, and credentials are re-created in Dynamiq, since exported workflows carry only credential names and IDs. AI Agent prompts and tool choices carry over as the starting configuration for each agent. ## Questions and answers ### Is n8n open source like Dynamiq's SDK? n8n is source-available under the Sustainable Use License (fair-code), which limits use to internal business, non-commercial or personal purposes and allows only free, non-commercial distribution. Dynamiq's Python SDK is open source under Apache-2.0 without those limits; the Dynamiq platform is not open source. ### Does n8n support voice agents? Not as a product. n8n's agents and workflows work across chat and messaging channels, but its site describes no voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### Does n8n have guardrails and evaluations? Yes. n8n has a Guardrails node, human approval for chosen agent tools, and quick and metric-based evaluations, per its docs. Dynamiq adds real-time evaluations on live deployments and an AI agent that surfaces issues from production traces. ### Can n8n run self-hosted for free? Yes. The Community Edition self-hosts for free under the Sustainable Use License; the self-hosted Business and Enterprise plans add SSO, LDAP and Git version control, per n8n's pricing page. ### Is n8n SOC 2 certified? n8n states it is SOC 2 audited, publishes a SOC 3 report and lists GDPR on its trust center. Dynamiq holds SOC 2 and CASA Tier 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. ## Sources - [n8n homepage](https://n8n.io/) - [n8n pricing](https://n8n.io/pricing/) - [n8n security](https://n8n.io/security/) - [n8n Trust Center](https://trust.n8n.io/) - [n8n Enterprise](https://n8n.io/enterprise/) - [n8n integrations](https://n8n.io/integrations/) - [n8n GitHub license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md) - [Build and manage agents (docs)](https://docs.n8n.io/build/build-and-manage-agents/) - [Guardrails node (docs)](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.guardrails/) - [Test and improve AI workflows (docs)](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/) - [Introducing n8n Agents (September 2026)](https://blog.n8n.io/introducing-n8n-agents/) - [n8n Gateway credits (September 2026)](https://blog.n8n.io/gateway-credits/) --- # Dynamiq vs StackAI URL: https://www.getdynamiq.ai/compare/stackai StackAI, now part of Asana and still run as its own product, is a closed, drag-and-drop platform for IT and operations teams to build, govern and scale agents, with deployment from multi-tenant cloud to air-gapped on-premises, 300+ integrations, and forward-deployed engineers who co-build weekly with every enterprise customer. Dynamiq covers the same deployment depth on a governed platform that also includes inbound voice agents, an Apache-2.0 open-source SDK for teams that build in code, and real-time evaluations on live deployments. IT teams that want a no-code builder with hands-on support fit StackAI; teams that also need voice agents, an open-source engine or evals on live traffic fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | StackAI | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | IT and operations teams that want a drag-and-drop, no-code platform to build, deploy, govern and scale agents across an enterprise, per StackAI's site. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Four deployment models from multi-tenant cloud to air-gapped on-premises, including VPC and dedicated infrastructure; on-prem installs run on AWS, GCP, Azure or your own servers, per its security page and docs. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. On-premises and virtual private cloud deployment are Enterprise features, per StackAI's pricing page, and existing deployments continue unchanged after the Asana acquisition, per its announcement. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed source; StackAI's GitHub organization has no public repositories, and no SDK is published. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Model-agnostic: providers such as OpenAI, Anthropic, Google and AWS Bedrock, models hosted on Azure or Bedrock, and local LLMs, with admins able to turn off StackAI-provided keys or block providers. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Agents and workflows, including Multi-Agent Teams (September 2026), delivered as forms, chat assistants, website chatbots, batch runs, APIs, and apps in Slack, Microsoft Teams, WhatsApp or SMS. Audio nodes transcribe and generate speech, but no real-time voice agent is offered. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A drag-and-drop builder with Python code and custom API nodes, a REST API, an MCP server for assistants such as Claude and Cursor, MCP client nodes, and JSON project export; no general-purpose SDK is published. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | PII detection and masking, retention controls, admin review before a project is published with version history and diffs, human-in-the-loop approval nodes, SSO with Okta or Entra ID, SAML, SCIM, RBAC and audit logs; DPAs keep customer data out of model training. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | An Evaluator that scores a test set with LLM-as-a-judge before a project goes live, plus analytics on usage, latency, tokens and users, run history and per-step traces. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | 300+ integrations plus MCP servers, in a public catalog on StackAI's site. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2 audited and ISO 27001 certified with continuous monitoring, per its trust center, plus HIPAA, GDPR and CCPA-CPRA, with a BAA available, per its site and docs. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Forward-deployed engineers lead weekly co-building sessions with every enterprise customer, alongside dedicated solution engineers, per StackAI's docs and pricing page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | A free plan at $0 a month (500 runs, 2 projects, 1 seat), and custom Enterprise pricing with on-prem or VPC deployment, dedicated infrastructure, solution engineers and SSO, per StackAI's pricing page. | ## Choose Dynamiq when - The agent engine is an open-source Python SDK under Apache-2.0, with the visual canvas on the same engine, so developers can build in code and read what their agents run on. - Voice Agents answer inbound calls over WebRTC, SIP and WebSocket on the same engine as chat and workflow agents. - Real-time evaluations on live deployments and an AI agent that surfaces issues from production traces, not only tests before launch. - AI Coworker works with its own cloud sandbox in the browser, Slack, Microsoft Teams and Telegram. ## Choose StackAI when - A drag-and-drop, no-code builder aimed at IT and operations teams standing up agents across many business units. - Forward-deployed engineers who co-build weekly with every enterprise customer, plus dedicated solution engineers. - 300+ integrations in a public catalog, and four deployment models up to air-gapped on-premises. ## Migration Agents and workflows built on StackAI's canvas export as JSON, which gives a map for rebuilding them as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK. Integrations reconnect as Dynamiq integrations or MCP servers, knowledge bases move to Dynamiq Knowledge, and human-in-the-loop nodes become Dynamiq approvals. ## Questions and answers ### Is StackAI open source? No. StackAI is closed source, and its GitHub organization has no public repositories. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Can both Dynamiq and StackAI run on premises? Yes for both. StackAI lists on-premises and virtual private cloud deployment as Enterprise features, per its pricing page. Dynamiq self-hosts on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem. ### What does the Asana acquisition mean for StackAI? Per StackAI's announcement, it continues to operate as its own product and brand after joining Asana, and existing deployments, including on-prem and VPC, remain functional. ### Does StackAI support voice agents? Not as a product. StackAI has audio nodes that transcribe recordings and generate speech, and text messaging over WhatsApp or SMS, but no real-time voice agent. Dynamiq includes Voice Agents for inbound calls. ### Does StackAI offer a free plan? Yes, a free plan at $0 a month with 500 runs, 2 projects and 1 seat, per its pricing page. Dynamiq also offers a free plan to start, plus custom Enterprise pricing. ## Sources - [StackAI homepage](https://www.stackai.com/) - [StackAI pricing](https://www.stackai.com/pricing) - [StackAI security](https://www.stackai.com/security) - [StackAI integrations](https://www.stackai.com/integrations) - [StackAI Trust Center](https://trust.stackai.com/) - [StackAI joins Asana (announcement)](https://www.stackai.com/blog/stackai-joins-asana-to-build-the-future-of-agentic-work-management) - [StackAI Multi-Agent Teams (September 2026)](https://www.stackai.com/blog/introducing-multi-agent-teams-turn-one-ai-agent-into-a-team-of-specialists) - [What makes StackAI unique (docs)](https://docs.stackai.com/welcome-to-stackai/overview/what-makes-stackai-unique) - [StackAI docs](https://docs.stackai.com/) - [StackAI on GitHub](https://github.com/stackai) --- # Dynamiq vs Langdock URL: https://www.getdynamiq.ai/compare/langdock Langdock is a closed platform for rolling out AI across an organization, with chat, agents, workflows, integrations and an API, hosted on Microsoft Azure in the EU by default, with single-tenant, bring-your-own-cloud and on-premises options from 2,000 or 5,000 Enterprise seats. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 open-source SDK, self-hosted on any cloud or on-prem, and forward-deployed engineers. Organizations rolling AI out to a whole workforce under EU hosting fit Langdock; teams that want voice agents, an open-source engine or engineers who ship agents with them fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Langdock | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Organizations that want to roll AI out across their whole workforce, with chat, agents and workflows hosted in Europe, per Langdock's own positioning. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Multi-tenant SaaS on Microsoft Azure servers inside the EU, single-tenant SaaS from 2,000 Enterprise seats, and bring-your-own-cloud or on-premises through Kubernetes and Helm charts from 5,000 seats, per its security page. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes, on higher Enterprise tiers: bring-your-own-cloud and on-premises Kubernetes deployment each start at 5,000 seats, per its security page. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed platform; Langdock's public GitHub holds an MIT-licensed cookbook of examples, not platform code. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Model-agnostic chat and one API across models from providers such as OpenAI, Anthropic, Google, Mistral, Meta and DeepSeek, with bring-your-own-key options; new models usually reach the EU a few weeks after the US. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Chat, agents with subagents, and workflows that are agentic or fully deterministic; a workflow node transcribes audio into text, but no voice agent product is offered. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A no-code builder for agents and workflows with a code node, APIs for completions, agents, workflows, knowledge and audit logs, MCP as client and server, A2A, and apps in Slack, Teams, Excel, Outlook and a desktop app. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | SSO, SAML and SCIM on its Business plan, audit logs, a Guardrails workflow node that detects PII, jailbreaks and hallucinations, human-in-the-loop approvals in workflows, and a Governance add-on for agent review and compliance rules; customer data is never used to train models, per its security page. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Agent evals that run test sets through an AI judge, tool checks or keyword checks before publishing, agent analytics, usage export, and workflow run history with versioning. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | 57 integrations with 754 native actions, per its integrations page, plus MCP and folder sync from SharePoint, Google Drive, OneDrive and Confluence. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | ISO 27001, a SOC 2 audit and GDPR alignment, per its security page; HIPAA is not mentioned there. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Enterprise includes a dedicated technical implementation team, enablement sessions and a remote or on-site kickoff workshop, per Langdock's enterprise page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | A 7-day free trial, then seats at 25 euro (Standard) or 99 euro (Business Max) a user a month for up to 1,000 users and custom Enterprise above that; Workflows packages, a Governance add-on (free until 2027, then 3.50 euro a user a month), and API usage billed per model by tokens, per Langdock's pricing page. | ## Choose Dynamiq when - The agent engine is an open-source Python SDK under Apache-2.0, so the code your agents run on is yours to read, audit and fork. - Voice Agents answer inbound calls on the same engine as chat and workflow agents. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, including EU regions and data centers you control. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Langdock when - A platform built around EU hosting by default, from a European company, for organizations whose policy requires it. - Ready-made chat, agents, workflows and office add-ins for rolling AI out to an entire non-technical workforce quickly. - A stated policy that customer data is never used to train AI models, and a Governance add-on with compliance rule templates for the EU AI Act and GDPR. ## Migration Agents and workflows built in Langdock can be read through its Agents and Workflows APIs, including instructions, models, actions, knowledge folders and workflow graphs, which gives a blueprint for rebuilding them in Dynamiq's Agent Builder, on the canvas or in the Python SDK. Integrations reconnect as Dynamiq integrations or MCP servers, knowledge folders are re-ingested into Dynamiq Knowledge, and Guardrails and human-in-the-loop nodes become Dynamiq guardrails and approvals. ## Questions and answers ### Can Dynamiq run entirely inside the EU like Langdock? Yes. Dynamiq self-hosts on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, including EU regions and data centers you control. Langdock hosts its multi-tenant service on Azure in the EU and notes that new models usually reach the EU a few weeks after the US. ### Is Langdock open source? No. Langdock's platform is closed; its public GitHub holds an MIT-licensed cookbook of examples. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Does Langdock support voice agents? No. Langdock offers chat, agents and workflows, and a workflow node that transcribes audio, but no voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### Does Langdock require a seat minimum to self-host? Yes. Per its security page, bring-your-own-cloud and on-premises deployment start at 5,000 Enterprise seats, and single-tenant SaaS at 2,000. Dynamiq self-hosting comes with its Enterprise plan. ### Is Langdock HIPAA compliant? Langdock's security page states ISO 27001 certification, a SOC 2 audit and GDPR alignment, but does not mention HIPAA. Dynamiq signs a BAA for HIPAA workloads. ## Sources - [Langdock homepage](https://www.langdock.com/) - [Langdock pricing](https://www.langdock.com/pricing) - [Langdock security](https://www.langdock.com/security) - [Langdock Enterprise](https://www.langdock.com/enterprise) - [Langdock integrations](https://www.langdock.com/products/integrations) - [Langdock API](https://www.langdock.com/products/api) - [Langdock Workflows](https://www.langdock.com/products/workflows) - [Langdock models](https://www.langdock.com/models) - [Langdock Guardrails node (docs)](https://docs.langdock.com/en/using-langdock/workflows/nodes/guardrails-node) - [Langdock agent evals (docs)](https://docs.langdock.com/en/using-langdock/agents/agent-evals) - [Langdock cookbook on GitHub](https://github.com/langdock/langdock-cookbook) --- # Dynamiq vs Glean URL: https://www.getdynamiq.ai/compare/glean Glean is a closed platform for enterprise search, an AI assistant and agents, built on a hybrid search index and Enterprise Graph across 275+ connectors that mirror source permissions, with a Model Hub of 40+ models, a hosted MCP server and per-seat pricing. It runs as single-tenant SaaS or as a tenant Glean operates inside your own GCP or AWS account. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 SDK, Knowledge Bases with permission-aware retrieval, AI Coworker in the browser, Slack, Microsoft Teams and Telegram, and self-hosting on any cloud or on-prem. Organizations whose first need is search across every app fit Glean; teams that want assistants and agents that act on company knowledge inside their own perimeter fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Glean | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Enterprises that want one AI assistant and search experience for every employee, grounded in a permission-aware index of company apps, with agents built on the same context, per Glean's positioning. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Glean Hosted, a single-tenant SaaS in multiple GCP and AWS regions, or Customer Hosted (formerly Cloud-Prem), a tenant Glean deploys and operates as a managed service inside your own GCP project or AWS account, per its docs. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | No on-premises option is documented. Customer Hosted runs in your GCP or AWS account, but Glean calls it not a traditional self-hosted model: Glean handles upgrades, patching and monitoring, keeps support access, and does not support customers deploying or patching its services. | | Isolated environments | Fully isolated and air-gapped environments are delivered with our engineers. | Not documented. In both models Glean keeps support access, and a Customer Hosted build on AWS pulls container images and control-plane bootstrap secrets from Glean's central AWS account, per its AWS deployment guide. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed platform. Glean publishes API clients in Python, TypeScript, Java and Go, a Web SDK and an Indexing SDK for building on it, per its developer site. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | A Model Hub of 40+ open and proprietary models: OpenAI GPT models through Azure or OpenAI, Gemini through Vertex AI, Claude through Vertex AI, Amazon Bedrock or Anthropic, Amazon Nova, and open models such as Nemotron 3 Ultra. Glean's Universal Model Key adds auto routing (beta); with your own keys, Vertex AI is limited to GCP-based deployments and Amazon Bedrock to AWS, per its docs. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Glean Assistant for chat, with real-time voice in the app, and Glean Agents: workflow agents with branching and loops, autonomous Auto Mode agents, sub-agents and triggers. Independent agents are in beta and the Glean Tau desktop workspace is coming soon, per its August 2026 launch. No phone-line voice agent product is described. | | Enterprise search and knowledge | Knowledge Bases built from uploads, websites and selected files and pages synced from Google Drive, Notion, Dropbox, OneDrive, SharePoint, Box and Confluence, re-synced in the background. Permission syncing on Google Drive and Confluence sources enforces source permissions at retrieval time; agents retrieve with hybrid search, metadata filters and optional knowledge graphs. | Its core product: a hybrid search index and Enterprise Graph across connected apps, with source permissions inherited and enforced, content and permission changes synced in real time, results personalized by role, and people and expert search, per its search and connectors pages. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A no-code Agent Builder, conversational or drag-and-drop, with Auto Mode, drafts, versioning and rollback; Search, Chat, Agents and Indexing APIs with clients in Python, TypeScript, Java and Go; a Web SDK; and a hosted MCP server that brings Glean search, chat and agents into ChatGPT, Claude, Cursor, VS Code and other MCP hosts. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | SSO is mandatory (OIDC or SAML), agents check source permissions on every read and write, write tools pause for confirmation in the web app, admins set agent sharing rules, Agent Access Policies are in beta, and audit events stream to a SIEM. AI security models for prompt injection, malicious code and harmful content come with Glean Protect+, licensed site-wide for an annual fee. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Debug and trace views for every agent run, with inputs, tool calls, model decisions and outputs; insights into adoption, error rates, feedback and ROI for agents; and lifecycle guidance for testing agents against golden test sets before launch. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | 275+ native and MCP-based connectors, per its connectors page, plus native read and write actions, custom actions from OpenAPI specs, an Indexing SDK for custom sources, and a growing library of MCP server templates. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2, ISO 27001, ISO 42001, HIPAA, GDPR and TX-RAMP Level 2, per its homepage; zero LLM data retention and encryption in transit and at rest are part of Glean Core Suite. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Glean describes its platform as deployable without the burden of heavy professional services; deployment and adoption help comes through the Glean Partner Network, launched in August 2026, whose partners hold more than 2,000 accreditations across sales, technical and delivery work. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | No public price list. Per-user, per-month seats: Glean Core Suite adds model usage on an upfront commit at provider API rates plus a management fee, and Enterprise Flex adds a pooled allowance of FlexCredits for agents, APIs and premium models. Seats and credits cost less when you host Glean in your own cloud or bring your own model keys, per its docs. | ## Choose Dynamiq when - Your perimeter, your operations: self-host on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, operated by your team, with isolated setups delivered with our engineers. Glean's own-cloud option covers GCP and AWS and stays operated by Glean. - Agents that act as well as answer: AI Coworker works with its own cloud sandbox computer in the browser, Slack, Microsoft Teams and Telegram, with personal or organization-wide connections to your apps, and deployed agents can ask each end user to connect their own account. - Workflow agents in Agent Builder, on a visual canvas or in the open-source Python SDK under Apache-2.0, with approvals on any tool step, guardrails that detect PII and prompt injection, and real-time evaluations on live deployments. - Voice Agents answer inbound calls on the same engine, and forward-deployed engineers ship the first production agent with your team, then hand over. ## Choose Glean when - Company-wide search is the main job: one permission-aware index and Enterprise Graph across 275+ connectors, with real-time sync of content and permissions, personalized results and people search. - An assistant for every employee, in browser, desktop and mobile apps and inside Slack and Microsoft Teams, licensed per seat and fully managed by Glean, even when it runs in your own GCP or AWS account. - Company context inside other AI tools: Glean's hosted MCP server brings its search, chat and agents into ChatGPT, Claude, Cursor, VS Code and other MCP hosts, and its AI Gateway, in beta, governs model and tool traffic across them. ## Migration Agents built in Glean's Agent Builder are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK, starting from the same instructions, steps and model choices. Content Glean indexes from Google Drive, SharePoint, OneDrive, Box, Dropbox, Notion and Confluence syncs into Dynamiq Knowledge Bases, with permission syncing on Google Drive and Confluence; other apps connect as Dynamiq integrations or MCP servers, and write steps that needed confirmation in Glean become Dynamiq approval steps. Teams that keep Glean for company-wide search can call its Search or Chat API from a Dynamiq workflow. ## Questions and answers ### Can Dynamiq replace Glean? In part. Dynamiq replaces Glean where the goal is an assistant and agents over company knowledge that run inside your own perimeter and take action: AI Coworker, workflow agents with approvals, and Knowledge Bases with permission-aware retrieval, self-hosted on any cloud or on-prem. For search across every app in the company, with 275+ connectors and real-time permission sync, Glean is the stronger choice, and the two can run side by side. ### Can Glean run in our own cloud or on-prem? In your own cloud, on GCP or AWS: Glean's Customer Hosted model deploys a tenant that Glean operates as a managed service, keeping upgrades, patching, monitoring and support access. No on-premises or Azure option is documented. Dynamiq self-hosts on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. ### Does Dynamiq enforce source permissions the way Glean does? For Google Drive and Confluence sources, yes: Dynamiq's permission syncing enforces each file's or page's permissions at retrieval time, so a user only gets content they can open in the source. Glean inherits and enforces source permissions across its connectors and syncs permission changes in real time, which matters most when the goal is one index over every app. ### Is Glean open source? No. Glean's platform is closed; it publishes API clients, a Web SDK and an Indexing SDK for building on it. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### How is Glean priced compared with Dynamiq? Glean publishes no price list. It licenses per-user, per-month seats, with model usage bought on an upfront commit under Glean Core Suite or drawn from pooled FlexCredits under Enterprise Flex, per its docs. Dynamiq offers a free plan to start and custom Enterprise pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. ## Sources - [Glean homepage](https://www.glean.com/) - [Glean enterprise search](https://www.glean.com/enterprise-search) - [Glean connectors](https://www.glean.com/platform/connectors) - [Glean Assistant](https://www.glean.com/ai-assistant) - [Glean Agents](https://www.glean.com/ai-agents) - [Glean Agent Builder](https://www.glean.com/ai-agents/agent-builder) - [Glean agent governance](https://www.glean.com/ai-agents/agent-governance) - [Glean Protect](https://www.glean.com/platform/security) - [Glean Model Hub](https://www.glean.com/platform/model-hub) - [Glean AI Gateway](https://www.glean.com/platform/ai-gateway) - [Glean APIs](https://www.glean.com/platform/api) - [Glean developer platform](https://developers.glean.com/) - [Glean deployment models (docs)](https://docs.glean.com/get-started/prepare/about-deployment) - [Glean AWS deployment guide (docs)](https://docs.glean.com/get-started/prepare/self-hosted-deployment/aws/deploy-aws) - [Glean supported LLMs (docs)](https://docs.glean.com/administration/llms) - [Glean SSO (docs)](https://docs.glean.com/administration/identity/sso/about) - [Glean confirmations for write tools (docs)](https://docs.glean.com/tools/human-in-the-loop-experience-for-tools) - [Glean AI Security (docs)](https://docs.glean.com/administration/protect/ai-security/introduction) - [Glean MCP server (docs)](https://docs.glean.com/administration/platform/mcp/about) - [Glean in Slack (docs)](https://docs.glean.com/administration/platform/embedded-integrations/slackbot/) - [Glean in Microsoft Teams (docs)](https://docs.glean.com/administration/platform/embedded-integrations/glean-in-teams/glean-in-microsoft-teams--admin-guide) - [Glean agent quality evaluation (docs)](https://docs.glean.com/agents/agent-development-lifecycle/evaluate-agent-quality) - [Glean Core Suite pricing (docs)](https://docs.glean.com/glean-core-suite-pricing) - [Glean Enterprise Flex pricing (docs)](https://docs.glean.com/glean-enterprise-flex-pricing) - [Glean release notes, September 15, 2026](https://docs.glean.com/release-notes/releases/2026-09-15-september-release) - [Glean Enterprise Agent Development Lifecycle (press release, May 2026)](https://www.glean.com/press/glean-introduces-the-enterprise-agent-development-lifecycle-codifying-how-enterprises-build-govern-and-measure-ai-agents) - [Glean adds NVIDIA Nemotron 3 Ultra (press release, June 2026)](https://www.glean.com/press/glean-adds-support-for-nvidia-nemotron-3-ultra-expanding-model-choice-for-cost-effective-enterprise-ai) - [Glean Partner Network (press release, August 2026)](https://www.glean.com/press/glean-launches-global-partner-network-to-scale-its-growing-enterprise-ai-ecosystem) - [Glean:GO 2026 announcements (press release, August 2026)](https://www.glean.com/press/glean-takes-on-enterprise-ais-biggest-bottlenecks-context-gaps-rising-costs-and-ai-sprawl) --- # Dynamiq vs Flowise URL: https://www.getdynamiq.ai/compare/flowise Flowise was an open-source visual builder for chatbots, agents and multi-agent workflows. Its team announced on July 29, 2026 that it was winding down and froze development that day, archived the GitHub repository on August 13, and reached end of life on August 31, 2026; the Apache 2.0 source stays available to self-host or fork, but npm packages and Docker images are marked deprecated. Dynamiq is an actively developed, governed platform for chat, voice and workflow agents, with a deployment team. Teams with a working Flowise install can keep it running on the archived source while they plan a move; teams choosing a platform going forward fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Flowise | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Not a fit for new production deployments: Flowise froze development on July 29, 2026 and reached end of life on August 31, 2026, per its sunset page. The Apache 2.0 source remains for teams that already run it or want to fork it. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Self-hosted from npm, Docker, Docker Compose or Kubernetes with Helm, with guides for AWS, Azure, GCP and DigitalOcean. Its homepage still lists Flowise Cloud tiers, and the sunset page does not say what happens to them. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes, and it is the only path forward: the source stays visible on GitHub indefinitely, but issues and pull requests are locked and npm packages and Docker images are marked deprecated, per Flowise's sunset page. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Apache License 2.0, except an enterprise directory under a commercial license, per its GitHub license file; the repository was archived on August 13, 2026. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | 100+ LLMs, embeddings and vector databases through its integrations, including OpenAI, Azure OpenAI, Anthropic, AWS Bedrock, Google Vertex AI, Mistral, Groq and Ollama. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Chatflows for chatbots and single agents with tools and RAG, and Agentflows for multi-agent workflows; no voice agent product, and with the project wound down, no new capabilities are coming. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A visual flow builder with a REST API, Python and TypeScript SDKs, a CLI, an embeddable chat widget and an MCP client; the archived repository receives no further updates. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | A Human Input node for approve or reject steps, moderation nodes, workspaces with role-based access on Cloud and Enterprise plans, and SSO with Microsoft, Google or Auth0 on its Enterprise plan. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Evaluations with text, numeric and LLM-based evaluators on Cloud and Enterprise plans, tracing integrations such as LangSmith, Langfuse, Arize and Phoenix, and Prometheus, Grafana and OpenTelemetry metrics. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | Not publicly documented. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | None going forward: official core team presence in Discord and GitHub ended at end of life on August 31, 2026, per its sunset page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Its homepage still lists cloud tiers (Free, Starter at $35 a month and Pro at $65 a month), but the sunset page leaves Flowise Cloud's future unstated; the source itself is free under Apache 2.0. | ## Choose Dynamiq when - Actively developed and supported: one engine for chat, voice and workflow agents, with a roadmap rather than a frozen codebase. - Guardrails, approvals, evals (including real-time) and observability with AI issue discovery are part of the same platform, not bolted on. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Flowise when - Teams with a working Flowise deployment can keep running it on the Apache 2.0 source while they plan a migration. - Teams willing to fork and maintain the codebase themselves keep a familiar visual builder with LangChain and LlamaIndex integrations. - The archived source is a complete reference for existing chatflows and agentflows during a migration. ## Migration Chatflows and Agentflows built in Flowise are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK, since the archived repository receives no updates. Tool nodes and API integrations reconnect as Dynamiq integrations or MCP servers, document stores and vector store configuration are re-ingested into Dynamiq Knowledge, and Human Input steps become Dynamiq approvals. Because Flowise reached end of life on August 31, 2026, teams still running it are migrating off a fixed, unmaintained version rather than a moving target. ## Questions and answers ### Is Flowise still actively maintained? No. Flowise announced on July 29, 2026 that it was winding down and stopped feature development and pull request reviews that day, archived its GitHub repository on August 13, 2026, and reached end of life on August 31, 2026, per its sunset page and GitHub. Its Apache 2.0 source remains on GitHub. ### Can I still self-host Flowise after its end of life? Yes. Flowise's sunset page says the source stays visible indefinitely and encourages teams to fork it, but issues and pull requests are locked and npm packages and Docker images are marked deprecated, so any fixes are up to you or a community fork. ### Why did Flowise wind down? Its sunset page says developers increasingly rely on coding agents for complex tasks, and that rigid low-code workflows quickly hit their limits. Dynamiq pairs a visual canvas with an open-source Python SDK on one engine, so teams can build visually, in code or both. ### Is Dynamiq actively developed and supported? Yes. Dynamiq is under active development, with forward-deployed engineers who ship your first production agent with your team and hand it over. Dynamiq holds SOC 2 and CASA Tier 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. ### What happens to a Flowise deployment if I migrate to Dynamiq? Flows, tools and knowledge sources move to Dynamiq's Agent Builder, integrations or MCP servers, and Dynamiq Knowledge respectively; the original Flowise source remains available under Apache 2.0 if you need to reference it during the migration. ## Sources - [Flowise homepage (sunset notice and cloud tiers)](https://flowiseai.com/) - [Flowise sunset announcement](https://flowiseai.com/sunset) - [Flowise GitHub repository (archived August 13, 2026)](https://github.com/FlowiseAI/Flowise) - [Flowise GitHub license](https://github.com/FlowiseAI/Flowise/blob/main/LICENSE.md) - [Flowise sunset timeline on GitHub](https://github.com/FlowiseAI/Flowise/discussions/6727) - [Flowise docs](https://docs.flowiseai.com/) - [Flowise deployment (docs)](https://docs.flowiseai.com/configuration/deployment) - [Flowise evaluations (docs)](https://docs.flowiseai.com/using-flowise/evaluations) - [Flowise SSO (docs)](https://docs.flowiseai.com/configuration/sso) --- # Dynamiq vs Langflow URL: https://www.getdynamiq.ai/compare/langflow Langflow is an open-source, MIT licensed, Python-based low-code builder for agentic and RAG applications that can also publish flows as MCP servers, running as a desktop app, in Docker or on Kubernetes, with guardrails, human-in-the-loop steps and built-in traces. Dynamiq is a full governed platform for chat, voice and workflow agents, with an Apache-2.0 SDK, built-in evals, inbound voice agents and a deployment team. Python developers who want a free, extensible visual builder fit Langflow; teams that need voice agents, built-in evals, enterprise SSO and a deployment team fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Langflow | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Python developers who want an open-source, low-code way to build and deploy agents, RAG applications and MCP servers, and to customize any component in Python. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Self-hosted: Langflow Desktop for macOS and Windows, a Python package, Docker or Docker Compose, and a headless runtime on Kubernetes with Helm. Its homepage also mentions a free cloud account without naming who runs it. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. Langflow runs locally, in a container or on your own Kubernetes cluster, per its documentation. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Open source under the MIT license, per its GitHub repository; IBM lists Langflow as a product of DataStax, an IBM company. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Many providers through its language model component and bundles, including OpenAI, Anthropic, Google Gemini, IBM watsonx.ai, Azure, Mistral, Groq, NVIDIA and Ollama, plus any OpenAI-compatible provider. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Chat apps, agents and multi-agent flows, with agent-to-agent (A2A) support and human-in-the-loop steps; its voice mode is deprecated as of Langflow 1.10, and no voice agent product is offered. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A visual editor where any component can be customized in Python, flows exported as JSON, run and workflow APIs with AG-UI streaming, an OpenAI-compatible endpoint, MCP as server and client, a CLI, a TypeScript client and an embeddable chat widget. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | A Guardrails component that checks for PII, credentials, jailbreaks and prompt injection, human-in-the-loop checkpoints, API keys, and external authentication through an OIDC proxy or SSO gateway; full SSO and RBAC enforcement are plugins that open-source Langflow does not include. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Built-in traces with latency and token usage, OpenTelemetry export (Langflow 1.12), and integrations such as LangSmith, Langfuse and Arize; no built-in evaluation suite is documented. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | Not publicly documented. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Professional services and premier support on request through Langflow's contact page; no named engineering engagement is described. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Free and open source to self-host; no cloud pricing is published. | ## Choose Dynamiq when - One engine for chat, voice and workflow agents, including Voice Agents for inbound calls, not only a builder for flows. - Built-in evals: datasets from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments. - Guardrails and approvals are part of the platform, and SSO (OIDC with Okta and Microsoft Entra ID) comes with Enterprise, rather than as plugins you source separately. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Langflow when - A free, MIT licensed builder with a very large community, where any component can be rewritten in Python. - A lightweight start as a desktop app, a Python package or a Docker container, with flows published as MCP servers for tools such as Claude Desktop or Cursor. - Frequent releases, recently adding guardrails, human-in-the-loop, A2A and OpenTelemetry, from DataStax, an IBM company. ## Migration Flows built in Langflow export as JSON, which maps directly onto agents and workflows rebuilt in Dynamiq's Agent Builder, on the canvas or in the Python SDK. Custom Python components move to SDK functions or MCP servers, and knowledge bases and vector stores are re-ingested into Dynamiq Knowledge. Prompts, model settings and component configuration carry over as the starting point for each agent. ## Questions and answers ### Is Langflow open source like Dynamiq's SDK? Yes. Langflow is open source under the MIT license, and Dynamiq's Python SDK is open source under Apache-2.0, per each project's GitHub repository. The Dynamiq platform around the SDK is not open source. ### Does Langflow support voice agents? No. Langflow's voice mode is deprecated as of version 1.10, per its documentation, and it offers no voice agent product. Dynamiq includes Voice Agents for inbound calls on the same platform. ### Can Langflow run fully self-hosted? Yes. Langflow runs as a desktop app, a Python package, a Docker container or a headless runtime on Kubernetes, per its documentation. ### Does Langflow include guardrails, SSO and evals? Langflow documents a Guardrails component, human-in-the-loop steps and built-in traces; full SSO and RBAC enforcement are plugins that open-source Langflow does not include, and no built-in evaluation suite is documented. Dynamiq includes guardrails, approvals and evals in one platform, with SSO on Enterprise. ### Does Dynamiq offer the same flow-building experience as Langflow? Dynamiq's Agent Builder is a visual canvas on the same engine as its open-source Python SDK, so developers can build visually, in code or both, and it can generate a workflow from a prompt. ## Sources - [Langflow homepage](https://www.langflow.org/) - [Langflow docs](https://docs.langflow.org/) - [Langflow on GitHub](https://github.com/langflow-ai/langflow) - [Langflow GitHub license (MIT)](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - [Install Langflow (docs)](https://docs.langflow.org/get-started-installation) - [Deploy Langflow on Kubernetes (docs)](https://docs.langflow.org/deployment-kubernetes-prod) - [Langflow guardrails (docs)](https://docs.langflow.org/guardrails) - [Langflow human-in-the-loop (docs)](https://docs.langflow.org/human-in-the-loop) - [Langflow external authentication (docs)](https://docs.langflow.org/external-authentication) - [Langflow authorization (docs)](https://docs.langflow.org/authorization) - [Langflow traces (docs)](https://docs.langflow.org/traces) - [Langflow MCP server (docs)](https://docs.langflow.org/mcp-server) - [Langflow voice mode (docs)](https://docs.langflow.org/concepts-voice-mode) - [Langflow 1.12 release](https://www.langflow.org/blog/langflow-1-12) - [DataStax, an IBM company](https://www.ibm.com/products/datastax) --- # Dynamiq vs Sierra URL: https://www.getdynamiq.ai/compare/sierra Sierra is a closed platform for customer experience AI agents across chat, SMS, WhatsApp, email, voice and ChatGPT, built with Ghostwriter, the no-code Agent Studio or its proprietary Agent SDK, and sold with outcome-based pricing and an expert agent development team. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 SDK, covering internal and back-office work as well as customer conversations, that runs inside your environment with forward-deployed engineers. Enterprises focused on customer experience that want outcome-based pricing fit Sierra; teams that need self-hosting, an open-source engine or internal agents fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Sierra | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Enterprises that want customer experience agents built and run with Sierra under outcome-based pricing, in industries such as financial services, healthcare, telecom, media, travel, retail and technology. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Not publicly documented. Sierra's pages describe a hosted platform across chat, SMS, WhatsApp, email, voice and ChatGPT, and list FedRAMP High, without stating self-hosted or on-premises options. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Not publicly documented. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed platform, and its Agent SDK is proprietary; Sierra's public GitHub holds open-source benchmarks such as tau-bench, not platform code. Each customer's agent code lives in a Git repository the customer can access and export. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | No customer model choice is stated. Sierra runs on a constellation of frontier, open-weight and proprietary models and switches between LLM providers automatically. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Customer-facing agents across chat, SMS, WhatsApp, email, voice and ChatGPT, plus Horizon (July 2026) for long-running goals such as originating a loan or getting prior authorization for a procedure. Internal or employee agents are not described. | | Voice | Inbound calls only, over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. | Inbound and outbound calls, with models tuned across 60+ locales, and Voice Sims that test languages, emotional states, background noise and interruptions. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Ghostwriter builds agents from SOPs, transcripts, audio or a plain-language goal; Agent Studio edits them without code; and the Agent SDK uses a declarative programming language inside your existing development lifecycle. | | Memory | Session memory scoped by user and session, plus long-term memory, where the agent decides when to save and recall durable, user-scoped facts that people can see and correct. Knowledge adds ingestion pipelines, permission-aware retrieval filters and knowledge graphs. | A Context Engine gives agents a memory layer the customer owns: permissioned, retained on the customer's terms and exportable to a data warehouse or another system. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Supervisory layers around the models to reduce hallucinations and prevent abuse, PII that is automatically encrypted and masked, PCI-certified infrastructure for payment data, and guardrails that set where human sign-off is required. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Simulations and regression test suites before release, Voice Sims, observability over tool calls, knowledge lookups and latency, Monitors and Pulse to surface issues, and multivariate Experiments. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | 40+ pre-built integrations plus custom integrations for proprietary systems; agents also work with ChatGPT Apps through MCP. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2, HIPAA, GDPR, PCI, FedRAMP High, CCPA, CSA STAR, ISO 27001 and ISO 42001 on its trust and reliability page, the EU AI Act among its homepage badges, and AIUC-1 certification (September 2026). | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | An expert agent development team that, per Sierra, has supported hundreds of deployments, with agent engineers and strategists who work directly with customers. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Outcome-based: customers pay for business outcomes delivered rather than tokens, per Sierra; rates are not published. | ## Choose Dynamiq when - One engine for chat, voice and workflow agents, covering internal and back-office work as well as customer conversations. - The agent engine is an open-source Python SDK under Apache-2.0, so the code your agents run on is yours to read, audit and fork. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - Model choice stays with you: 29 model providers in the SDK, the most-used models behind one OpenAI-compatible endpoint, plus open models on vLLM and LoRA fine-tuning. ## Choose Sierra when - Ghostwriter builds production agents directly from SOPs, call transcripts, audio or a plain-language goal, and Agent Studio lets CX teams edit them without code. - Outcome-based pricing tied to business results, for buyers who prefer it to seat or usage pricing. - Outbound calling, long-running customer journeys with Horizon, and an agent development team that builds alongside you. ## Migration Journeys, policies and personas built in Sierra, whether with Ghostwriter, Agent Studio or the Agent SDK, are rebuilt as agents and workflows in Dynamiq's Agent Builder; because Sierra agent code lives in a Git repository customers can export, it is a practical reference for the rebuild. Channels reconnect through AI Coworker, Voice Agents for inbound calls, and Dynamiq integrations or MCP servers, and knowledge sources move to Dynamiq Knowledge. Outbound calling has no direct equivalent, because Dynamiq voice agents are inbound only. ## Questions and answers ### Is Sierra open source? No. Sierra's platform and Agent SDK are proprietary; its public GitHub holds research benchmarks. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Can Sierra run self-hosted or on-premises? Not publicly documented; Sierra's site describes a hosted platform without stating a self-hosted or on-premises option. Dynamiq self-hosts on your own cloud or on-prem. ### How is Sierra priced? Sierra uses outcome-based pricing, where customers pay for business outcomes delivered rather than tokens, without published rates. Dynamiq offers a free plan to start and custom Enterprise pricing. ### Does Sierra handle outbound calls? Yes. Sierra's channels page describes agents across inbound and outbound calls. Dynamiq Voice Agents handle inbound calls only. ### Does Dynamiq build agents from existing procedures or transcripts like Sierra's Ghostwriter? Partly. Dynamiq's Agent Builder can generate a workflow from a prompt, which you then refine on the canvas or in the Python SDK. Building an agent directly from uploaded SOPs, transcripts or audio, the way Ghostwriter does, is not a stated Dynamiq feature. ## Sources - [Sierra homepage](https://sierra.ai/) - [Sierra product](https://sierra.ai/product) - [Sierra: Trust and reliability](https://sierra.ai/product/trust-and-reliability) - [Sierra Ghostwriter](https://sierra.ai/product/ghostwriter) - [Sierra Agent Studio](https://sierra.ai/product/agent-studio) - [Sierra Agent SDK](https://sierra.ai/product/agent-sdk) - [Sierra Horizon](https://sierra.ai/product/horizon) - [Sierra Context Engine](https://sierra.ai/product/context-engine) - [Sierra Insights](https://sierra.ai/product/insights) - [Sierra channels](https://sierra.ai/product/channels) - [Introducing Horizon (July 2026)](https://sierra.ai/blog/horizon) - [Sierra AIUC-1 certification (September 2026)](https://sierra.ai/blog/sierra-achieves-aiuc-1-certification) - [Sierra on agent transparency and export](https://sierra.ai/blog/your-agent-laid-bare-and-why-it-matters) - [Sierra on voice AI and Voice Sims](https://sierra.ai/blog/what-is-voice-ai) - [Sierra research on GitHub](https://github.com/sierra-research) --- # Dynamiq vs Decagon URL: https://www.getdynamiq.ai/compare/decagon Decagon is a closed, hosted platform for customer-facing AI agents across web chat, SMS, WhatsApp, voice and email, including outbound calls, configured through Agent Operating Procedures that combine natural language and code, with simulations, A/B experiments and an always-on QA system called Watchtower. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 SDK that covers internal and back-office work as well as customer conversations, runs inside your environment, and comes with a deployment team. Enterprises focused on customer experience, including outbound campaigns, fit Decagon; teams that need self-hosting, an open-source engine or agents beyond customer conversations fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Decagon | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Enterprises that want one customer-facing agent platform across chat, voice and email, with built-in simulations, experiments and quality monitoring, in industries such as financial services, travel, health, retail and telecom. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Hosted by Decagon on multi-region infrastructure, per its security page, and available to buy through AWS Marketplace since July 2026; no self-hosted or on-premises option is described on its site. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Not publicly documented; Decagon describes only its hosted platform. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed source; no open-source repository or SDK is published. Decagon offers APIs, MCP support, and Git-based versioning for its Agent Operating Procedures. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | No customer-facing model choice is described. Decagon says over 80% of its model traffic runs on models it trained itself (March 2026), and it enforces zero data retention with AI providers such as OpenAI and Anthropic. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Customer-facing agents across web chat, SMS, WhatsApp, voice and email, plus outbound campaigns such as payment reminders by SMS or call, a copilot for human representatives, and browser actions that operate internal systems during a conversation. Internal or back-office agents outside customer conversations are not described. | | Voice | Inbound calls only, over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. | Inbound and outbound calls in more than 70 languages with automatic language detection, SIP trunking into existing telephony, and voice authentication for callers, per Decagon's voice, integrations and security pages. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Agent Operating Procedures that mix natural language and code, versioned in Git, with Duet, an AI partner for building and improving agents; simulations run in CI/CD pipelines, and APIs and MCP connect tools. No general-purpose SDK or CLI is published. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | SSO through identity providers such as Okta and Microsoft Entra, role-based access, short-lived scoped tokens for agent access to customer systems, tamper-protected audit logs, bad actor detection, a supervisor that revises hallucinated responses before they send, and automatic redaction of sensitive data in logs and transcripts after a conversation ends. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Simulations that test agents before deployment and with every update, A/B experiments on live traffic against a control group, Watchtower QA that reviews every AI and human conversation against custom criteria, and reporting with generative analytics. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | Pre-built integrations such as Salesforce, Zendesk, Intercom, Kustomer, Amazon Connect and RingCentral, plus APIs and MCP; no total count is published. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | Its security page shows SOC 2, HIPAA, GDPR, CCPA, ISO, PCI and EU AI Act badges, and a September 2026 product post states the platform meets PCI, SOC 2, HIPAA, GDPR and CCPA. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | No named engineering engagement. Decagon says its tools remove the need for lengthy configuration cycles and large forward-deployed engineering teams, and it runs Decagon University certifications. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Not published on Decagon's site, which routes buyers to a demo; it is also sold through AWS Marketplace. | ## Choose Dynamiq when - One engine for chat, voice and workflow agents, covering internal and back-office work as well as customer conversations. - The agent engine is an open-source Python SDK under Apache-2.0, so the code your agents run on is yours to read, audit and fork. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - Model choice stays with you: 29 model providers in the SDK, the most-used models behind one OpenAI-compatible endpoint, plus open models on vLLM and LoRA fine-tuning. ## Choose Decagon when - Customer experience is the whole scope, including outbound calls and SMS or WhatsApp campaigns such as payment reminders, which Dynamiq's inbound-only voice agents do not cover. - A built-in quality loop for customer conversations: simulations in CI/CD, A/B experiments on live traffic and Watchtower QA across AI and human agents. - CX teams define agent behavior in Agent Operating Procedures without waiting on engineering, and can buy through AWS Marketplace. ## Migration Agent Operating Procedures and conversation designs built in Decagon are rebuilt as agents and workflows in Dynamiq's Agent Builder; procedures versioned in Git give a clear starting point for instructions and tool calls. Chat and email channels reconnect through Dynamiq integrations or MCP servers, inbound calls move to Dynamiq Voice Agents over SIP, and knowledge sources are re-ingested into Dynamiq Knowledge. Outbound calling and campaign flows have no direct equivalent, because Dynamiq voice agents are inbound only. ## Questions and answers ### Is Decagon open source? No. Decagon publishes no open-source repository or SDK; it offers APIs and MCP support for integrations. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Is Decagon SOC 2 and HIPAA compliant? Decagon's security page shows SOC 2, HIPAA, GDPR, CCPA, ISO, PCI and EU AI Act badges, and a September 2026 product post states the platform meets PCI, SOC 2, HIPAA, GDPR and CCPA. Dynamiq holds SOC 2 and CASA Tier 2, and signs a BAA for HIPAA workloads and a DPA under GDPR. ### Can Decagon run self-hosted or on-premises? Not per its site, which describes a hosted platform on multi-region infrastructure. Dynamiq self-hosts on your own cloud or on-prem. ### Does Decagon make outbound calls? Yes. Decagon's voice product initiates outbound calls for sales, support and customer updates, per its voice page. Dynamiq Voice Agents handle inbound calls only. ### Does Decagon publish pricing? Not on its own site, which routes buyers to a demo; it is also sold through AWS Marketplace. Dynamiq offers a free plan to start and custom Enterprise pricing. ## Sources - [Decagon homepage](https://decagon.ai/) - [Decagon security](https://decagon.ai/security) - [Decagon voice](https://decagon.ai/product/voice) - [Decagon chat](https://decagon.ai/product/chat) - [Agent Operating Procedures](https://decagon.ai/product/aop) - [Decagon Watchtower](https://decagon.ai/product/watchtower) - [Decagon testing and QA](https://decagon.ai/product/testing-qa) - [Decagon experiments](https://decagon.ai/product/experiments) - [Decagon integrations](https://decagon.ai/product/integrations) - [Decagon Duet](https://decagon.ai/product/duet) - [Introducing Decagon Labs (March 2026)](https://decagon.ai/blog/introducing-decagon-labs) - [Decagon on AWS Marketplace (July 2026)](https://decagon.ai/blog/decagon-and-aws-marketplace) - [Decagon Campaign Composer (September 2026)](https://decagon.ai/blog/campaign-composer) - [Decagon Assist (August 2026)](https://decagon.ai/blog/decagon-assist) - [Decagon Browser Actions (August 2026)](https://decagon.ai/blog/browser-actions) - [Decagon Series D announcement](https://decagon.ai/blog/series-d-announcement) --- # Dynamiq vs Wonderful URL: https://www.getdynamiq.ai/compare/wonderful Wonderful is a closed enterprise AI platform for agents that engage customers, assist employees and run business processes across voice, chat, email and documents, with an AI gateway, AI-native business systems, four deployment modes up to fully air-gapped on-premises, and forward-deployed teams it calls deployment pods. Dynamiq is a governed platform for chat, voice and workflow agents with an Apache-2.0 open-source SDK, published model and integration coverage, a free plan to start, and forward-deployed engineers; fully isolated and air-gapped environments are delivered with our engineers. Enterprises that want Wonderful's delivery model and business systems fit Wonderful; teams that want an open engine they can read and a platform they can start on themselves fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Wonderful | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Enterprises in industries such as banking, telecom, healthcare, utilities, insurance and retail that want agents and AI-native business software built with local teams, with leadership across 30+ markets, per Wonderful's site. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Multi-tenant SaaS, single-tenant, bring-your-own-cloud on AWS, Azure or GCP, or fully air-gapped on-premises, per Wonderful's deployment and product pages. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. Its on-premises option is described as completely air-gapped, with zero external dependencies and full on-site maintenance by Wonderful's engineers. | | Isolated environments | Fully isolated and air-gapped environments are delivered with our engineers. | Single-tenant (fully isolated, dedicated infrastructure) and on-premises (air-gapped) options are both described on its deployment page. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed; no open-source repository is published. An SDK is referenced on its AI Gateway page, and its documentation sits behind a sign-in. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | A separately named AI Gateway routes each request by task complexity, from smaller to heavier models, with role-based access to models and audit logs; specific providers or a count are not published. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Agents that engage customers, assist employees and execute business processes end to end across voice, chat, email and documents, plus Wonderful Systems: AI-native business software, such as CRM and HCM, designed, built and run by Wonderful. | | Voice | Inbound calls only, over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. | Voice is one of its stated channels, and its telecom page describes negotiating overdue payments across voice, WhatsApp and SMS. Telephony details, and inbound versus outbound specifics, are not published. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | A CLI and an SDK, reusable skills, a catalog of agents, skills and tools, versioning with rollback, and hand-offs between agents on different platforms through the A2A protocol. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Permissions down to the resource, role-based model access, an immutable audit trail, enterprise MCP gateways, prompt injection protection and agent guardrails; SSO specifics are not stated on the pages checked. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Evals and Observability are named platform components; A/B tests between agent variants and automatic conversation tags are described, with little further detail published. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2, ISO 27001, GDPR, PCI DSS and EU AI Act alignment on its product page, plus zero data retention with AI providers, optional automatic PII redaction and AES-256 encryption. HIPAA is not in that list, though its healthcare page calls the platform HIPAA-ready. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | Forward-deployed engineers and deployment strategists work inside the client's environment as a dedicated deployment pod, then move the client toward full ownership, per its deployment page. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Not published. Wonderful's site has no pricing page. | ## Choose Dynamiq when - The agent engine is an open-source Python SDK under Apache-2.0, so the code your agents run on is yours to read, audit and fork. - Coverage you can check before you buy: 29 model providers in the SDK, the most-used models behind one OpenAI-compatible endpoint, and 2,100+ app integrations plus MCP servers. - Start free at $0 and build on your own, then move to Enterprise for self-hosting, SSO and forward-deployed engineers. - Documented evals: datasets built from traces, LLM-as-a-judge, predefined and code metrics, real-time evaluations on live deployments, and voice simulations. ## Choose Wonderful when - Four deployment modes described in detail, up to fully air-gapped on-premises with on-site maintenance by Wonderful's own engineers. - Deployment pods of forward-deployed engineers and strategists, working inside your environment with local teams. - AI-native business systems, such as CRM and HCM, designed, built and run by Wonderful alongside its agents. ## Migration Agents built on Wonderful's platform are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK. Skills map to Dynamiq skills or workflow steps, tools reconnect as Dynamiq integrations or MCP servers, and knowledge moves to Dynamiq Knowledge. Wonderful Systems are business applications rather than agent flows, so replacing them is a separate systems decision; Dynamiq agents connect to the systems of record you choose. ## Questions and answers ### Is Wonderful open source? No. Wonderful publishes no open-source repository; an SDK is referenced on its site, with documentation behind a sign-in. Dynamiq's Python SDK is open source under Apache-2.0; the Dynamiq platform is not. ### Can both Dynamiq and Wonderful run fully air-gapped on-premises? Yes. Wonderful describes a completely air-gapped on-premises option with on-site maintenance by its engineers, per its deployment page. Dynamiq delivers fully isolated and air-gapped environments with our engineers. ### Does Wonderful support voice agents? Yes. Wonderful lists voice among the channels its agents work across, alongside chat, email and documents. Dynamiq Voice Agents answer inbound calls over WebRTC, SIP and WebSocket. ### Does Wonderful publish pricing? No. Wonderful's site has no pricing page. Dynamiq offers a free plan to start and custom Enterprise pricing. ### How many model providers does Dynamiq support compared with Wonderful? Dynamiq supports 29 model providers in its SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint. Wonderful's AI Gateway routes requests across models by task complexity but does not publish its providers or a count. ## Sources - [Wonderful homepage](https://www.wonderful.ai/) - [Wonderful product](https://www.wonderful.ai/product) - [Wonderful deployment](https://www.wonderful.ai/deployment) - [Wonderful agents](https://www.wonderful.ai/agents) - [Wonderful Systems](https://www.wonderful.ai/systems) - [Wonderful AI Gateway](https://www.wonderful.ai/ai-gateway) - [Wonderful safety](https://www.wonderful.ai/safety) - [Wonderful for healthcare](https://www.wonderful.ai/industries/healthcare) - [Wonderful for telecommunications](https://www.wonderful.ai/industries/telecommunication) - [About Wonderful](https://www.wonderful.ai/about-us) --- # Dynamiq vs Kore.ai URL: https://www.getdynamiq.ai/compare/kore-ai Kore.ai's Artemis, launched in May 2026, is a closed enterprise agent platform for customer and employee experience: voice and digital agents, long-running workflows and process automation, defined in its proprietary Agent Blueprint Language through visual and code authoring, deployable on public cloud, sovereign regions, private cloud or on-premises, with custom pricing. Dynamiq runs chat, voice and workflow agents on one engine, with an open-source Python SDK and forward-deployed engineers who hand the work over to your team. Enterprises that want a vertical-tuned CX and EX suite with outbound voice and contact-center connectors fit Kore.ai; teams that want an open-source engine they can read and a named deployment engagement fit Dynamiq. Facts checked 2026-09-29. | Dimension | Dynamiq | Kore.ai | | --- | --- | --- | | Best for | Enterprises in regulated industries that need chat, voice and workflow agents on one governed stack, deployed inside their own environment. | Enterprises, especially in banking, wealth management, insurance, healthcare, retail, telecom and travel, that want a vertical-tuned platform for customer experience and employee experience agents; Kore.ai cites more than 500 Global 2000 organizations. | | Deployment options | Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Fully isolated or air-gapped setups are delivered with our engineers. | Public cloud (multi-tenant SaaS), sovereign regions, private cloud (a dedicated single-tenant VPC) or on-premises, with data residency by region; its docs name Microsoft Azure, Google Cloud and AWS, and it is listed on the Azure and AWS marketplaces. | | Self-hosted or on-prem | Yes. Self-hosted on your cloud or on-prem, or Dynamiq Cloud; both run the same engine. | Yes. On-premises deployment is offered for organizations that require it, per Kore.ai's trust center and docs. | | License and source code | The Python SDK is open source under Apache-2.0 (github.com/dynamiq-ai/dynamiq), and the visual canvas runs on the same engine. The platform around it is not open source. | Closed platform. Agents are defined in Kore.ai's Agent Blueprint Language (ABL), a typed, schema-driven language; its public GitHub repositories hold client SDKs, not platform source. | | Model choice | 29 model providers in the SDK, the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. | Providers on its supported models page include OpenAI, Anthropic, Azure OpenAI, Microsoft Foundry, Google, AWS Bedrock, Mistral, Cohere and xAI, plus custom or self-hosted endpoints and fine-tuned models, with per-agent model overrides. | | Agent types (chat, voice, workflow) | Chat, inbound voice and workflow agents on one engine. AI Coworker works with its own cloud sandbox computer, in the browser, Slack, Microsoft Teams and Telegram; Voice Agents answer inbound calls; Agent Builder runs workflow agents. | Voice and digital agents for customer experience, employee experience agents for HR, IT and recruiting, and long-running workflows with human-in-the-loop steps, approval gates and durable execution, across web, voice, messaging, email, Microsoft Teams and API channels. | | Voice | Inbound calls only, over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. | Inbound and outbound calls (outbound added in September 2026), realtime speech models from several providers, connectors for contact centers such as Genesys, NICE CXone, Five9, Talkdesk and Amazon Connect, and automated quality reviews on every call. | | Developer surface | Agent Builder pairs a visual canvas with the open-source Python SDK on one engine, with Choice and rules nodes, workflows generated from a prompt, and versions with rollback. | Visual and code authoring of ABL in one workspace, an AI architect (Arch) that drafts ABL from business goals, a CLI, an MCP server that lets coding tools such as Claude Code, Cursor and Codex author agents, Git sync, REST APIs and a Web SDK for embedding. No open-source, general-purpose SDK. | | Governance (guardrails, approvals, SSO) | Guardrails that detect PII and prompt injection, policy detection, validators and approvals with editable fields on any tool step, on every plan; SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. | Real-time PII tokenization so models never receive plaintext PII, input guardrails, policies enforced outside the LLM, tenant isolation, immutable audit trails, SSO over SAML 2.0 or OpenID Connect on its Enterprise plan, custom roles, human-in-the-loop steps and bring-your-own-key encryption. | | Evals and observability | Evals with datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. | Evaluations built from scenarios, personas and variations with an LLM judge, which can run on a schedule; traces of every agent run across reasoning, tools, guardrails and handoffs; and Agent Insights that tie agent performance to cost and outcomes. | | Integrations | 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. | 300+ integrations and 40+ voice and digital channels, per its Artemis announcement (its docs count 27 channels), plus MCP tools and a Kore.ai Marketplace of prebuilt agents, templates and integrations. | | Compliance | SOC 2 and CASA Tier 2, with a BAA for HIPAA workloads and a DPA under GDPR. | SOC 2, ISO 27001, PCI DSS, GDPR, CCPA, the EU AI Act and DESC on its trust center; its Artemis announcement adds FedRAMP Moderate authorization, HITRUST and HIPAA alignment. | | Delivery help (engineers) | Forward-deployed engineers run a bootcamp in your environment, ship the first production agent with your team, then hand over. | No named engineering engagement is published; Kore.ai offers customer success, an Academy and systems integrator partners. | | Pricing model | Free to start ($0), and Enterprise with custom pricing. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. | Custom pricing by quote, through a sales conversation or an RFP; self-serve account sign-up exists, but no price list is published. | ## Choose Dynamiq when - The agent engine is an open-source Python SDK under Apache-2.0, so agents are defined in Python you can read, audit and fork, not in a vendor-specific language. - One engine for chat, voice and workflow agents, with guardrails, approvals, evals and observability built in from the start. - Runs inside your environment: Dynamiq Cloud, or self-hosted on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem, with isolated setups delivered with our engineers. - Forward-deployed engineers ship the first production agent with your team, then hand over. Your team keeps the code and the environment. ## Choose Kore.ai when - A vertical-tuned CX and EX suite with a long track record in banking, insurance, healthcare, retail and telecom. - Outbound voice campaigns and ready connectors into contact centers such as Genesys, NICE CXone, Five9, Talkdesk and Amazon Connect. - A long published certification list, including FedRAMP Moderate authorization and DESC, useful when a vendor review needs it today. ## Migration Agents and flows defined in Kore.ai's Agent Blueprint Language are rebuilt as agents and workflows in Dynamiq's Agent Builder, on the canvas or in the Python SDK; project exports from Kore.ai's API are a useful reference, but ABL is a Kore.ai language, so agents are rebuilt rather than imported. Integrations and MCP tools reconnect as Dynamiq integrations or MCP servers, knowledge sources are re-ingested into Dynamiq Knowledge, and inbound voice flows move to Dynamiq Voice Agents. Outbound campaigns have no direct equivalent, because Dynamiq voice agents are inbound only. ## Questions and answers ### Is Dynamiq open source, unlike Kore.ai's platform? Dynamiq's Python SDK is open source under Apache-2.0 on GitHub (github.com/dynamiq-ai/dynamiq); the Dynamiq platform around it is not. Kore.ai's Artemis platform is closed, with agents defined in its Agent Blueprint Language; its public GitHub repositories hold client SDKs. ### Can both Dynamiq and Kore.ai run on premises? Yes for both. Dynamiq self-hosts on your AWS, Azure, GCP, IBM Cloud, OpenShift or any Kubernetes 1.32+ cluster, or on-prem. Kore.ai states Artemis can deploy on public cloud, sovereign regions, private cloud or on-premises. ### Does Kore.ai support outbound calls? Yes. Kore.ai added outbound calling for voice agents in September 2026, per its release notes. Dynamiq Voice Agents handle inbound calls only. ### Does Kore.ai publish pricing? No. Kore.ai's site routes buyers to a sales conversation or an RFP and offers self-serve sign-up, without a price list. Dynamiq publishes a free plan to start and custom Enterprise pricing. ### Does Dynamiq offer the same breadth of prebuilt integrations as Kore.ai? Dynamiq connects to 2,100+ app integrations plus MCP servers; Kore.ai states 300+ integrations and 40+ channels in its Artemis announcement. Both figures come from each vendor's own materials. ## Sources - [Kore.ai homepage](https://www.kore.ai/) - [Kore.ai Trust Center](https://trust.kore.ai) - [Kore.ai launches Artemis (May 2026 announcement)](https://www.kore.ai/news/kore-ai-launches-artemis-the-new-generation-of-the-kore-ai-agent-platform-for-building-governing-and-optimizing-enterprise-ai) - [Artemis agent platform](https://www.kore.ai/ai-agent-platform) - [Kore.ai for customer experience](https://www.kore.ai/customer-experience) - [Kore.ai for employee experience](https://www.kore.ai/employee-experience) - [Kore.ai AI for Service (Marketplace)](https://www.kore.ai/ai-for-service) - [Kore.ai partners and marketplaces](https://www.kore.ai/partners) - [Kore.ai contact and RFP](https://www.kore.ai/contact-us) - [Agent Platform overview (docs)](https://docs.kore.ai/agent-platform/) - [Agent Blueprint Language (docs)](https://docs.kore.ai/agent-platform/abl) - [Supported models (docs)](https://docs.kore.ai/agent-platform/supported-models) - [Evaluations (docs)](https://docs.kore.ai/agent-platform/evals) - [PII protection (docs)](https://docs.kore.ai/agent-platform/pii-protection) - [Workspace, SSO and roles (docs)](https://docs.kore.ai/agent-platform/administration/workspace-and-team) - [Voice channels (docs)](https://docs.kore.ai/agent-platform/channels/voice) - [MCP server (docs)](https://docs.kore.ai/agent-platform/mcp-server) - [Artemis CLI (docs)](https://docs.kore.ai/agent-platform/cli) - [Agent Platform release notes](https://docs.kore.ai/agent-platform/release-notes/recent-updates) - [Kore.ai on GitHub](https://github.com/Koredotcom) --- # Agent memory URL: https://www.getdynamiq.ai/glossary/agent-memory The system that lets an AI agent recall earlier conversation turns or durable facts across separate runs, scoped to a user and session. Agent memory is what turns a stateless model call into a system that remembers. Short-term memory stores the messages of a conversation, scoped by a user identifier and a session identifier, and replays the relevant history at the start of the next turn. Long-term memory is a separate, durable store of facts about a user that an agent can write to and read from across every session, not just one conversation thread. In a regulated enterprise, memory is a data governance question as much as a product feature. Conversation history can contain account numbers, case details or health information, so where it lives, how long it is kept, and whether one caller's history can ever reach another caller matters as much as whether the agent sounds coherent across turns. A support agent that remembers a caller mentioned an order number earlier in the same call, but never carries that number into a different caller's session, is agent memory working correctly. Two sessions of the same user do not see each other's history either, by design. **In Dynamiq**, memory is configured on the Agent node with a choice of 8 memory backends, from the managed platform store to Postgres, DynamoDB, Qdrant, Pinecone and Weaviate, a save mode that controls how much of a run is persisted, and a retrieval strategy of the most recent messages, the most relevant ones, or both. A separate long-term memory store holds durable, user-scoped facts across sessions, which the agent updates rather than duplicates. --- # Agent orchestration URL: https://www.getdynamiq.ai/glossary/agent-orchestration The layer that decides which agent or step runs next, in what order and with what shared state, when a task needs more than one agent. Agent orchestration coordinates multiple agents or steps toward one outcome. One pattern is manager-led delegation, where a parent agent decides at run time which specialist agent to call, treating each one like a tool. The other is an explicit orchestrator, a graph of states and edges defined ahead of time, where the control flow is fixed and only the content varies. Enterprises that automate a real process, not just a single question, need orchestration to stay predictable and reviewable. A loan review or a claims workflow already has a defined sequence of steps in the business, and the automation should mirror that sequence rather than leave the order of operations to a single model's judgement on every run. A request comes in, an intake agent classifies it, a document agent extracts the relevant fields, and a policy agent checks them against the rules, each step handing off to the next in a fixed order. **In Dynamiq**, both patterns are available: attach an agent as a tool of another agent for delegation, with call limits and a setting that lets a specialist's answer pass straight back as the final one, or use the Graph Orchestrator to define states, edges and conditions explicitly. Every step, in either pattern, appears in the run's trace. --- # Agent sandbox URL: https://www.getdynamiq.ai/glossary/agent-sandbox An isolated, disposable virtual machine an agent can use to run shell commands, execute code and write files, separate from production systems. An agent sandbox gives an agent a real, isolated computer instead of only text in and text out. Inside it, the agent can install packages, write and run scripts, produce file artifacts such as spreadsheets or charts, and even start a web server to preview what it built. The machine is provisioned only when the agent needs it and is disposed of afterward. Running arbitrary, model-generated code is inherently risky, and a regulated enterprise cannot let that code touch the systems that hold customer data or run production workloads. A sandbox contains that risk in a disposable machine with no path back into the systems that matter. A research agent writes a script to analyze a spreadsheet and produce a chart, running the script inside the sandbox and returning only the finished chart, never touching the company's internal network. **In Dynamiq**, the sandbox is backed by E2B or Daytona, provisioned per run in your own provider account and isolated from Dynamiq's infrastructure and from every other run. It adds shell, file and preview tools to the agent automatically, environment variables passed to it are excluded from traces, and it is mutually exclusive with the lighter file store, since the sandbox already provides its own filesystem. --- # Agentic AI URL: https://www.getdynamiq.ai/glossary/agentic-ai AI systems that pursue a goal by planning steps, calling tools and acting on the results, rather than only generating a response to a single prompt. Agentic AI describes systems that act, not just answer. Given a goal, an agentic system breaks it into steps, chooses tools such as a database query, a search or an API call, observes what comes back, and decides what to do next until the goal is met or it needs a person. Generative AI produces content from a prompt; agentic AI puts that same model inside a loop that takes actions. An AI agent is one such system, and agentic AI is the broader approach, which can include several agents working in one workflow. The distinction matters most where actions have consequences. A drafted reply can be wrong without harm; an agent that updates an account, files a report or moves money needs limits on what it can reach, checks on what it produces, and a person at the points where a mistake is expensive. Agentic AI moves the governance question from what the model says to what the system does. A customer asks to move a card payment date. An agentic system looks up the account, checks the policy for allowed dates, makes the change through the card system's API and confirms it, handing off to a person if the request falls outside policy. **In Dynamiq**, agentic systems run on one platform: AI Coworker for everyday work with its own cloud sandbox, Voice Agents for inbound calls, and Agent Builder for workflows that mix agent steps with deterministic rules, approvals and guardrails, with every run traced. --- # Agentic memory URL: https://www.getdynamiq.ai/glossary/agentic-memory Memory an AI agent manages itself, deciding which facts to save, update or recall across sessions, instead of only replaying past conversation. Agentic memory gives the agent a say in what it remembers. Conversation memory replays earlier messages from the same session so a reply stays coherent. Agentic memory adds a long-term store the agent writes to and reads from on its own: it notices a durable fact, such as a customer's preferred language or an analyst's reporting format, saves it, and recalls it in a later session when it becomes relevant. Because the agent decides what to keep, the store has to stay clean and scoped. Facts must belong to one user so they never surface for another, near-duplicates should merge rather than pile up, and a stale fact needs to be replaceable. In a regulated organization, what an agent chose to remember is also data that retention and access rules have to cover, the same as any other record about a customer. A support agent learns in one call that a customer prefers a callback on their mobile number and saves it. In a new session the following week, it recalls the preference without being told again, and it never shows that fact to anyone else. **In Dynamiq**, memory combines session memory, the recent or most relevant messages of a conversation stored on a choice of 8 memory backends, with long-term, user-scoped facts the agent saves and recalls itself, updating a fact rather than saving a near duplicate. --- # Agentic workflow URL: https://www.getdynamiq.ai/glossary/agentic-workflow A business process modeled as a graph of steps where at least one step is an agent that reasons and chooses its own actions, not a fixed script. An agentic workflow mixes deterministic steps, a database lookup, a validation, a branch, with at least one agent step that plans and acts rather than following instructions written in advance for every case. The result is a process with the structure a regulated organization needs and the adaptability a purely scripted automation cannot offer. Traditional automation handles the predictable parts of a process well but breaks the moment a case does not fit the script. An agentic workflow keeps the fixed checkpoints a regulated process requires, intake, validation, sign-off, while letting the agent step handle the judgement calls in between. A document review workflow has fixed steps for intake and filing, with an agent step in the middle that reads a contract and decides which clauses need a closer look from a reviewer. **In Dynamiq**, the workflow canvas mixes deterministic nodes, such as choice, map and validator nodes, with agent nodes in a single graph, versioned on every save. Describe the workflow in plain language to get a first draft instead of a blank canvas, and every run of the finished workflow is traced end to end. --- # AI agent URL: https://www.getdynamiq.ai/glossary/ai-agent A system that pairs a large language model with tools, memory and a control loop so it can plan and carry out multi-step tasks toward a goal. An AI agent is more than a chatbot that answers one prompt. It runs a loop: reason about what to do next, take an action such as calling a tool, observe the result, and repeat until it has a final answer. That loop is what lets it look something up, take an action in another system, and adjust its plan based on what it finds, instead of only producing a single response from a single input. What an agent can actually do is defined by the tools it is given, not by the model underneath it. That is exactly the question a regulated enterprise needs answered before it approves one: which systems can this agent read from, which can it write to, and who reviews its decisions. An agent that reads an insurance claim, checks it against policy in a knowledge base, and drafts a decision for a person to approve is doing agent work. A single-turn chatbot answering a policy question is not. **In Dynamiq**, an agent is one node: a model, a role, a set of tools and, optionally, memory, running in a configurable loop with a maximum number of steps. The same agent is available on the visual canvas and in the open source Python SDK, one engine either way, and every run is traced from the first thought to the final answer. --- # AI agent platform URL: https://www.getdynamiq.ai/glossary/ai-agent-platform Software for building, deploying, governing and monitoring AI agents in one place, from models and tools to approvals, evaluations and traces. An AI agent platform is the layer between models and the business processes agents run. It covers the whole life of an agent: building it with a visual builder, an SDK or both, connecting it to models, knowledge and business systems, governing what it can do, deploying it to real channels, and measuring how it behaves once people rely on it. A framework gives developers building blocks; a platform adds the operations, security and governance a production agent needs. For an enterprise, the platform decision outlasts any single model. Models change every few months, while the controls, integrations, audit trail and deployment model are what a security review approves and what a team lives with. That is why evaluations usually center on where the platform runs, which models it can use, how approvals and guardrails are enforced, and whether the agents can move if the vendor relationship ends. A bank shortlisting platforms checks whether each one can run in its own cloud account, route only to approved models, require approval before any payment step, trace every run, and export its agents in an open format. **In Dynamiq**, one platform covers chat, voice and workflow agents: AI Coworker, Voice Agents, and Agent Builder with its visual canvas and Python SDK on one engine, with Knowledge, 2,100+ app integrations, 29 model providers and an AI Gateway, guardrails and approvals, evals and observability. It runs on Dynamiq Cloud or self-hosted, and the SDK is open source under Apache-2.0. --- # AI gateway URL: https://www.getdynamiq.ai/glossary/ai-gateway A single endpoint that routes chat completion requests to many model providers behind one credential, so applications do not hold provider keys. An AI gateway sits between an application and the model providers it calls, exposing one API and one credential no matter which provider actually serves the request. Instead of a client holding a key for each provider it might call, it holds one key for the gateway, and the gateway resolves it to the right provider on the server side. Centralizing model access this way gives a regulated enterprise one place to revoke access, one place to see usage across every model in use, and the ability to change which providers are even reachable without touching every application that calls a model. Switching an internal tool from one model to another becomes a one-line change to a model name, with no client code change, because both models are reached through the same endpoint and the same credential. **In Dynamiq**, the AI Gateway exposes an OpenAI-compatible chat completions endpoint for the most-used models, so existing OpenAI SDK code works by changing only the base URL and the key. The same gateway also parses and extracts structured data from documents and collects traces from workflows running outside the platform. --- # AI governance URL: https://www.getdynamiq.ai/glossary/ai-governance The policies, controls and evidence an organization uses to decide how AI may be used, who is accountable, and how its behavior is checked over time. AI governance turns principles into operating rules. It decides which uses of AI are allowed, who owns each system, what data it may touch, what review happens before launch, and how performance and risk are monitored afterward. Frameworks such as the NIST AI Risk Management Framework and laws such as the EU AI Act provide structure; governance is how an organization applies them to its own systems. For AI agents, governance has to reach past the model to the actions. A policy that says a person approves every payment only matters if the platform actually stops the payment step until someone approves it. The same holds for data access, logging and change control: governance that lives only in a document is not enforced when a model does something unexpected. A governance committee requires that any agent touching customer accounts screens input for prompt injection, gets approval before changing a record, keeps a replayable record of every run, and is evaluated again whenever its model or prompt changes. **In Dynamiq**, those requirements map to platform controls: SSO on Enterprise, organization roles and private projects for access, detectors and validators on input and output, approval gates on consequential steps, versioned deployments with rollback, every run traced and replayable, and real-time evals with an AI agent that surfaces issues from production traces. The platform supports a governance program; it does not certify a customer's compliance. Sources: [NIST, AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) --- # AI guardrails URL: https://www.getdynamiq.ai/glossary/ai-guardrails Checks placed around a model or agent, such as input screening and output validation, that flag or block unsafe behavior before it reaches anyone. AI guardrails are enforced checkpoints, not just instructions written into a prompt. On the input side, a detector screens a message before it reaches a model, for example for personal data or an attempt to override the system prompt. On the output side, a validator checks that what the model produced is well-formed before it is used, for example that it is valid JSON or matches an allowed set of values. Prompt instructions describe what a model should do, but nothing stops a model from ignoring them under the right pressure. A regulated enterprise needs checkpoints that hold regardless of what the model decides, and an auditable record of every check that ran. Before an agent's drafted email goes out, a validator confirms the message is well-formed and a detector confirms nothing in the conversation triggered an injection attempt, before the send step ever runs. **In Dynamiq**, guardrails are ordinary workflow nodes: detectors for personal data, prompt injection and policy violations, and validators for JSON, Python, allowed choices and regular expressions, placed anywhere in a workflow and branched on with a choice node. Every verdict is recorded in the run's trace, and a flagged branch can route to a refusal or to a person for review. --- # AI observability URL: https://www.getdynamiq.ai/glossary/ai-observability The practice of recording every step an agent or workflow takes so its behavior, cost and failures can be inspected after the fact. AI observability captures what happened during a run, not only what it produced. That includes which model answered, what it was asked, which tools it called and with what arguments, what came back, how many tokens were used, and how long each step took. For an agent that reasons across several steps, the final answer alone does not explain why it arrived there. A wrong or contested answer in a regulated workflow needs to be explainable after the fact. Being able to reconstruct exactly which document an agent retrieved, or which tool it called, is what turns a dispute into a short investigation instead of an unanswerable question. A support lead investigating an incorrect refund opens the run and finds the exact policy passage the agent retrieved and the tool call that issued the refund, rather than guessing at what happened. **In Dynamiq**, every run of a deployed App is recorded with monitoring charts for cost, tokens, requests and latency, and a full node-by-node trace browsable as a graph, a tree or a timeline, or downloadable as JSON. An agent that reviews production traces for issues and regressions runs alongside this, surfacing failures before a person has to go looking for them. --- # Air-gapped deployment URL: https://www.getdynamiq.ai/glossary/air-gapped-deployment A deployment with no network path to the public internet, used when data must never leave a physically or logically isolated environment. An air-gapped deployment has no route out to the public internet at all, not a firewalled connection, no connection. Everything the system needs, models, dependencies, updates, has to be brought in through a controlled process rather than fetched live from the network. This is a stricter requirement than a private cloud or an on-premises install that still reaches the internet for updates or provider APIs. Some regulated environments, defense and intelligence systems, certain government networks, the most sensitive banking workloads, require this level of isolation as a baseline, not an option. A vendor that can only offer a shared cloud service has no path into that kind of environment at all. The strictest form is a classified or otherwise physically isolated network that cannot exchange traffic with any outside system under any circumstances. Running an LLM air-gapped follows the same rule. A hosted model API is out of reach, so the model weights are brought in through the same controlled process and served inside the environment, which in practice means open-weight models on an inference server such as vLLM. Fine-tuning, evaluations and logs stay inside too. **In Dynamiq**, fully isolated and air-gapped environments are delivered with Dynamiq's engineers rather than offered as a self-service option, built on top of the standard self-hosted path: your own AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, or any Kubernetes 1.32+ cluster, or on-prem. --- # Approval gate URL: https://www.getdynamiq.ai/glossary/approval-gate A control on a specific workflow step that pauses execution until a person approves, edits or rejects it, every time that step runs. An approval gate is a hard, platform-enforced stop on one node, not a request the agent can choose to skip. Every time execution reaches that node, the run pauses, a reviewer sees exactly what is about to happen, and nothing proceeds until they approve, edit an allowed field, or reject it with feedback. Some actions in a regulated process, sending money, filing a report, emailing a customer, should never execute unattended, no matter how confident the agent is. An approval gate is how that rule is enforced by the system itself rather than left to a prompt instruction. Before a drafted client report is delivered, a reviewer sees the exact payload that would be sent, approves it as written, edits one allowed field, or rejects it with a note that flows back into the agent's next attempt. **In Dynamiq**, execution approval is enabled per node with a message template rendered from the node's own input, an optional list of fields a reviewer may edit, and a timeout after which the run checkpoints and waits, sometimes minutes, sometimes days, until a reply arrives. The decision, and any feedback, becomes part of the run's trace. --- # Barge-in and turn detection URL: https://www.getdynamiq.ai/glossary/barge-in-and-turn-detection How a voice agent knows a caller has stopped talking, and how it stops speaking immediately if the caller interrupts. Turn detection decides when a caller is done talking, so the agent knows it is safe to reply. Barge-in is the reverse: letting the caller interrupt the agent mid-sentence and having the agent stop immediately rather than talking over them. Getting either wrong reads as bad service on a call, an agent that jumps in too early or one that talks over an interruption. Perceived call quality is mostly timing, and a regulated contact center cannot have its voice agent cutting off a caller reading out a case number or a balance. Getting turn taking wrong is a customer experience problem with a compliance shadow: a caller interrupted before finishing a required disclosure or confirmation. A caller reading out a long account number needs a more patient pause allowance than a caller giving a quick yes or no answer. **In Dynamiq**, turn detection resolves in tiers: a speech-to-text model's own native end-of-turn detection where available, then a semantic detector that judges whether the sentence sounds finished, then silence-based detection as the fallback. End-of-speech silence and turn sensitivity are both tunable, interruption is on by default, and it is always on for realtime voice agents. --- # Chunking URL: https://www.getdynamiq.ai/glossary/chunking Splitting a document into smaller pieces before it is embedded and indexed, so retrieval returns focused passages instead of whole files. Chunking decides the unit of retrieval. A document can be split by character count, by word, by sentence, by page, by paragraph, or by heading, and each chunk keeps a link back to the document it came from. Smaller chunks match a query more precisely but lose surrounding context; larger chunks carry more context but dilute what any single vector represents. The chunking strategy, decided once when a knowledge base is set up, shapes every answer that base ever produces. A policy document chunked badly returns fragments that are technically related but not actually the answer, which is a worse failure mode in a compliance search than returning nothing. A long policy manual chunked by section heading returns one complete, citable rule when queried, instead of a paragraph that starts mid-sentence and ends before the rule's exception clause. **In Dynamiq**, Knowledge Bases default to a configurable splitter, by character, word, sentence, page, paragraph or title, with adjustable overlap between chunks, and structure- or meaning-aware splitters, a markdown header splitter, a semantic splitter, a recursive character splitter, are available when you customize the ingestion workflow directly. --- # Computer use agent URL: https://www.getdynamiq.ai/glossary/computer-use-agent An agent that operates a computer the way a person would, controlling a browser or desktop through clicks, typing and screenshots. A computer use agent interacts with software visually and by input, the same way a person does, rather than through a programmatic interface. It is the fallback for systems that were never built with an API: a legacy portal, a form-driven internal tool, a vendor site that only offers a human interface. Regulated back offices often run on exactly these systems, decades old, no API, still load-bearing. A computer use agent is frequently the only automation path available, which makes containing what it can reach as important as what it can accomplish. An agent logs into an older claims portal, fills out a form field by field, submits it, and downloads the confirmation page, one step at a time, the way an analyst would. **In Dynamiq**, browser automation is a tool that drives a real remote browser from natural-language instructions: navigate to a page, act on it, extract data from it, or list what is on it. A separate desktop tool gives full graphical control beyond the browser. Both run inside an isolated agent sandbox, never against production infrastructure directly. --- # Context window URL: https://www.getdynamiq.ai/glossary/context-window The maximum amount of text, measured in tokens, a model can read and generate in one call, including the prompt, history and output. Every model has a fixed context window: a limit on how many tokens of prompt, conversation history and generated output fit into a single call. Once a conversation or an agent's working history exceeds that limit, something has to give, older content gets dropped, truncated, or the call fails outright, depending on how the system is built. An agent that runs many tool calls, reads several documents and reasons across many steps fills its context window far faster than a single chat message would. If a system quietly drops the part of the context holding a policy constraint or an earlier instruction, that is not just a quality problem in a regulated workflow, it is a control that silently stopped applying. An agent running several web searches and reading multiple retrieved documents in one task can approach a smaller model's window well before the task is actually finished. **In Dynamiq**, an agent can enable context summarization: once token usage crosses a configurable share of the model's window, older history is compressed into a summary while the most recent messages are kept verbatim, and the original request is always pinned so a long run never loses track of the task it started with. --- # Conversational AI URL: https://www.getdynamiq.ai/glossary/conversational-ai Technology that lets people talk with software in natural language, by text or voice, and have it understand the request and respond or act on it. Conversational AI covers the systems behind chat assistants, voice bots and phone agents: understanding what a person says or types, keeping track of the conversation, and replying in text or speech. Earlier conversational systems matched messages to predefined intents and scripted flows; newer ones use large language models to understand open-ended requests and, as agents, call tools to look things up and take actions. The difference shows up in the contact center. An intent-based bot handles the questions it was scripted for and sends everything else to a queue. An agent can check an order, change a booking or reset a password in the same conversation, which is also why it needs rules for what it may change, a transcript of what it said, and a path to a person. A caller says, in their own words, that they were charged twice. A conversational agent finds both charges, explains that one is a pending authorization that will drop off, and offers a transfer to billing if the caller still wants to dispute it. **In Dynamiq**, conversational agents run over chat and voice on one platform. Voice Agents answer inbound calls over SIP, WebRTC or a WebSocket, with every turn's latency measured and simulations before go-live, and chat agents deploy as a widget or an API. Both use the same tools and knowledge, and every conversation is traced. --- # Data residency URL: https://www.getdynamiq.ai/glossary/data-residency A guarantee about which country or region a system stores and processes its data in, used to meet legal or contractual requirements. Data residency is a claim about geography: where data physically sits and where it is processed, not just who owns it. It usually exists to satisfy a specific legal requirement, a banking secrecy law, a data localization statute, a public sector rule that citizen data must stay within national borders, regardless of which region would otherwise be cheapest or fastest to run in. For a regulated enterprise, data residency is often a precondition for using a system at all, not a preference. A vendor that only offers one fixed region cannot serve a customer whose law requires a different one. A government agency required to keep resident records inside its own country's borders, no matter which cloud region a vendor's default deployment would otherwise use. **In Dynamiq**, self-hosting the platform inside your own AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, or any Kubernetes 1.32+ cluster, or on-prem, keeps your data and the models you host inside the location you choose, and any outside service, such as a hosted model, is one you choose to use. Fully isolated or air-gapped environments, where even that is not enough, are delivered with Dynamiq's engineers. --- # Document parsing URL: https://www.getdynamiq.ai/glossary/document-parsing Converting a file such as a PDF, scan or slide deck into clean, structured text, typically Markdown, that a model or index can use. Document parsing turns a file format into text a model can actually reason over, while preserving the structure that gives the content its meaning: headings, tables, lists, emphasis. That is different from plain OCR, which recovers characters but not the relationships between them, so a table becomes a row of numbers with no columns. A contract or a policy document is not just words, it is words in a specific structure, clause numbers, defined terms, tables of figures. Flattening that structure into a wall of text loses information a regulated reader, human or model, actually depends on. A scanned invoice is parsed into Markdown with its line-item table intact, so an amount can still be matched to the specific item it belongs to, instead of a stream of numbers with no columns to anchor them. **In Dynamiq**, Document Parse in the AI Gateway runs a vision-capable model over a PDF or image and returns Markdown that keeps headings, tables and lists, while skipping running headers, footers and blank pages, callable from a playground or a single API endpoint. --- # Embeddings URL: https://www.getdynamiq.ai/glossary/embeddings Numeric vectors that represent the meaning of a piece of text, so pieces with similar meaning end up close together in vector space. An embedding is a list of numbers produced by a model that has learned to place similar meanings near each other in that numeric space. Two sentences that use different words but mean nearly the same thing end up close together; two sentences that share words but mean different things can end up far apart. That is what makes semantic search possible: search by meaning, not by matching keywords. A regulated search over policies or case files written in inconsistent language needs to find the right passage even when the query and the document do not share vocabulary. Keyword search alone misses exactly the paraphrased or differently-worded version of the same rule. A query about client onboarding retrieves a passage titled new account setup, because the two are close together in embedding space even though they share almost no words. **In Dynamiq**, 8 embedding providers sit behind one interface, OpenAI, Cohere, Bedrock, Mistral, Gemini, Hugging Face, watsonx and Vertex AI, and the platform enforces the one rule that actually matters: whichever model embedded the documents at ingestion time is the same one used to embed the query at search time, automatically, inside a Knowledge Base. --- # EU AI Act URL: https://www.getdynamiq.ai/glossary/eu-ai-act The European Union's risk-based regulation of AI systems, Regulation (EU) 2024/1689, with obligations that scale to how risky a given use is. The EU AI Act classifies an AI system by risk rather than regulating AI as one category. Unacceptable-risk practices are banned outright. High-risk systems, generally those used in areas such as employment, credit, law enforcement or critical infrastructure, carry obligations around risk management, data governance, technical documentation, human oversight and conformity assessment. Limited-risk systems carry transparency duties, such as disclosing that a person is interacting with AI, and the obligation attaches to the provider or deployer of a specific system, not to AI as a category, so a bank or a hospital has to classify its own use case. The Act entered into force on August 1, 2024. Prohibitions on unacceptable-risk practices have applied since February 2, 2025, and obligations for general-purpose AI models since August 2, 2025. Transparency duties, disclosing AI interaction and labeling synthetic content, apply from August 2, 2026. High-risk system obligations were originally due in August 2026, but the Digital Omnibus on AI postponed them: standalone high-risk systems under Annex III now apply from December 2, 2027, and AI embedded in regulated products under Annex I from August 2, 2028. **In Dynamiq**, the platform does not certify that a customer's use of it complies with the Act. Toward these obligations, every run is traced and replayable, which supports record-keeping duties, approval gates put a named person in the loop before a consequential action, which supports human oversight, and self-hosted deployment keeps data and models inside infrastructure the customer already controls. Whether a specific system is high-risk, and what its provider or deployer must do about it, depends on the system and the organization's own assessment. Sources: [European Commission, AI Act regulatory framework](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) --- # Fine-tuning URL: https://www.getdynamiq.ai/glossary/fine-tuning Further training a pretrained model on a smaller, specific dataset so it performs better on a narrower task, tone or domain. Fine-tuning changes a model's own weights using examples from a specific task or domain, rather than changing what the model is given at the moment it answers. That makes it different from prompting or retrieval, which shape a single call without touching the model itself. The result is a model that has, in effect, practiced the exact kind of task you need repeatedly. A regulated enterprise that wants a model to consistently match its own approved tone, terminology or classification scheme often gets there faster with fine-tuning than with an ever-longer prompt, and without sending a stream of live examples to a third party's training process. Fine-tuning a smaller open model on a firm's own approved response templates so it drafts client-facing replies that already match house style, rather than prompting for that style on every single call. **In Dynamiq**, deploy and fine-tune open models, including Llama, Gemma, Mistral and Qwen, on the platform's own or your self-hosted Kubernetes infrastructure, then serve the resulting LoRA adapters dynamically over one shared base model, without a separate deployment for every adapter. --- # Forward-deployed engineer URL: https://www.getdynamiq.ai/glossary/forward-deployed-engineer An engineer who works inside a customer's own environment to build and ship the customer's first production use of a platform. A forward-deployed engineer builds a real, running system inside the customer's own environment, using the customer's own tools and data, rather than advising from the outside or shipping a generic demo. The work is measured by what goes into production, not by a report or a set of recommendations handed over at the end. A regulated organization is often unwilling, or unable, to hand its processes and data to an outside team without direct oversight. Having an engineer embedded, working alongside the customer's own developers inside the customer's own environment, changes both the risk profile of the engagement and how much the customer's team learns from watching the build happen. An engineer pairs with a bank's own developers to connect an agent to the bank's existing case management system and get it running in production, rather than presenting a proof of concept that never leaves a demo environment. **In Dynamiq**, forward-deployed engineers run a bootcamp inside the customer's own environment, define success criteria and evaluations before writing anything, ship the first production agent together with the customer's team, then hand over. What gets built along the way feeds back into the platform's templates and connectors for the next deployment. --- # GDPR URL: https://www.getdynamiq.ai/glossary/gdpr The EU's General Data Protection Regulation, Regulation (EU) 2016/679, governing how personal data of people in the EU and EEA is handled. The GDPR sets rules for processing the personal data of people in the EU and EEA: a lawful basis is required before processing can begin, individuals have rights to access, correct, delete and receive a copy of their data, and a breach that risks people's rights must generally be reported to the relevant supervisory authority within 72 hours of discovery. It applies regardless of where the organization doing the processing is based, so a vendor outside the EU handling an EU resident's data is still in scope. The regulation was adopted in 2016 and has applied since May 25, 2018. Fines can reach 20 million euros or 4 percent of an organization's global annual turnover, whichever is higher. For an enterprise buying an AI system, GDPR turns into a specific, checkable requirement: any vendor processing personal data on the company's behalf needs a signed Data Processing Agreement setting out what it does with that data and under what safeguards. A support agent that retrieves a customer's account details needs a documented lawful basis for that processing and a Data Processing Agreement covering the vendor whose system performs the retrieval. **In Dynamiq**, a Data Processing Agreement is available, and self-hosted deployment lets an organization keep data processing inside the region its own GDPR analysis requires. Sources: [European Commission, Data protection in the EU](https://commission.europa.eu/law/law-topic/data-protection/data-protection-eu_en) --- # GraphRAG URL: https://www.getdynamiq.ai/glossary/graphrag A retrieval method that answers questions by walking a knowledge graph of entities and relationships, not only by matching text similarity. GraphRAG retrieves facts by following relationships between entities rather than by matching text similarity alone. A vector search answers what a passage sounds like; a graph answers how two things are connected, by walking from a named entity across one or more relationships to the facts that matter. For a question that spans more than one hop, who reports to whom, which vendor is linked to which counterparty, a graph can reach an answer a similarity search never will. Compliance and investigations questions are frequently relationship questions: ownership chains, reporting lines, shared addresses between entities that should be unrelated. Text search alone tends to answer these poorly, because the connection that matters is not written down as a sentence anywhere. Answering what system does this person's employer use requires following a works-at relationship to an organization, then a uses relationship from that organization to a system, not a single similarity match against any one document. **In Dynamiq**, a Knowledge Base can build a graph alongside its vector store, with entity and relationship types defined in an editable ontology, retrieval bounded by a beam-search traversal so it does not explode across a highly connected entity, and access control enforced on the graph's own edges, so a caller only ever sees facts drawn from documents they are entitled to see. --- # Grounding URL: https://www.getdynamiq.ai/glossary/grounding Tying a model's answer to specific, retrievable source material, so the answer can be traced back to a document rather than training data. Grounding is the practice of giving a model source material to answer from, and having its answer stay tied to that material, rather than to whatever it happened to learn during training. A grounded answer can point to the passage it came from; an ungrounded one can only be checked by an expert who already knows the right answer. A regulated decision, a coverage determination, a policy interpretation, a compliance finding, needs a source a reviewer can actually check, not a plausible sentence with nothing behind it. Grounding is what makes an AI-generated answer reviewable rather than something to be taken on faith. An agent answering a question about a refund policy quotes the exact clause and the exact document it came from, so a reviewer can verify the answer in one click rather than re-researching the question from scratch. **In Dynamiq**, Knowledge Base retrieval and GraphRAG both return the source passages or facts an agent used alongside its answer, guardrail validators can check an output's structure before it reaches a user, and evaluation metrics such as Faithfulness and Context Recall score whether an answer actually reflects what was retrieved, rather than drifting from it. --- # Hallucination URL: https://www.getdynamiq.ai/glossary/hallucination An answer a model states with confidence that is not supported by its source material or by fact, presented as if it were true. A hallucination is not a model admitting uncertainty and guessing wrong, it is a model stating something false with the same confidence as something true, because it is predicting plausible text, not verified text. Nothing about how a hallucinated answer reads distinguishes it from a correct one. In a regulated workflow, a hallucinated citation, figure or policy clause does not announce itself. It can misinform a customer or a regulator, and by the time anyone questions it, the answer has already been acted on as if it were fact. An agent invents a policy clause number that does not exist anywhere in the actual document, phrased with exactly the same confidence as a real citation would carry. **In Dynamiq**, several layers work against this: guardrail validators check structure before an output ships, retrieval grounds answers in cited source passages instead of memory, and evaluation metrics, including an LLM-as-a-judge hallucination rubric and the predefined Faithfulness evaluator, score how well an answer matches what was actually retrieved, on demand before a release, or continuously against live traffic with online evaluations. --- # HIPAA URL: https://www.getdynamiq.ai/glossary/hipaa The US Health Insurance Portability and Accountability Act of 1996, whose Privacy and Security Rules govern protected health information. HIPAA governs protected health information handled by covered entities, health plans, providers and clearinghouses, and by their business associates, vendors that handle protected health information on a covered entity's behalf. A vendor acting as a business associate must sign a Business Associate Agreement setting out how it will safeguard that data, and the Department of Health and Human Services' Office for Civil Rights enforces the requirement. HIPAA was enacted on August 21, 1996. Covered entities have had to comply with the Privacy Rule since April 14, 2003, and with the Security Rule since April 21, 2005. Any AI system that reads, stores or summarizes patient information is handling protected health information. A clinical intake agent summarizing patient notes on a hospital's behalf is doing exactly the kind of processing a Business Associate Agreement is meant to cover, whether or not anyone thought of it as a HIPAA question at the time. **In Dynamiq**, HIPAA is supported with a signed Business Associate Agreement available, and PII detection can flag identifiable information before it reaches a model or leaves a workflow, so a healthcare team can see where protected health information enters a process before deciding how that step should be handled. Sources: [HHS.gov, Summary of the HIPAA Privacy Rule](https://www.hhs.gov/hipaa/for-professionals/privacy/laws-regulations/index.html), [HHS.gov, Summary of the HIPAA Security Rule](https://www.hhs.gov/hipaa/for-professionals/security/laws-regulations/index.html) --- # Human in the loop URL: https://www.getdynamiq.ai/glossary/human-in-the-loop A design pattern where a person reviews, confirms or supplies information at a defined point in an otherwise automated process. Human in the loop covers two shapes of the same idea. In one, the agent itself decides, based on its instructions, that it needs to ask a person something, a clarifying question, a confirmation, before it continues. In the other, a specific step is set up so that it always pauses for a person, regardless of what the agent decides, the stricter pattern usually called an approval gate. Regulated processes typically require a named, accountable person at specific points in a decision, and human in the loop is how that accountability gets built into the system's design rather than assumed after the fact. An agent drafts a client communication and asks a person to confirm the intent is right before it proceeds, and a separate, mandatory gate still holds the final send step regardless of that earlier confirmation. **In Dynamiq**, a Human Feedback tool lets an agent ask a question and wait for a reply, or send a status update without waiting, while execution approval on a node pauses that node every time it is about to run. Both pause the run, surface the prompt over a live connection, and resume, sometimes minutes or days later, from a checkpoint the moment a reply arrives. --- # Intelligent document processing URL: https://www.getdynamiq.ai/glossary/intelligent-document-processing Using AI to read unstructured documents such as scans, PDFs and forms and turn them into structured, validated data fields. Intelligent document processing goes a step past OCR: it classifies a document, extracts specific fields against a schema you define, and returns structured data rather than a block of recognized characters. The output is something a downstream system can use directly, a named field with a value, not a paragraph a person still has to read and re-key. Claims, applications, invoices and onboarding forms arrive as PDFs and scans in exactly the high-volume, high-stakes processes a regulated back office runs, and manually re-keying them is slow and error-prone at the volumes those processes handle. An invoice's vendor name, line items and total are extracted into fields that match a firm's own accounting schema, instead of a person retyping each value from a scanned page. **In Dynamiq**, Document Extract in the AI Gateway runs an OCR pass over a PDF or image, then a structured output pass that pulls the fields defined by a JSON template you write, returning data shaped exactly like that schema, callable from a playground or a single API endpoint. --- # Knowledge graph URL: https://www.getdynamiq.ai/glossary/knowledge-graph A database of entities and the relationships between them, such as a person working at an organization, used to answer connection questions. A knowledge graph stores facts as entities and the relationships connecting them, a person, an organization, a works-at relationship between them, rather than as blocks of text to search. Each fact carries provenance back to the document it was drawn from, so a graph is not just a diagram, it is a queryable, source-linked record. Regulated investigations, know-your-customer checks, ownership questions, are relationship questions by nature. Without a knowledge graph, an analyst reconstructs those connections by hand, reading document after document and keeping the map in their head. Which vendors share a registered address with this counterparty is a graph query across shared attributes and relationships, not something a keyword search over documents can answer directly. **In Dynamiq**, a Knowledge Base can extract entities and relationships into a graph alongside its vector store, constrained by an editable ontology, entity types, relationship types and the patterns allowed between them, so extraction stays on-schema rather than free-form. Every fact remains traceable to the document and the access list it came from. --- # KYC and AML automation URL: https://www.getdynamiq.ai/glossary/kyc-and-aml-automation Using software, including AI, to run the identity checks and transaction monitoring that anti-money-laundering and know-your-customer rules require. KYC, know your customer, is verifying who a customer is when a relationship begins. AML, anti-money laundering, is watching transactions afterward for patterns that suggest laundering, sanctions exposure or other financial crime, and reporting what is found. Automating either one means the decision and the evidence behind it both need to hold up, not just the output. These are among the highest-volume, most heavily regulated review processes in finance, and a regulator does not just want the right answer, it wants to see how the answer was reached and who signed off on it. Reconciling a batch of flagged transactions against a firm's own reporting policies and drafting the write-up an analyst used to produce by hand, then requiring a named reviewer's sign-off before anything leaves the firm. **In Dynamiq**, an agent can query case and transaction systems through tools, ground its policy reasoning in a knowledge base of the firm's own rules, run under per-analyst database credentials so it never widens data access beyond what that analyst already has, and route its draft through an approval gate before delivery, with every query and decision recorded in the run's trace. --- # LLM as a judge URL: https://www.getdynamiq.ai/glossary/llm-as-a-judge Using a large language model to score another model's output against a written rubric, in place of, or alongside, a human reviewer. An LLM-as-a-judge metric is a model given a written rubric and asked to score an answer against it, returning a structured score rather than free-form commentary. It is best suited to qualities that are hard to check with code: tone, relevance, completeness, whether an answer actually reflects its source material. Regulated teams often need to check far more agent responses for quality and policy adherence than a human review team can read one at a time. A judge model applying the same written rubric to every sampled response is what makes that review scale without lowering the bar on any single check. A rubric scores whether a support reply reflects the retrieved policy passage it was supposed to be grounded in, applied consistently to every conversation sampled for review, not just the ones a person happens to read. **In Dynamiq**, an LLM-as-a-judge metric is defined once, with a chosen judge model and a rubric written using placeholders such as question, answer and context, then reused across evaluation runs and online evaluations. Every score stays tied to the exact rubric version that produced it, so a later rubric change never rewrites history. --- # LLM evaluation URL: https://www.getdynamiq.ai/glossary/llm-evaluation Scoring an LLM or agent's outputs against test data with defined metrics, so changes in quality are measured rather than judged by feel. LLM evaluation is built from three pieces: a metric that defines how to score, a dataset that defines what to score against, and a run that puts the two together and produces results. Comparing a candidate version against what is already in production, on the same dataset, with the same metric, turns a subjective impression into a number you can actually compare. Regulated teams cannot ship a prompt or model change on a hunch. They need a documented, reviewable comparison of before and after, on a dataset that does not change out from under them. Comparing a new prompt against the current one on the same released dataset of real questions, scored by the same metric, before deciding whether to promote the new version. **In Dynamiq**, Evals combine versioned metrics, an LLM-as-a-judge rubric, a predefined RAG evaluator such as Faithfulness or Context Recall, or your own Python function, with versioned datasets you can build directly from production traces, and evaluation runs that score a workflow's fresh outputs or a dataset as recorded. Every run pins the exact metric and workflow version it used, so it can be rerun and trusted later. --- # LLM inference URL: https://www.getdynamiq.ai/glossary/llm-inference Running a trained model to produce an output for a given input, as opposed to training the model in the first place. Inference is what happens every time a model actually answers something: tokenizing the input, generating output tokens one at a time, and returning the result. It is distinct from training, which happens once, ahead of time, to produce the model's weights in the first place. Inference is also where ongoing cost and latency come from, since it scales with every call, not with a one-time training run. Where inference happens, on a shared public endpoint or on infrastructure a regulated enterprise controls, is as much a governance question as it is a performance one, alongside the more familiar concerns of cost and response time. An agent's every reasoning step and tool decision is a separate inference call, so one long agent run can mean many calls to a model before it produces a final answer. **In Dynamiq**, call the most-used models through one AI Gateway endpoint, use any of 29 model providers from the SDK, or run inference on your own infrastructure by self-hosting an open model on the vLLM engine behind an OpenAI-compatible endpoint, with autoscaling replicas and no call ever leaving the cluster it runs in. --- # LLM tracing URL: https://www.getdynamiq.ai/glossary/llm-tracing A recorded, node-by-node execution tree of one run, including every prompt, tool call, retrieved document and token count. A trace is the record of one run, not a summary of it: every node that executed, every tool call and its arguments, every document retrieved, every prompt sent to a model and the tokens it used, in order, with timing. Aggregate observability tells you a system is slow or expensive on average; a trace tells you exactly what one specific run did. When an outcome is challenged, a customer, a regulator, an internal reviewer, someone needs to reconstruct exactly what the system did on that one run, not a statistical picture of typical behavior. A disputed refund decision is reconstructed by opening the run's trace and reading the exact policy passage the agent retrieved and the exact tool call that issued the refund. **In Dynamiq**, every run of a deployed App is recorded as a trace, browsable as a graph, a tree or a timeline, with each node's input, output, timing, rendered prompt and token usage available individually, and downloadable as JSON on its own or in bulk. Workflows built with the open source SDK can ship the same traces into the platform with a single callback handler. --- # LoRA URL: https://www.getdynamiq.ai/glossary/lora Low-Rank Adaptation, a fine-tuning method that trains a small set of extra weights on top of a frozen base model, not the whole model. LoRA trains a small number of additional weights that sit alongside a base model's original, frozen weights, rather than updating every parameter in the model. That makes it far cheaper and faster to produce than full fine-tuning, and because the base model itself never changes, several LoRA adapters can be trained for different tasks and swapped in over the same underlying model. A regulated team that wants a model adapted to its own tone, terminology or a specific classification task, without the cost of training and separately hosting a whole model per use case, gets there with a LoRA adapter instead. Training one adapter for a firm's client-facing tone and a second, separate adapter for internal ticket triage, both served from the same base model rather than two full models. **In Dynamiq**, fine-tune LoRA adapters and serve them dynamically over a shared base model, with no separate deployment required per adapter. A request is routed to a specific adapter by name, on the same OpenAI-compatible endpoint the base model itself serves. --- # Model Context Protocol (MCP) URL: https://www.getdynamiq.ai/glossary/model-context-protocol An open protocol that lets an AI agent discover and call tools exposed by an external server over a standard interface. Model Context Protocol, MCP, defines how an agent discovers what tools a server offers and calls them, over a standard interface rather than a bespoke integration written for every tool. A server advertises its tools with a name, a description and an input schema; an agent asks what is available, then calls whichever tool the task requires, using the schema to know what to send. An enterprise that already runs internal tool servers, a case lookup service, an internal search index, can expose them to an agent once, over MCP, instead of writing a custom connector for every agent that needs them, and can control exactly which tools are visible to which agent. An internal MCP server exposing a case lookup tool is connected once and then reused by any agent in the organization that needs to look up a case. **In Dynamiq**, an MCP Server tool connects an agent to a remote MCP server over the current standard transport or the legacy one, discovers its tools at run time, and expands each into a typed agent tool automatically. Include and exclude lists control exactly which tools an agent can see, alongside a built-in catalog of 2,100+ app integrations. --- # Model routing URL: https://www.getdynamiq.ai/glossary/model-routing Automatically or programmatically directing a request to one of several possible models, based on name, cost, capability or availability. Model routing sends a request to whichever model should actually serve it, rather than hardwiring an application to one specific model. That can mean choosing a cheaper model for a simple task and a stronger one for a complex task, or falling back to a second provider automatically when the first one is degraded or unavailable. An enterprise that wants to standardize on one credential and one endpoint, while still choosing the most appropriate model for each task or reacting when a provider has an outage, needs routing rather than one model wired directly into every application. Routing simple classification requests to a smaller, cheaper model and complex drafting requests to a larger one, both reached through the same endpoint with no client code aware of the difference. **In Dynamiq**, the AI Gateway resolves a router model slug to whichever provider actually serves it and relays the request, so switching between the models it serves is a one-line change to a model name. The SDK's fallback configuration can also retry a secondary model automatically when the primary one hits a rate limit or a connection error. --- # Multi-agent system URL: https://www.getdynamiq.ai/glossary/multi-agent-system Two or more agents working together on one task, each with a narrower role, coordinated by delegation or an explicit control flow. A multi-agent system splits one broad task across several agents, each with a smaller role and a smaller tool list, instead of one agent trying to do everything. Two coordination styles cover most cases: a manager agent that delegates to specialists at run time, or an explicit orchestrator that defines the control flow between agents ahead of time. A regulated workflow often already mirrors a real organizational structure, an analyst, a reviewer, a specialist for a narrow question, and splitting agents the same way keeps each one small enough to reason about and audit on its own, rather than one agent whose behavior spans an entire process. A manager agent routes a request to a research specialist and a writing specialist, then either combines their answers itself or returns a specialist's answer directly when no further editing is needed. **In Dynamiq**, attaching any agent as a tool of another creates a sub-agent, with a limit on how many times it can be called per run and a setting that lets its answer pass straight back as the parent's final answer. The Graph Orchestrator handles the other pattern, an explicit state machine of agents and conditions, and every child run appears as a nested span in the parent's trace. --- # Online evaluation URL: https://www.getdynamiq.ai/glossary/online-evaluation Automatically scoring a sampled share of a deployed system's live production traffic on an ongoing basis, not a one-time test. An online evaluation attaches a metric to a system that is already live and scores a sampled share of its real traffic continuously, rather than a fixed dataset run once before a release. Nothing is re-executed for this: the metric reads what the system already produced. A model, a prompt or an upstream data change can degrade quality after launch in ways a pre-release test never surfaces. A regulated operation needs continuing evidence that a live system still performs the way it did when it was approved, not only a report from the day it launched. Sampling a share of a deployed agent's real conversations and scoring each one for policy adherence as they happen, catching a drift in quality quickly instead of waiting for a periodic manual audit to find it. **In Dynamiq**, attach a metric to a deployed App's evaluations with a sample rate and a path that pulls the metric's input from each recorded trace, and every scored trace links back to its full execution trace, so a low score is one click from the run that produced it. --- # Open-weight model URL: https://www.getdynamiq.ai/glossary/open-weight-model A model whose trained weights are published for anyone to download and run, unlike a model only reachable through a hosted API. An open-weight model publishes its trained weights for download, so anyone can run it on their own infrastructure. That is a narrower claim than fully open source, the training data and code behind it may not be published, but it is enough to run the model yourself instead of calling someone else's API for every request. It sits opposite a closed, API-only model that only ever runs on its provider's servers. A regulated enterprise that cannot send prompts or documents to a third-party API at all, for residency, confidentiality or air-gap reasons, needs a model it can run entirely inside its own infrastructure, and an open-weight model is what makes that possible. Running Llama, Mistral, Gemma or Qwen inside a bank's own Kubernetes cluster, so no prompt and no document it processes ever leaves the bank's own network. **In Dynamiq**, deploy open-weight models from the catalog on the platform's own or your self-hosted Kubernetes infrastructure using the vLLM engine, behind an OpenAI-compatible endpoint, and the Dynamiq SDK itself ships under Apache-2.0 on GitHub. --- # PII detection URL: https://www.getdynamiq.ai/glossary/pii-detection Automatically flagging personally identifiable information, such as names, emails or account numbers, before it reaches a model. PII detection classifies whether a piece of text contains personal data and what kind, a name, an email address, an account number, before that text goes any further. It is a classifier, not an editor: it tells you that personal data is present and what category it falls into, it does not rewrite or remove it. Regulated data handling rules generally require knowing when personal data is about to enter a prompt, be logged, or cross a boundary, before it happens, not after a review turns it up later. A support message containing a customer's email and account number is flagged before an agent ever sees it, so the workflow can route to a stricter handling path instead of processing it the same way as any other message. **In Dynamiq**, a PII detector node screens a message and reports whether personal data was found and which categories, placed anywhere in a workflow, typically ahead of an agent, with a flagged branch routed to a refusal, a transformation step, or a person for review. Dynamiq detects PII and prompt injection; it does not redact or mask text. --- # Prompt injection URL: https://www.getdynamiq.ai/glossary/prompt-injection An attempt, hidden in user input or a retrieved document, to override a model's instructions and make it act against its role. A prompt injection tries to make a model ignore its actual instructions and follow different ones smuggled in through the input instead. A direct injection is typed straight into a chat, telling the model to ignore its previous instructions. An indirect one is hidden somewhere the agent will read it anyway, inside a scraped web page or an uploaded document, where the agent never expected instructions to appear. An agent with tools that can send a message, query a database, or take an action is a much bigger target than a chatbot, because a successful injection does not just produce an unwanted sentence, it can trigger a real action in a real system. A scraped web page contains hidden text instructing the agent to reveal its system prompt or call a tool it was never asked to use, buried in content the agent was only supposed to summarize. **In Dynamiq**, a prompt injection detector node classifies a message before it reaches an agent, backed by a dedicated classifier, and flagged input can route to a refusal or a human review branch instead of the agent, with every verdict recorded in the run's trace. --- # Reranking URL: https://www.getdynamiq.ai/glossary/reranking Re-scoring and reordering a first pass of retrieved documents with a more precise model, so the most relevant ones reach the model first. Reranking takes the results of an initial, fast retrieval pass and re-scores them with a slower, more precise model, so a broader first pass can afford to over-fetch and still hand the model only the strongest results. The common pattern is to retrieve generously with vector search, then rerank down to a small, focused set. Vector similarity alone is a rough filter: it finds passages that are topically related, not necessarily the specific one that actually answers the question. In a regulated answer, the model needs to see the passage that actually applies, not merely one that mentions similar terms. Retrieving a broad set of passages that all mention a termination clause and reranking so the one from the contract that is actually in question ends up first, ahead of similar clauses from unrelated agreements. **In Dynamiq**, a Cohere cross-encoder reranker, an LLM-based ranker, or a time-weighted ranker that favors recent documents can sit between any retriever and the model, on plain vector search results or on GraphRAG's returned facts, with over-fetching on the retriever and trimming on the reranker as the standard pattern. --- # Retrieval-augmented generation (RAG) URL: https://www.getdynamiq.ai/glossary/retrieval-augmented-generation Retrieving relevant passages from your own documents and giving them to a model as context, so its answer is grounded in that material. Retrieval-augmented generation, RAG, has two halves. Indexing converts documents into chunks, embeds each chunk, and writes the result into a vector store, done once, in advance. Retrieval embeds an incoming query, searches that store for the closest chunks, and hands them to a model as context for its answer, done on every request. A general-purpose model does not know a firm's internal policies, contracts or case history, and a regulated answer needs to be grounded in the firm's own current documents, not in whatever the model happened to learn during training, which may be outdated or simply not about this firm at all. An agent asked about a firm's travel policy on business-class flights retrieves the actual current policy passage and answers from it, instead of producing a plausible-sounding guess drawn from general knowledge of how travel policies typically work. **In Dynamiq**, a Knowledge Base bundles conversion, chunking, embedding and vector storage into one managed pipeline with a search endpoint, or the same components can be wired by hand with vector store search and writer nodes against 8 vector stores. Either way, the result attaches to an agent as a retrieval tool. --- # Role-based access control (RBAC) URL: https://www.getdynamiq.ai/glossary/role-based-access-control Granting permissions based on a person's assigned role, such as admin or member, rather than configuring access for each person. Role-based access control grants permissions by role rather than by person: everyone holding a given role gets the same set of permissions, so access is reasoned about, and audited, at the level of a small number of roles instead of a long list of individual grants that drift over time. A regulated organization needs to be able to show who could do what, and role-based grants make that answerable directly: list the people in a role, and you know what they could all do, without inspecting every account one by one. Only an admin role can invite a new member or change another member's role, while every role, including the most basic one, can view usage and create access keys for calling deployed resources. **In Dynamiq**, every organization member holds one role, Owner, Admin or Member, and management actions such as inviting members or changing roles require Owner or Admin. Every API request is authorized by walking the resource's project and organization membership before anything is returned, so a private project stays invisible to anyone outside it, regardless of what they might otherwise be able to see. --- # Self-hosted AI platform URL: https://www.getdynamiq.ai/glossary/self-hosted-ai-platform An AI platform installed and run inside an organization's own infrastructure, cloud account or data center, not a vendor's shared cloud. A self-hosted AI platform runs inside infrastructure the customer already owns and operates, their own cloud account or their own data center, rather than a vendor's shared, multi-tenant service. Data does not leave that infrastructure by default, and the customer controls upgrade timing, scaling and exactly what the deployment can reach on the network. Some regulated workloads cannot go into a shared cloud at all, whatever its certifications, because of a data residency requirement, a confidentiality obligation, or an internal policy that predates any particular vendor. For those workloads, self-hosting is not a preference, it is the only option that clears the bar. A bank runs the platform inside its own Kubernetes cluster, so every prompt, document and trace stays on infrastructure it already controls, monitors and audits, with no new vendor network path to account for. **In Dynamiq**, self-host the full platform in your own AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, or any Kubernetes 1.32+ cluster, or on-prem, from an official Helm chart, with installation guides for AWS EKS, IBM Cloud and OpenShift, or run it as Dynamiq Cloud instead. Fully isolated or air-gapped environments are delivered with Dynamiq's engineers. --- # Single sign-on (SSO) URL: https://www.getdynamiq.ai/glossary/single-sign-on Letting people log in to a platform using the identity and credentials their organization already manages, not a separate password. Single sign-on connects a platform's login to the identity provider an organization already runs, so a person authenticates with the same corporate account they use everywhere else, rather than a password created specifically for this one system. The identity provider, not the platform, remains the place where access is actually decided. Regulated organizations typically require every application to authenticate through the corporate identity provider, both for a consistent access policy across every system and so that disabling one employee's account in one place revokes their access everywhere at once, immediately. A new analyst logs into the platform with the same corporate account they already use for email, and loses access to both the moment IT disables that one account, with nothing else to revoke separately. **In Dynamiq**, SSO is part of the Enterprise plan and works over OIDC, shipped today with Okta and Microsoft Entra ID, so sign-in flows through the identity provider an organization already runs rather than a separate Dynamiq-specific password to manage and eventually forget to revoke. --- # SIP trunking URL: https://www.getdynamiq.ai/glossary/sip-trunking A connection between a phone network and a voice application over the internet using the Session Initiation Protocol, not analog lines. SIP trunking carries phone calls over the internet using the Session Initiation Protocol, connecting a phone number at a carrier to an application, here, a voice agent, instead of a physical phone line. The carrier relationship, the phone numbers and the rates all stay exactly what they already were; only where the call is answered changes. A regulated contact center's phone line is often the one channel its customers actually use, and connecting a voice agent has to work with the numbers and carriers the institution already has, not force a number migration or a new carrier relationship just to add an agent. A contact center points its existing carrier trunk's origination at a voice agent, and customers keep calling the exact same numbers they always have, with no number to relearn. **In Dynamiq**, phone numbers are not sold by Dynamiq. Connect numbers you already own at a SIP provider by pointing that provider's trunk origination at the SIP URI a deployed voice agent provides, with optional digest authentication for providers that require it. Voice agents answer inbound calls; they do not place outbound calls. --- # SOC 2 URL: https://www.getdynamiq.ai/glossary/soc-2 An audit standard from the AICPA evaluating a service organization's controls against the Trust Services Criteria: security, plus optional categories. A SOC 2 report is produced by an independent auditor, not by the organization being audited, evaluating its controls against the AICPA's Trust Services Criteria. Security is required in every SOC 2 report; availability, processing integrity, confidentiality and privacy are each included at the discretion of the organization being audited, depending on which categories are relevant to what it does. It is the audit security and procurement teams most commonly ask a software vendor to produce before approving a purchase, precisely because it comes from an independent third party against a defined, external standard, not a vendor's own description of its own controls. A bank's vendor risk team requests a SOC 2 report as a condition of approving a new AI vendor, reviewed alongside its own security questionnaire before any contract is signed. **In Dynamiq**, a SOC 2 audit is complete; the report is available under NDA on request, alongside HIPAA and GDPR documentation and details of the platform's credential storage, encryption and access controls. Sources: [AICPA & CIMA, SOC 2](https://www.aicpa-cima.com/topic/audit-assurance/audit-and-assurance-greater-than-soc-2) --- # Sovereign AI URL: https://www.getdynamiq.ai/glossary/sovereign-ai AI that an organization or country controls end to end: where it runs, which models it uses, where its data and records live, and who operates it. Sovereign AI is about control rather than a specific technology. A sovereign setup answers four questions on the owner's terms: where the compute runs, which models are used and who can inspect them, where prompts, documents and logs are stored, and who operates and can change the system. The term covers national AI programs, such as in-country compute and local language models, and organizations that must keep AI inside their own perimeter. Governments, central banks, defense and critical infrastructure operators, and regulated firms in countries with data localization rules often cannot send sensitive data to a shared AI service abroad. Sovereignty also covers dependency: if agents are locked into one vendor's proprietary builder, the organization does not fully control them even when the servers are local. A ministry runs citizen-service agents on a cluster in its own data center, uses open-weight models it hosts itself, keeps every trace in that environment, and can move the agents to another cluster without rebuilding them. **In Dynamiq**, the full platform runs in your own cloud account or data center: AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes 1.32+ cluster, or on-prem. Open models such as Llama, Gemma, Mistral and Qwen run on vLLM inside your environment so agents need no outside model provider, and the SDK is open source under Apache-2.0. Fully isolated or air-gapped environments are delivered with Dynamiq's engineers. --- # Speech to text URL: https://www.getdynamiq.ai/glossary/speech-to-text Converting spoken audio into written text in real time, the first stage of a pipeline voice agent's turn. Speech-to-text transcribes a caller's spoken words into text a language model can reason over, the first of three stages in a pipeline voice agent's turn, alongside the language model and text-to-speech. Models differ on latency, accuracy, language coverage, and whether they can detect the end of a turn themselves as part of transcription. A misheard word in a regulated call, an account number, a medication name, a dollar figure, can propagate into a wrong action downstream in a way that matters far more than it would in a text chat, where a person can reread what they typed. A phone caller's narrowband audio is transcribed by a model tuned for telephony, while a clearer browser call uses a different model tuned for that audio instead. **In Dynamiq**, choose from a curated catalog spanning Deepgram, AssemblyAI, Cartesia, ElevenLabs, Speechmatics, OpenAI, Groq and self-hosted options for the speech-to-text stage of a pipeline voice agent, each listed with its measured latency, price and word error rate, or skip the stage entirely with a single realtime speech-to-speech model instead. --- # Structured output URL: https://www.getdynamiq.ai/glossary/structured-output Constraining a model's final answer to a JSON schema, so downstream code receives typed, parseable data instead of free-form prose. Structured output forces a model's final answer into a shape a schema defines, a JSON object with named, typed fields, instead of a sentence a person or a script has to parse and interpret. What downstream code receives is guaranteed to have the fields it expects, not a plausible paragraph that says the same thing a different way on every run. Any downstream system that consumes an agent's answer directly, a case management tool, a ledger, a reporting pipeline, needs a value it can read as data, not text it has to interpret before it can act on it. A triage agent returns a category, a priority and a short summary as named fields, instead of a paragraph a separate script would have to read and guess the same three values out of. **In Dynamiq**, set a response format JSON schema on an agent node and its inference mode switches automatically to one that can guarantee schema-conformant JSON. If the model's answer does not validate against the schema, the agent appends a correction instruction and retries before the run ends, rather than returning something that fails downstream instead. --- # Text to speech URL: https://www.getdynamiq.ai/glossary/text-to-speech Converting written text into spoken audio, the final stage of a pipeline voice agent's turn, or one channel of a realtime model. Text-to-speech turns a model's written reply into audio a caller actually hears, the last stage of a pipeline voice agent's turn. Models differ on how quickly they produce the first audio a caller hears, how natural the voice sounds, and whether a specific cloned or custom voice can be used instead of a stock one. A regulated contact center needs its voice to sound consistent and professional on every single call, and the wait before the agent starts speaking is exactly what a caller experiences as the agent being slow, regardless of how fast the reasoning behind it actually was. A legal disclosure that must be read in full without interruption calls for a different balance of speed and naturalness than a quick conversational reply where responsiveness matters more than a fully natural voice. **In Dynamiq**, choose from a curated catalog spanning ElevenLabs, Cartesia, Deepgram, OpenAI, Groq and self-hosted options for the text-to-speech stage, each with measured latency and a curated voice lineup, plus a custom voice ID field for a cloned or fine-tuned voice from your own provider account. --- # Tool calling URL: https://www.getdynamiq.ai/glossary/tool-calling A model's ability to invoke a defined function or external action mid-response, then read the result back before continuing. Tool calling gives a model a list of available tools, each with a name, a description and an input schema, and lets it decide during its own reasoning which one to call and with what arguments, then read the result back before deciding what to do next. It is what turns a model from something that only talks into something that can look something up or actually do something. That capability is also exactly where risk enters: a tool call can touch a real system, send a message, run a query, take an action, not just produce a sentence. What an agent can call is the real, practical boundary of what it can do, more than anything written in its instructions. An agent calls a CRM lookup tool mid-answer to check a real order's actual status, instead of producing a plausible-sounding guess about what the status probably is. **In Dynamiq**, any workflow node, including another agent, can be attached to an Agent node as a tool, drawing from web search, code execution, databases, HTTP calls, 2,100+ app integrations, and Model Context Protocol servers. The model only ever sees a tool's name and description, and runtime values that must never reach the model are injected separately, invisibly to it, at the moment the tool actually runs. --- # Vector database URL: https://www.getdynamiq.ai/glossary/vector-database A database built to store embeddings and search them by similarity, returning the nearest vectors to a query, not exact matches. A vector database indexes embeddings so that, given a query vector, it can quickly return the nearest ones by similarity, at a scale and speed a general-purpose relational database is not built for. It is the storage layer most retrieval-augmented generation is built on top of, holding the embedded chunks a retriever actually searches. Because it holds the same sensitive content as the documents it was built from, everything about it, where it runs, who can reach it, how it is backed up, is part of the same governance question as the documents themselves, not a separate infrastructure decision. A firm's policy manual, once embedded, lives in a vector index that a retrieval query searches in milliseconds, instead of scanning every document in the manual on every question. **In Dynamiq**, Knowledge Bases provision managed vector storage by default, backed by Weaviate, or can point at 8 vector stores you operate yourself, including Pinecone, Qdrant, Milvus, pgvector, Elasticsearch, OpenSearch and Chroma, with the same retrieval and access control options either way. --- # Voice agent URL: https://www.getdynamiq.ai/glossary/voice-agent A real-time conversational agent that talks with people over audio, on a phone call, in a browser, or over a raw audio connection. A voice agent holds a real-time conversation over audio rather than text, either as a pipeline of separate speech-to-text, language model and text-to-speech stages, or as one realtime speech-to-speech model handling all of it directly. Either way, it has to decide what to say, when to say it, and when to stop talking, inside the fraction of a second a caller will tolerate before a pause starts to feel wrong. A regulated contact center's phone line carries disclosures, recorded consent, and moments that need a human, and all of it happens live, in a channel that cannot be edited before a caller hears it, unlike a chat message that can be reviewed before it is sent. An inbound support line where the voice agent verifies an account, answers routine questions, and transfers to a person for anything it should not handle on its own. **In Dynamiq**, build a voice agent from written instructions, a choice of 15 voice model vendors across speech-to-text, language model and text-to-speech, or a single realtime model, plus tools and transfer destinations, deployed over phone numbers you already own, a browser client, or a raw WebSocket connection. Voice agents answer inbound calls; they do not place outbound calls, and calls can be recorded, with a transcript and per-turn latency. --- # WebRTC URL: https://www.getdynamiq.ai/glossary/webrtc An open, browser-native protocol for real-time audio and video that lets a web page or app stream microphone audio to a voice agent. WebRTC is a browser and mobile-native protocol for real-time audio and video, built to handle the parts that make live audio hard over an ordinary network connection: adapting to changing network conditions, recovering from a dropped packet, reconnecting cleanly, all without a browser plugin. For a voice agent, it is what lets a web page or an app stream a caller's microphone audio directly, with no phone call involved at all. An enterprise's own product, a customer portal, a mobile app, often needs a voice agent embedded directly in its own interface, not routed out through a phone number and back in through a carrier. A customer clicks talk to support inside a web app and starts talking through their microphone immediately, with no phone number dialed and no call placed on any carrier. **In Dynamiq**, web and mobile callers connect to a voice agent over WebRTC through a client SDK that captures the microphone, and it is what the platform's own browser test tab uses to let you talk to an agent before deploying it. A backend that already holds the audio itself, a SIP trunk or a telephony stack, uses the WebSocket transport instead. --- # How to automate loan file QC with rules your analysts can change URL: https://www.getdynamiq.ai/post/automate-loan-file-qc-rules-analysts-can-change Published: 2026-10-05 Automate loan file QC with rules your analysts can read and change, a calibrated judge for names, a routing table and test cases behind every release. A loan file QC workflow on Dynamiq reads the fields your extraction step pulled from each document, runs your checklist as plain rules, asks a calibrated judge whether two names are the same party, and routes the loan to a queue. It holds up in an audit because every finding quotes the rule, shows the values it read, and runs on a saved, tested and traced version of the workflow. We built the example in this post on synthetic data so every screenshot is real: eleven invented loans, one fictional watchlist, and the same workflow you can build in a free account. ## Why do loan QC checklists outgrow rule engines and spreadsheets? A mortgage file is a few hundred pages: the note, the deed of trust, the Closing Disclosure, the loan application, the appraisal, the automated underwriting findings, the credit report, flood and hazard insurance, and more. Quality control checks that file against a checklist, whether it runs before funding, after closing, in due diligence, before a purchase or at servicing boarding. Each failed check is a defect, and defects are graded by severity. The checklist is rarely the hard part. Keeping the automation in step with it is: - **Checklists change often.** Agencies and investors update their guidelines several times a year, and each change has to reach whatever runs the checks. - **The automation sits where analysts cannot reach it.** In a Java rule engine, a BRMS, or spreadsheets and macros, changing one threshold is a developer ticket, and checks that are awkward to code stay manual. - **Extracted data is messy.** OCR and IDP return "$412,500.00", "6.375%", "TBD" or a blank, each with a confidence. A check that quietly reads "TBD" as zero passes a loan it should stop. - **Some checks need judgment.** "Jon Reyes" and "Jonathan Reyes" are the same borrower. A loan officer one letter off a watchlist entry probably is not the listed person, but someone should look. String matching gets both wrong. - **A model on its own is hard to defend.** The same file can get two answers, and an auditor wants to know which value failed which rule. ## What does an automated loan file review look like? ![The loan-file-qc-demo workflow on the Dynamiq canvas: input and documents feed three branches (policy-checks; name-pairs into name-consistency; screen-parties into watchlist-review, each Map holding a same-party-judge), which meet in identity-findings, route and output.](https://cdn.sanity.io/images/csaezwuw/production/2b3e496deafc6b1447b995c4b2024a8442d7f54e-1600x911.webp "The whole review is one workflow: deterministic policy checks on top, two judged identity checks below, one routing table at the end.") The workflow takes the documents an extraction step already produced. Each document has a type such as `PROMISSORY_NOTE` or `CLOSING_DISCLOSURE`, a `signed` flag, a received time, and extracted fields, each with a value and a confidence. 1. **documents** (Expression) picks one copy of each document type: the latest signed one. Drafts and duplicates stop here, in one place. 2. **policy-checks** (Rules) runs the checklist: amounts, rates, dates, ages, ratios, required documents and extraction confidence. 3. **name-pairs** (Expression) lists every borrower name that differs, as text, between the note and the other documents. 4. **screen-parties** (Custom Function) collects every party in the file (borrower, lender, loan officer, appraiser, appraisal company, employer, insurer) and returns close calls against the lender's watchlist. It matches loosely on purpose; it decides nothing. 5. **name-consistency** and **watchlist-review** (Map) send each pair to a Judgement node, a few at a time. 6. **identity-findings** (Rules) turns the judge's answers into findings. 7. **route** (Decision table) picks the queue and priority: `clear`, `analyst-review`, `cure` or `reject`. ## Rules your QC team can read and change ![The policy-checks rules editor in Dynamiq: a searchable list filtered by severity and category, each rule shown with its name, its check expression, a fail severity badge, a coverage badge such as 1x or never from the last test run, and an on/off switch.](https://cdn.sanity.io/images/csaezwuw/production/cda1ab2fa04f23869b38e164bcb652507170a1df-1600x1507.webp "The checklist as rules: each one names the check it runs, its severity, how often the test cases made it fire, and a switch to turn it off.") A rule in the Rules node reads like the checklist row it automates. It has a name, a category, a severity (fail, warn or info), an optional condition for when it applies, the check itself, a message that inserts the values it read, a reason code and references. Derived values are named facts computed once per loan, so rules stay short: ```text note_amount = number(note.loan_amount.value) cd_amount = number(cd.loan_amount.value) note_date = date(note.note_date.value) loan_type = text(application.loan_type.value) | lower ltv = number(aus.ltv.value) ``` Three rules from the demo: ```yaml - name: Note and Closing Disclosure loan amounts match severity: fail check: note_amount == cd_amount message: > Loan amount is {{ note.loan_amount.value }} on the Note but {{ cd.loan_amount.value }} on the Closing Disclosure. - name: Mortgage insurance is in the file above 80% LTV (conventional) severity: fail applies_when: loan_type == 'conventional' and ltv > 80 check: mi is present message: > Conventional loan at {{ aus.ltv.value }} LTV, but no mortgage insurance certificate is in the file. - name: Closing Disclosure received at least 3 days before closing severity: fail check: days_between(date(cd.issue_date.value), date(cd.closing_date.value)) >= 3 references: ["CFPB TRID, 12 CFR 1026.19(f)(1)(ii)"] ``` The demo's thresholds are illustrations, not compliance advice. The point is who can change them: an analyst who can read `ltv > 80` can maintain it. ![The mortgage insurance rule open in Dynamiq: severity Fail, when data is missing set to report the rule as not evaluated, the check mi is present, applies when loan_type == 'conventional' and ltv > 80, a message that inserts the LTV, category, reason code, effective dates, references, and its last result on each test case, including a fail on DEMO-0003 that reads "Conventional loan at 90.05% LTV, but no mortgage insurance certificate is in the file."](https://cdn.sanity.io/images/csaezwuw/production/4e523d851882010bf083cd69ad85f20448bc15d8-1600x2298.webp "One rule, everything an auditor asks about: what it checks, when it applies, what missing data means, its source, and what it said about each test loan.") Four details decide whether rules hold up on real extracted data: - **Missing is not pass.** Each rule, or the whole node, says what missing data means: report the rule as not evaluated, fail it, or treat it as not applicable. A loan never clears because a value was absent. - **Unreadable is not zero.** `number()` reads "$412,500.00" and "6.375%", and reports "TBD" as not a number. In the demo, a Closing Disclosure amount of "TBD" makes the amount check not evaluated, and the loan goes to an analyst. - **Rules have dates.** Each rule carries effective from and until dates and an on/off switch, so a guideline change can be scheduled rather than deployed at midnight. - **Rules travel as CSV.** Export the list, review it in a spreadsheet with compliance, and import it back. ![A rule's check being typed in the Dynamiq expression editor: note_amount == cd_amount && deed is present, with && underlined, the message "Use 'and' instead of '&&'." under the field and in the rule header, and an autocomplete list offering present.](https://cdn.sanity.io/images/csaezwuw/production/24cb3c4c1b4dedad821ca8894cab77f29cf43976-1600x588.webp "The editor checks expressions as you type: here it flags && and offers the present test.") The expression editor highlights syntax, colors bracket pairs, completes the names of inputs, derived values and helpers, and lints as you type: write `&&` and it says to use `and` instead. ## When text alone cannot decide: a calibrated judge ![The same-party-judge inside the name-consistency Map, selected on the canvas, with its settings: the same_party yes-or-no question, its instructions, yes when and no when guidance in plain English, a yes threshold of 0.5 and a minimum confidence of 0.75.](https://cdn.sanity.io/images/csaezwuw/production/e58fa1e6ab0568b510e5e3c04cca555f4515a5a1-1600x1764.webp "One narrow question in plain English. Below the minimum confidence, the answer goes to a person.") Two checks in a loan file are about identity, not arithmetic: is the borrower the same person on every document, and is any party on the lender's watchlist? We ask one narrow yes-or-no question, "Do these two names identify the same person or company?", with plain-English guidance on when the answer is yes (a nickname, a middle initial, word order, a typo, a corporate suffix) and when it is no (a different first name, a different surname, a different business that only shares common words). The judge is TypeSafe System One, which returns calibrated probabilities, so each answer comes with a confidence. Below a minimum confidence, set to 0.75 in the demo, the answer is flagged for review instead of being used. Nothing is cleared on a guess. The Judgement node can also use an LLM or an agent as the judge. In the demo this resolves three different cases: - **Jonathan on the note, Jon on the Closing Disclosure.** The judge says same person, and the loan clears. - **A different person on the deed of trust.** The judge says different, and the loan goes to `cure`. - **A loan officer one letter off a watchlist entry, and an appraisal company that is on the list.** The screening step sends both to the judge. The listed company is confirmed with a confidence of 0.98, so the loan is rejected. For the loan officer the judge leans toward "same person" but only with a confidence near 0.7, below the minimum, so that match is flagged for an analyst rather than decided. ## Your routing policy as a table ![The route decision table in Dynamiq: seven rows with conditions on policy, identity and watchlist_hits, outputs for queue and priority, a Matched column (1x, 4x, 1x, 1x, 1x, never, 3x) from the last test run, and an on/off switch per row.](https://cdn.sanity.io/images/csaezwuw/production/f32ba984101d411abc9ecfb7c9a6155d159ce7cc-1600x467.webp "Rows are read in order and the first match wins. Matched shows how many test loans each row decided.") Queue and priority come from a decision table, not code. The demo's table, row by row: | Row | When | Queue | Priority | Test loans matched | | --- | --- | --- | --- | --- | | watchlist hit | a party is confirmed on the watchlist | reject | high | 1 | | policy failed | a policy check failed | cure | high | 4 | | borrower differs | the judge says a document names a different borrower | cure | high | 1 | | check not run | a policy check could not be evaluated | analyst-review | normal | 1 | | policy warning | a policy check warned | analyst-review | normal | 1 | | judge unsure | a match was below the minimum confidence | analyst-review | normal | never | | clean | anything else | clear | low | 3 | Columns are typed, rows can be reordered and switched off, and the table imports and exports as CSV. The Matched column is worth a look: "judge unsure" never decided a loan in this set, because the only unsure match came with a confirmed watchlist hit, which wins. A row no test loan reaches needs a test loan. ## How do you know it works? ![The Test cases tab in Dynamiq: 11 cases, 11 passed, listing DEMO-0001 clean through DEMO-0011 watchlist, each with a Passed badge and its run time, plus Import CSV, Export CSV, New case and Run all.](https://cdn.sanity.io/images/csaezwuw/production/30dd0458e642a91a65983f230775eb261317ede7-1452x1254.webp "Each synthetic loan is a saved test case with the queue it must reach. Run all replays them on every change.") Every loan we care about is a saved test case with the output it must produce. **Run all** replays them after every change. We include a planted case: a copy of a clean loan with deliberate defects, a loan officer one letter off a watchlist entry and a listed appraisal company, so screening is tested on every release. | Loan | What it shows | Queue | | --- | --- | --- | | DEMO-0001 | Every check passes | clear | | DEMO-0002 | Closing Disclosure amount $1,000 higher than the note | cure | | DEMO-0003 | Conventional loan at 90% LTV, no mortgage insurance certificate | cure | | DEMO-0004 | Jonathan on the note, Jon on the Closing Disclosure | clear | | DEMO-0005 | A different person on the deed of trust | cure | | DEMO-0006 | Closing Disclosure loan amount reads "TBD" | analyst-review | | DEMO-0007 | Note interest rate extracted with 41% confidence | analyst-review | | DEMO-0008 | Credit report 150 days old at the note date | cure | | DEMO-0009 | Flood zone AE, no flood insurance in the file | cure | | DEMO-0010 | An unsigned draft and a signed final Closing Disclosure | clear | | DEMO-0011 | Watchlist near-miss and a listed appraisal company | reject | All eleven pass on the platform with the real judge, and the outcomes in the table are what it produced. ![Test case DEMO-0006 opened in Dynamiq: expected output queue analyst-review, and findings that read 1 not evaluated, 3 not applicable, 14 pass; the loan amount rule is not evaluated with the message "check could not be evaluated: not a number: 'TBD'" and the values it read, note_amount 628500 and cd_amount unreadable: not a number: 'TBD'.](https://cdn.sanity.io/images/csaezwuw/production/2371e00f21bbe7779c84bf3fa2b3a7e29df4ef25-1396x1336.webp "DEMO-0006: the amount check reports TBD as unreadable and the loan goes to an analyst. Nothing is guessed.") ![The By rule view in Dynamiq: 28 rules, 14 never fired across the 11 cases run, with each rule's Fired badge (1x or never) and the passed and failed counts of the cases it fired in.](https://cdn.sanity.io/images/csaezwuw/production/d23637eb5270d846069b3269a5d8bb67d3cfc886-1452x1294.webp "By rule: which rules the saved cases make fire, and which they never reach.") The By rule view shows, for every rule and every row of the routing table, how many saved cases it fired in. In our demo it reads 28 rules, 14 never fired across the 11 cases. That is the point of the view: each rule marked never is one no test loan has exercised yet, and the next test case to write. ## From canvas to production: versions, the API and traces ![The Integration tab of the deployed loan-file-qc-demo app in Dynamiq: ready-made Python for a regular HTTP request with the documents and review_date input, plus sections for server-sent events streaming, WebSocket, async callback and streaming with human feedback, and a Python or TypeScript switch. The endpoint address is blurred.](https://cdn.sanity.io/images/csaezwuw/production/dddff123ad928b36c3843b7c10079242ba3850ed-1516x2172.webp "Deployed, the workflow is an API with ready-made code in Python or TypeScript.") Every save is a version, and you deploy the version the API runs. The integration tab gives ready-made Python and TypeScript for regular HTTP requests, server-sent events streaming, WebSocket, async callbacks and streaming runs that pause for human feedback, and access keys can be scoped to a project or the organization, with an expiry. Every run is traced with each step's inputs and outputs, so months later you can show exactly why a loan went where it went. See [observability](/product/observability) and [deployment options](/product/deployments). ## How we approach a migration from a legacy rule engine When a lender or review firm already has a rule engine, we start from the written checklist, not the code: 1. **The checklist is the source of truth.** The old engine shows how each test was implemented. Where the two disagree, we follow the checklist and list every difference for the business to confirm. 2. **Account for every row.** Each checklist row ends up as one of four things: automated; waiting for a value the extraction does not produce yet; needing an outside system, such as a fraud tool or a custodian; or a process step that is not a data check. That coverage map turns "how much is automated?" into a precise answer and gives the extraction team a to-do list. 3. **Deterministic first.** Comparisons, dates and presence checks become rules. Judgment is reserved for the questions text cannot settle. 4. **Audit against the source documents.** Before go-live, we check each finding on the test loans against the actual pages. Some findings are real defects; some are extraction errors, and those go back to the extraction team. ## Where else does this pattern fit? | Use case | What is checked | Nodes doing the work | | --- | --- | --- | | Prefunding QC | Note and Closing Disclosure agree; disclosure timing; insurance in place; AUS data matches the file | Expression, Rules, Decision table | | Post-closing QC | The full checklist on a sample or every loan, with defect severity | Rules, Judgement, Decision table, test cases | | Due diligence and pre-purchase review | Investor guidelines; data integrity across documents; parties screened | Rules, Map and Judgement, Custom Function, Decision table | | Servicing boarding | Boarded data against the note and Closing Disclosure; required documents present | Expression, Rules | | Watchlist and party screening | Borrowers, loan officers, appraisers, vendors and employers against internal lists | Custom Function, Map and Judgement, Rules | | Insurance underwriting audit | Application against policy documents; required forms; underwriter authority | Rules, Judgement | | Accounts payable | Invoice against purchase order and receipt; vendor screening | Expression, Rules, Judgement | The shape is the same each time: extracted documents, a checklist, a few judgment calls and a routing decision. More in [document review](/use-cases/document-review), [back-office operations](/use-cases/back-office), [KYC and onboarding](/use-cases/kyc-onboarding), [financial services](/industries/financial-services) and [insurance](/industries/insurance). ## How does this compare with other approaches? | | Java rule engine or BRMS | LLM-only agent | RPA bots | Dynamiq workflow | | --- | --- | --- | --- | --- | | Who changes a rule | A developer | Whoever edits the prompt | A developer | A QC analyst, in plain expressions | | Same input, same answer | Yes | Not guaranteed | Yes | Yes for rules and tables; the judge's answers carry a confidence | | Names and near-matches | Hand-coded | Possible, without a calibrated score | Usually not | A calibrated judge, with a review threshold | | Tested before release | Separate tooling | Rarely built in | Rarely built in | Test cases in the editor, with rule coverage | | Explaining a finding | Logs | Hard to reconstruct | Screenshots | Each finding shows the values it read; every run is traced | ## How does Dynamiq compare with Copilot Studio, n8n and LangGraph for loan QC? Each is a strong tool for the job it was built for. For a loan QC checklist, the questions that matter are where the files are processed, who owns the rules, and how you prove a release is right. - **[Microsoft Copilot Studio](/compare/microsoft-copilot-studio)** suits teams standardized on Microsoft 365 and Power Platform, where business users build agents for Teams. It runs only as a Microsoft-hosted service. Dynamiq runs on Dynamiq Cloud, in your own AWS, Azure, GCP, IBM Cloud or OpenShift, or on-prem, so loan files stay in the environment your auditors already approved. - **[n8n](/compare/n8n)** is a mature workflow automation platform with a very large integration library, and it adds guardrails, human approvals and evaluations to those workflows. Dynamiq treats the checklist itself as the product: rules with severities, a missing-data policy, effective dates and coverage from test cases, a calibrated judge for identity questions, and a routing table the business reads. - **[LangGraph](/compare/langgraph)** gives developers low-level control over agent orchestration in code. There, every checklist rule is code a developer owns. On Dynamiq, a QC analyst edits the rule on the canvas, and developers still get the same engine as an Apache-2.0 Python SDK. ## How to start Build the workflow in a free account: [Start free](https://app.getdynamiq.ai/signup), then follow the docs for the [Rules node](https://docs.getdynamiq.ai/docs/platform/workflows/orchestration/rules-node), [Judgement node](https://docs.getdynamiq.ai/docs/platform/workflows/orchestration/judgement-node), [Decision table](https://docs.getdynamiq.ai/docs/platform/workflows/orchestration/decision-table-node), [Expression node](https://docs.getdynamiq.ai/docs/platform/workflows/orchestration/expression-node), [Map node](https://docs.getdynamiq.ai/docs/platform/workflows/orchestration/map-node) and [test cases](https://docs.getdynamiq.ai/docs/platform/workflows/test-cases). If you want to see it on your own files and checklist, [talk to the team](/book-a-demo). ## FAQ ### Can our analysts change rules without a developer? Yes. Rules are plain expressions such as `note_amount == cd_amount`, with a name, severity, message and references. Analysts edit them in the canvas, export and import them as CSV, and run the saved test cases before the change is deployed. ### What happens when extracted data is missing or unreadable? Each rule decides what missing data means: not evaluated, fail or not applicable. Text that is not a number, such as "TBD", is reported as unreadable rather than read as zero. In the demo, both route the loan to an analyst. ### How does the judge decide, and when does a person review? The judge answers one narrow question, such as whether two names identify the same person, with a calibrated probability and a confidence. Below the minimum confidence you set, the answer is flagged and the loan goes to an analyst. ### Can it run on our extraction vendor's output? The workflow reads documents as structured fields with values and confidences. If your OCR or IDP vendor produces fields, the documents step maps them; changing vendors changes that mapping, not the rules. ### How is every decision audited? Every save is a version, the API runs the version you deploy, and every run is traced with each step's inputs and outputs. Each finding carries the rule's message, its reason code and the values it read. --- # How Dynamiq built a cost-aware legal research workflow with IBM watsonx URL: https://www.getdynamiq.ai/post/how-dynamiq-built-a-cost-aware-legal-research-workflow-with-ibm-watsonx Published: 2026-04-15 How a three-agent legal research workflow on IBM watsonx routes every question by cost: Granite triages, deep research runs only when needed, all traced. Dynamiq built a legal research workflow for a European insurance company that answers contract, policy and compliance questions at a controlled cost. A small IBM Granite model classifies every request first, and an expensive research agent runs only when a question needs it. Every reasoning and tool step is traced, and the whole workflow is available across the company as an agent in IBM watsonx Orchestrate. This post walks through the requirements, the three-agent design, how models were chosen, and the patterns other legal and compliance teams can reuse. ## What problem did the legal team need to solve? Legal teams are expected to give fast, defensible answers across contracts, policies and compliance questions, while the volume of documents keeps growing. At this insurer, the work was spread across disconnected tools, handoffs and manual review: summarizing several contracts at once, answering policy questions against internal agreements, comparing internal language with external sources, and checking compliance across jurisdictions. Speed matters here because one overlooked clause can change liability, payment terms or compliance obligations, and lawyers often compare dozens of clauses across agreements under tight deadlines. The goal was faster, better-informed answers for the business without raising costs. ## What were the requirements? The constraints are what turned a demo into a production project: - **Tight cost control.** Expensive reasoning could not be applied to every request. - **Partly deterministic routing.** Every query had to be classified before any agent answered it. - **Full traceability.** Each reasoning step, tool call and model response had to be logged. - **Existing infrastructure.** The workflow had to use the insurer's search and document systems. - **Response time.** A query had to complete within a few minutes. - **Internal and external sources.** Proprietary legal document stores, plus external legal research through [Exa](https://exa.ai/docs/reference/search-api-guide). - **Access control.** Permissions aligned with the insurer's SOC 2 requirements. ## How does the three-agent architecture work? The design has three agents with distinct roles and cost profiles: an orchestrator, a low-cost classification agent and an advanced research agent. The orchestrator makes sure the cheapest adequate path is chosen, the classifier triages and answers simple queries, and the research agent does deep work only when the classifier recommends it. ![Architecture diagram: a legal query passes from IBM watsonx Orchestrate to a Dynamiq orchestrator agent on Grok 4 Fast, which asks a Granite 4 Small classifier how to route it and, when recommended, calls a Grok 4 Fast research agent that searches internal contracts in Milvus on watsonx.data and external legal research in Exa](/blog/how-dynamiq-built-a-cost-aware-legal-research-workflow-with-ibm-watsonx/1-dynamiq-runtime-1.webp "Every query goes through a low-cost classifier before any deep research runs") ### The orchestrator: routing, synthesis and budget control The orchestrator follows a route, decide, synthesize pattern. It sends each request to the classifier before answering, then decides whether the low-cost path is enough or the request should escalate. It also watches token consumption across the run and applies two budget controls: - When the classifier reports low confidence on a request that would be expensive to research, the orchestrator asks for the assumptions to be confirmed first. - When costs spike, it caps the research agent's tokens per step, returns a constrained summary and offers a "continue?" control. It also treats phrases such as "our terms", "my contracts" and "internal policies" as internal queries by default. For the orchestrator model, the team found Claude Sonnet 4 and GPT-5 capable but slower and more expensive for this role. Grok 4 Fast gave the needed balance of speed, cost and quality. ### The classifier: IBM Granite for low-cost triage Every request stops first at the classification agent, which runs on [IBM Granite 4 Small](https://www.ibm.com/granite/docs/models/granite). It returns structured output with six fields: complexity, data sources, task type, reasoning, recommendation and confidence. That output decides whether the workflow answers locally, escalates to the research agent or takes a hybrid path, and whether external calls are allowed at all. When the classifier marks a request INTERNAL_ONLY, access to Exa is turned off. - **Simple queries**, such as legal definitions, straightforward policy questions and status inquiries, are answered without escalation. - **Moderate queries**, such as contract term explanations, single-document analysis and basic compliance checks, may use the research agent. - **Complex queries**, such as comparisons across documents, clause-level compliance scoring and multi-source research, always escalate. The classifier is both a routing layer and a cost-control layer. For this triage task, Granite 4 Small delivered better price-to-performance than the Grok 4 Fast model used elsewhere in the system. ### The research agent: internal retrieval plus external legal sources The research agent runs on Grok 4 Fast and activates only when the classifier recommends an EXTERNAL_ONLY or HYBRID path. It searches the insurer's proprietary legal documents through a retrieval-augmented generation (RAG) pipeline on Milvus in IBM watsonx.data, and uses Exa for public legal sources and case law. It produces four kinds of output: 1. **Semantic contract search** for related clauses, obligations and rights. 2. **Comparative analysis** across documents, with differences in language and business impact. 3. **Clause-level compliance scoring** on a 1 to 10 risk scale, with rationale and mitigation advice. 4. **A cited summary** that blends internal and external results and labels which is which. Each answer includes an executive summary, detailed analysis, recommendations, citations and a risk assessment that flags ambiguities, conflicts, outdated references and coverage gaps. When sources contradict each other, the agent shows the conflicting passages side by side. When retrieval comes back weak, it broadens the search and records a coverage_gap flag instead of guessing. ## How is every step traced? To meet the traceability requirement, the agents run in an XML-structured [ReAct](https://www.ibm.com/think/topics/react-agent) inference mode, which keeps every reason, act and observe step in the same shape and easy to debug. The runtime produces timestamped logs of every reasoning decision, tool call and model response in a run. That makes postmortems straightforward across the orchestrator, agents, tools and models, including the prompts and responses exchanged with watsonx Orchestrate when the workflow is called through its API. Model and embedding versions are stored with each trace, so an answer can be tied to the exact configuration that produced it. See how [observability](/product/observability) works in Dynamiq. ## Why run the workflow through watsonx Orchestrate? Once the multi-agent workflow was built in Dynamiq, the team added it to IBM watsonx Orchestrate as an external agent through the API, using a [bearer token](https://www.ibm.com/docs/en/watsonx/w-and-w/2.1.0?topic=keys-generating-bearer-token) and a service instance URL generated by the Dynamiq platform. That step turned a stand-alone workflow into a company-wide capability. Every authorized user can call the legal research agent from chat, alongside agents connected to systems such as SAP, Salesforce and ServiceNow. The agent appears in a governed catalog, so IT and compliance teams can see what is running, who uses it and how it connects to the rest of the stack, without Dynamiq rebuilding any integration. ## What did the insurer gain? Contract review, clause identification and answers to business questions all became substantially faster; IBM's [case study of the project](https://www.ibm.com/case-studies/dynamiq) reports the measured reductions. Cost stayed predictable because routine triage runs on Granite and never touches the more expensive reasoning model. Legal teams across the company use the agent through the interface they already work in. The answers are easier to trust: citations separate internal and external sources, reasoning steps are logged for audit, and uncertainty is flagged rather than hidden. The architecture can be deployed in the cloud or on-premises as it moves further into production. ## What can other legal and compliance teams reuse from this design? - **Route before you reason.** A small model that classifies requests keeps most traffic away from expensive reasoning. - **Make routing output structured and enforceable.** Fields such as complexity, sources and confidence let code, not the model, decide what happens next. - **Turn off external access for internal questions.** An INTERNAL_ONLY route removes a whole class of data-leak risk. - **Cap spend per step.** A token budget with a "continue?" control prevents runaway costs on long research tasks. - **Store versions with traces.** Auditors and lawyers can tie an answer to the model and index that produced it. - **Surface conflicts.** Showing contradictory sources side by side builds more trust than a confident single answer. The same pattern applies to [document review](/use-cases/document-review) in underwriting, claims and contracts. Read more about our [partnership with IBM](/partners/ibm), including watsonx models and data and deployment on IBM Cloud. ## FAQ ### What is a cost-aware agent workflow? It is a workflow that matches the cost of the model to the difficulty of the task. A cheap model classifies and handles simple requests, and expensive reasoning runs only for requests that need it, with token budgets that cap spend on long tasks. ### Why use a small model like IBM Granite to classify queries? Classification is a narrow, structured task that a small model handles well. Running it first on every request costs little and keeps the larger model for work that needs deep reasoning, which lowers the average cost per question. ### How does the workflow keep internal documents private? Internal contracts are retrieved from Milvus in IBM watsonx.data, and when the classifier marks a request as internal only, access to the external research service is turned off. Citations label internal and external sources separately. ### How does a Dynamiq agent run inside IBM watsonx Orchestrate? The workflow is added to watsonx Orchestrate as an external agent through the API, using a bearer token and a service instance URL generated by the Dynamiq platform. Authorized users then call it from Orchestrate's chat like any other agent. ### Can this workflow run on-premises? Yes. The architecture can run in the cloud or on-premises, and Dynamiq itself can be self-hosted on IBM Cloud, Red Hat OpenShift, other major clouds, any Kubernetes cluster or on-prem. --- # Best Dify alternatives in 2026: open-source and paid options URL: https://www.getdynamiq.ai/post/best-dify-alternatives-for-ai-workflow-automation Published: 2025-09-17 Dify is a source-available platform for agentic workflows and RAG. Compare the best Dify alternatives in 2026 on license, deployment, voice agents and evals. Dify has become a go-to for product teams that want to move from prototype to production quickly. It now calls itself a platform for production-ready agentic workflows: a visual builder for workflows, chat apps and agents, Knowledge Pipelines for RAG, a large plugin marketplace, and deployment on Dify Cloud, a self-hosted Community Edition, or Dify Enterprise in a VPC, on-prem or air-gapped. Teams compare Dify alternatives for a handful of reasons: its modified Apache License restricts multi-tenant hosting, it has no voice agent product, we found no built-in evaluation suite in its docs, and its Enterprise support is customer success rather than engineers who build with you. This guide compares the main options as of September 2026, using each vendor's own pages, listed at the end. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Fast prototyping of LLM apps and RAG | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Workflow automation with AI as one part | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [Sana](https://sanalabs.com/) | Proprietary | Managed SaaS by default; can deploy on your cloud; single-tenant on Enterprise | No-code agents, workflows and enterprise search across 100+ apps | Analytics dashboard, response feedback, knowledge verification | Company-wide knowledge assistant and learning | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | Flowise, which earlier versions of this guide listed, reached end of life on August 31, 2026; see [below](#what-happened-to-flowise). ## What to look for in a Dify alternative 1. **License terms that fit how you ship.** Dify's license bars operating a multi-tenant environment without written authorization and requires keeping its logo and copyright in the console and apps. If you plan to host agents for many customers or business units, check this first. 2. **Agentic orchestration that scales.** Planning, tool use, memory, retries and guardrails should be first-class, along with human approvals at risky steps. 3. **Knowledge and RAG depth.** Look at ingestion pipelines, retrieval filters that respect permissions, citations, and how easily you can swap vector stores and models. 4. **Evaluations and observability.** Offline and live evaluations, traces with per-step latency and cost, and alerts keep quality high and spend predictable. 5. **Deployment.** SaaS is simple, but many teams need a private VPC, on-prem Kubernetes or an isolated network, with SSO and audit trails. ## Top Dify alternatives As a baseline, Dify's strengths are real: a free cloud Sandbox, a self-hosted Community Edition, a marketplace of plugins for models, tools, data sources and MCP, apps that publish as MCP servers, human input steps in workflows, and SOC 2 and ISO 27001 per its homepage. Here is how the alternatives compare. ### Langflow Langflow is an MIT licensed, Python-based builder for agents and RAG applications that can also publish flows as MCP servers and call other MCP servers. It runs as a desktop app, a Python package, a Docker container or a headless runtime on Kubernetes, and recent releases added a Guardrails component, human-in-the-loop checkpoints, built-in traces and OpenTelemetry export. It is the most permissive license in this list. Plan for gaps around built-in evaluations, which are not documented, and enterprise SSO and RBAC enforcement, which are plugins outside open-source Langflow. ### n8n n8n is fair-code workflow automation with AI built in: an AI Agent node with tools and sub-agents, n8n Agents in preview since September 2026, a Guardrails node, human approval before an agent runs a chosen tool, quick and metric-based evaluations, and knowledge bases on n8n Cloud. It fits teams that automate processes across many apps. Its Sustainable Use License limits use to internal business, non-commercial or personal purposes. ### Zapier Zapier is the no-code option, connecting 9,000+ apps. Zapier Agents draw on files and app knowledge sources, call other agents and pause for approvals, and chatbots embed as a widget on your site. Zapier runs only in its own cloud, hosted on AWS in the United States, and we found no evaluation datasets or scoring for agents in its documentation, so complex AI pipelines may outgrow it. ### Sana Sana, part of Workday since November 2025, is an AI platform for knowledge work: no-code agents, scheduled workflows, enterprise search across 100+ apps, and Sana Learn for training content. It offers SSO, SCIM and audit logging, managed hosting by default with deployment on your own cloud or a single-tenant option for Enterprise, and says it does not use customer data to train models. It fits when the goal is trusted answers and automation for every employee rather than building custom agents for products or operations. ### Microsoft Copilot Studio Within Microsoft 365, Copilot Studio stands out: agents and workflows that publish to Teams and Microsoft 365 Copilot, multi-agent orchestration, OpenAI and Anthropic models, agent evaluations and a Microsoft Entra Agent ID for each agent. The trade-off is hosting: it runs only as a Microsoft cloud service. If your stack is standardized on Microsoft, it is the fastest path to internal agents. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. For a row-by-row view, see [Dynamiq vs Dify](/compare/dify). ### What happened to Flowise? Flowise, a visual builder for chatbots and agents, announced on July 29, 2026 that it was winding down and froze development that day. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026. The Apache 2.0 source remains available, but npm packages and Docker images are marked deprecated, so it is not a fit for new projects. ## Where Dify fits and what to choose next Dify is a strong, community-backed foundation for LLM apps and RAG, and a sensible default for prototypes. If your roadmap includes voice agents, evaluations on live traffic, hosting agents for many tenants, or a deployment in your own environment with engineers who ship the first agent with your team, put Dynamiq on your shortlist alongside the alternatives above. ## FAQ ### Is Dify open source? Dify's code is public on GitHub under a modified version of the Apache License 2.0. It adds two conditions: you need Dify's written permission to run it as a multi-tenant service, where each workspace counts as a tenant, and you may not remove or change the Dify logo or copyright in its console or apps. That makes it source-available rather than open source in the strict sense. Dynamiq's SDK uses the standard Apache-2.0 license without those conditions; the Dynamiq platform around it is not open source. ### Is Dify free? Dify Cloud has a free Sandbox plan with 200 message credits for trying its core features, and the Community Edition is free to self-host as a single workspace, per Dify's pricing page. Paid Professional, Team and Enterprise plans add credits, workspaces and support. Dynamiq offers a free plan to start, plus custom Enterprise pricing. ### What is the best open-source alternative to Dify? It depends on what you need to be open. Langflow is MIT-licensed and suits Python developers who want a visual builder they can extend. If you want an Apache-2.0 agent engine, with voice agents, evaluations on live traffic and self-hosting in your own environment available on top, look at Dynamiq. Flowise, once a common pick, reached end of life in August 2026. ### Can you self-host Dify? Yes. The Community Edition runs on your own servers as a single workspace. Running Dify for many workspaces or customers needs Dify's written authorization under its license. ### How does Dynamiq compare with Dify? See the [Dynamiq vs Dify comparison](/compare/dify) for license, deployment, voice agents, evaluations and how to migrate. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Dify: [homepage](https://dify.ai/), [pricing](https://dify.ai/pricing), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE), [Human Input node](https://docs.dify.ai/en/cloud/use-dify/nodes/human-input), [MCP publishing](https://docs.dify.ai/en/cloud/use-dify/publish/publish-mcp) - Langflow: [homepage](https://www.langflow.org/), [guardrails](https://docs.langflow.org/guardrails), [traces](https://docs.langflow.org/traces), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [security and compliance](https://zapier.com/security-compliance) - Sana: [homepage](https://sanalabs.com/), [Sana and Workday](https://sanalabs.com/workday), [security](https://sanalabs.com/products/sana/security) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), [what's new](https://learn.microsoft.com/en-us/microsoft-copilot-studio/whats-new) - Flowise: [sunset announcement](https://flowiseai.com/sunset) --- # Best Flowise alternatives for AI workflow automation URL: https://www.getdynamiq.ai/post/best-flowise-alternatives-for-ai-workflow-automation Published: 2025-09-17 Flowise reached end of life on August 31, 2026. Compare the best Flowise alternatives for agents, RAG and chatflows, and plan the move off the archived code. Flowise made it easy to sketch chatbots and agents on a visual canvas, and many teams used it for proofs of concept and internal tools. That chapter has closed. On July 29, 2026 the Flowise team announced it was winding down and froze development the same day. The GitHub repository was archived on August 13, and Flowise reached end of life on August 31, 2026, when official core team presence in Discord and GitHub ended. The Apache 2.0 source stays visible and you can keep running or forking it, but issues and pull requests are locked and npm packages and Docker images are marked deprecated. In practice, every Flowise deployment is now a fixed version that only you maintain. This guide compares the replacements as of September 2026, using each vendor's own pages, listed at the end, and closes with a migration plan. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Flowise](https://flowiseai.com/) (end of life) | Apache 2.0, with a commercially licensed enterprise directory; repository archived | Self-hosted from the archived source | Chatflows and Agentflows; no voice; no further development | Evaluations and tracing integrations as last released | Existing installs planning a migration | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | The closest open-source visual builder | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Visual workflows and RAG with a large plugin marketplace | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Workflow automation with AI as one part | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [Sana](https://sanalabs.com/) | Proprietary | Managed SaaS by default; can deploy on your cloud; single-tenant on Enterprise | No-code agents, workflows and enterprise search across 100+ apps | Analytics dashboard, response feedback, knowledge verification | Company-wide knowledge assistant and learning | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | ## What to look for in a Flowise replacement 1. **A maintained project.** The first requirement after an end of life is a vendor or community that ships fixes. Check release notes and support terms, not only features. 2. **Agentic orchestration that scales.** Flowise's Agentflows covered multi-agent patterns; your replacement should cover planning, tools, memory, retries and human approvals just as well. 3. **Knowledge and RAG depth.** Flowise's document stores and loaders need a home: look for ingestion pipelines, permission-aware retrieval and support for the vector stores you already run. 4. **Evaluations and observability.** Offline and live evaluations, traces with per-step latency and cost, and alerts keep quality high after the move. 5. **Deployment and license.** Decide whether you need self-hosting, a private VPC or an isolated network, and read the license: MIT, Apache 2.0, fair-code and modified Apache terms differ in what you may host and resell. ## Top Flowise alternatives ### Langflow Langflow is the closest match to Flowise's visual, developer-friendly style: an MIT licensed, Python-based builder for agents and RAG applications that can also publish flows as MCP servers and call other MCP servers. It runs as a desktop app, a Python package, a Docker container or a headless runtime on Kubernetes, and it ships frequent releases; recent ones added a Guardrails component, human-in-the-loop checkpoints, built-in traces and OpenTelemetry export. Plan for gaps around built-in evaluations, which are not documented, and enterprise SSO and RBAC enforcement, which are plugins outside open-source Langflow. ### Dify Dify is a visual platform for agentic workflows, chat apps and agents, with Knowledge Pipelines for RAG and a marketplace of plugins for models, tools, data sources and MCP. Apps publish as web apps, API endpoints, embeds or MCP servers, and Dify Enterprise runs in your VPC, on-prem or air-gapped. Check its modified Apache License, which bars running Dify as a multi-tenant service without written permission, and note that we found no built-in evaluation suite in its docs. ### n8n n8n is fair-code workflow automation with AI built in: an AI Agent node with tools and sub-agents, n8n Agents in preview since September 2026, a Guardrails node, human approval before an agent runs a chosen tool, and quick and metric-based evaluations. It fits teams whose Flowise flows were mostly automations with an LLM step. Its Sustainable Use License limits use to internal business, non-commercial or personal purposes. ### Zapier Zapier is the no-code option, connecting 9,000+ apps, with Zapier Agents that draw on knowledge sources, call other agents and pause for approvals, plus chatbots that embed as a widget. It runs only in Zapier's cloud, hosted on AWS in the United States, which rules it out when agents must stay inside your environment. ### Sana Sana, part of Workday since November 2025, is an AI platform for knowledge work: no-code agents, scheduled workflows, enterprise search across 100+ apps, and Sana Learn. It fits if your Flowise use was an internal knowledge chatbot and the goal is now a company-wide assistant rather than custom agents. ### Microsoft Copilot Studio Copilot Studio builds agents and workflows for Teams, Microsoft 365 Copilot and websites, with OpenAI and Anthropic models, agent evaluations and a Microsoft Entra Agent ID for each agent. It runs only as a Microsoft cloud service, so it suits organizations standardized on Microsoft 365. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. For a row-by-row view, see [Dynamiq vs Flowise](/compare/flowise). ## How to plan the move off Flowise 1. **Inventory what runs.** List every chatflow and agentflow in production, the tools and credentials each one uses, and who depends on it. 2. **Freeze and reference the source.** Keep the archived repository and your exported flows as the reference for behavior; nothing upstream will change under you. 3. **Rebuild the flows.** Recreate chatflows and agentflows as agents and workflows in your new platform, starting with the highest-traffic flow. 4. **Move knowledge.** Re-ingest document stores and vector store data, and decide which users and teams each knowledge source should serve. 5. **Recreate controls.** Human Input steps become approvals; moderation nodes become guardrails; add SSO and role-based access where your new plan offers them. 6. **Evaluate before cutover.** Build a test set from real conversations, score the old and new flows side by side, then switch traffic. ## The bottom line Flowise was a friendly canvas for prototyping agents and chatflows, and its Apache 2.0 code will keep running where it is installed. For anything new, choose a maintained platform: Langflow or Dify if you want an open-source or source-available builder, n8n if your flows are really automations, and Dynamiq if you need governed chat, voice and workflow agents in your own environment, with engineers who ship the first one with your team. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Flowise: [homepage](https://flowiseai.com/), [sunset announcement](https://flowiseai.com/sunset), [archived repository](https://github.com/FlowiseAI/Flowise), [license](https://github.com/FlowiseAI/Flowise/blob/main/LICENSE.md), [docs](https://docs.flowiseai.com/) - Langflow: [homepage](https://www.langflow.org/), [install](https://docs.langflow.org/get-started-installation), [guardrails](https://docs.langflow.org/guardrails), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [security and compliance](https://zapier.com/security-compliance) - Sana: [homepage](https://sanalabs.com/), [Sana and Workday](https://sanalabs.com/workday) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio) --- # Best Langflow alternatives for AI workflow and agent orchestration URL: https://www.getdynamiq.ai/post/best-langflow-alternatives-for-ai-workflow-agent-orchestration Published: 2025-09-17 Langflow is an MIT licensed builder for agents, RAG and MCP servers. Compare the best Langflow alternatives in 2026 for evals, governance and deployment. Langflow makes it fast to build agentic AI with a visual editor, batteries-included components and MCP support in both directions. It describes itself as a low-code builder for agentic and RAG applications, and because it is Python-based and MIT licensed, developers can open any component and change it. Many teams start on Langflow for experiments and internal tools. As a project moves toward production, the questions change: how do we evaluate it, who can sign in and change it, where does it run, and who supports it? This guide keeps Langflow front and center, then compares the best alternatives as of September 2026, using each vendor's own pages, listed at the end. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Visual workflows and RAG with a plugin marketplace | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Workflow automation with AI as one part | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [Sana](https://sanalabs.com/) | Proprietary | Managed SaaS by default; can deploy on your cloud; single-tenant on Enterprise | No-code agents, workflows and enterprise search across 100+ apps | Analytics dashboard, response feedback, knowledge verification | Company-wide knowledge assistant and learning | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | ## What to look for beyond the canvas - **Agentic orchestration that scales** - Plan, act and reflect loops, multi-tool routing, retries, memory and fallbacks. - Multi-agent patterns (specialist agents, hand-offs, arbitration) without spaghetti graphs. - **Knowledge and RAG depth** - Ingestion pipelines, chunking and embeddings, retrieval configuration, source governance and citations. - The ability to swap vector stores and model providers without rewiring the app. - **Evaluations and quality loops** - Test sets before launch and metric-based evaluations after it (groundedness, answer correctness, tool-use correctness). - Continuous scoring on live traffic, regression gates, and comparisons across prompts and models. - **Enterprise deployment and security** - SSO and SCIM, roles and workspaces, audit logs. - A choice of SaaS, private VPC, on-prem or an isolated network. ## Top Langflow alternatives, in depth As a baseline, Langflow is an open-source, Python-based framework for building AI applications. It supports agents and the Model Context Protocol as both client and server, so flows can call external tools or be exposed as tools themselves. Here is how it and the alternatives compare. ### Langflow - **Best for:** developers who want an open-source, visual way to build agents, RAG applications and MCP servers, and to customize any component in Python. - **Use when:** you want an MIT license, a local or containerized install, and fast iteration with a large community. - **Skip if:** you need a built-in evaluation suite, which is not documented, or enterprise SSO and RBAC enforcement out of the box; both are plugins outside open-source Langflow. Its voice mode is deprecated as of version 1.10. - **Notable features:** agents and multi-agent flows; knowledge bases on vector databases such as Chroma, OpenSearch or pgvector; a Guardrails component for PII, jailbreaks and prompt injection; human-in-the-loop checkpoints; built-in traces with OpenTelemetry export; agent-to-agent (A2A) support; flows exported as JSON. - **Implementation notes:** Langflow Desktop is the recommended install; Docker, Docker Compose and a headless Kubernetes runtime cover servers. Langflow is listed by IBM as a product of DataStax, an IBM company. ### Zapier - **Best for:** non-technical teams that want agents to act across thousands of SaaS apps with minimal setup. - **Use when:** you need quick automations with company knowledge attached, across Zapier's 9,000+ apps. - **Skip if:** you need self-hosting or a formal evaluation loop; Zapier runs only in its own cloud, hosted on AWS in the United States, and we found no evaluation datasets or scoring for agents in its documentation. - **Notable features:** Zapier Agents with file and app knowledge sources, agent-to-agent calls and approval steps; AI by Zapier steps inside Zaps; embeddable chatbots; an MCP server for AI clients; SOC 2 and SOC 3 reports. ### n8n - **Best for:** automation teams that want Zapier-style integrations plus AI agents, self-hosted or on n8n's cloud. - **Use when:** agents need to call many business apps; you value quick and metric-based evaluations; you already run n8n for operations. - **Skip if:** you need an open-source license without use restrictions; the Sustainable Use License limits use to internal business, non-commercial or personal purposes. - **Notable features:** AI Agent node with tools and sub-agents; n8n Agents in preview since September 2026; a Guardrails node; human approval before an agent runs a chosen tool; MCP client and server; OpenTelemetry tracing. - **Implementation notes:** self-host the free Community Edition or buy Business or Enterprise for SSO and Git version control; knowledge bases are available on n8n Cloud. ### Sana - **Best for:** companies that want a company-wide AI assistant with no-code agents grounded in internal knowledge, plus AI-assisted learning. - **Use when:** the goal is trusted answers and automated knowledge work for every employee, especially for Workday customers; Sana has been part of Workday since November 2025. - **Skip if:** you need open-source code or published pricing; Sana is proprietary, and its pricing page lists custom Enterprise pricing. - **Notable features:** no-code agents; scheduled workflows triggered by time, meetings or email; enterprise search across 100+ apps; an MCP client on Enterprise; Sana Learn for content authoring; SSO and SCIM; managed hosting by default, with deployment on your own cloud or a single-tenant option for Enterprise. ### Dify - **Best for:** product teams that want a visual builder for agentic workflows and RAG with a large plugin marketplace. - **Use when:** you want a free cloud Sandbox or self-hosted Community Edition to prototype, with a path to Enterprise in a VPC, on-prem or air-gapped. - **Skip if:** you plan to host agents for many tenants; Dify's modified Apache License bars multi-tenant operation without written permission. - **Notable features:** Knowledge Pipelines; apps published as web apps, API endpoints, embeds or MCP servers; a Human Input node; apps exported as a YAML DSL; SOC 2 and ISO 27001 per its homepage. ### Microsoft Copilot Studio - **Best for:** enterprises standardized on Microsoft 365 that want agents inside Teams and Microsoft 365 Copilot with Microsoft governance. - **Use when:** Entra identity, Power Platform connectors and Microsoft analytics are requirements; you need orchestration across agents, tools and knowledge sources. - **Skip if:** you need hosting outside Microsoft's cloud; Copilot Studio runs only as a Microsoft online service. - **Notable features:** multi-agent orchestration; OpenAI and Anthropic models; agent evaluations; a Microsoft Entra Agent ID for each new agent; a Visual Studio Code extension for editing agents with Git; Copilot Credits pricing. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. For a row-by-row view, see [Dynamiq vs Langflow](/compare/langflow). ### What happened to Flowise? Flowise, the other visual builder earlier versions of this guide covered, announced on July 29, 2026 that it was winding down. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026, with npm packages and Docker images marked deprecated. It is not a fit for new projects. ## Where Langflow fits and what to choose next Langflow is a strong canvas for building agents, tools and MCP servers in Python, and its recent releases add real production features. If your roadmap includes evaluations on live traffic, enterprise SSO without separate plugins, voice agents, or deployment in your own environment with engineers who ship the first agent with you, pair Langflow with a production platform or pick an alternative above. For teams that want one stack from prototype to production, put Dynamiq on the shortlist. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Langflow: [homepage](https://www.langflow.org/), [docs](https://docs.langflow.org/), [install](https://docs.langflow.org/get-started-installation), [guardrails](https://docs.langflow.org/guardrails), [authorization](https://docs.langflow.org/authorization), [traces](https://docs.langflow.org/traces), [voice mode](https://docs.langflow.org/concepts-voice-mode), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE), [DataStax, an IBM company](https://www.ibm.com/products/datastax) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [security and compliance](https://zapier.com/security-compliance) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/), [Guardrails node](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.guardrails/) - Sana: [homepage](https://sanalabs.com/), [Sana and Workday](https://sanalabs.com/workday), [security](https://sanalabs.com/products/sana/security), [pricing](https://sanalabs.com/products/sana/pricing) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), [Visual Studio Code extension](https://learn.microsoft.com/en-us/microsoft-copilot-studio/visual-studio-code-extension-overview) - Flowise: [sunset announcement](https://flowiseai.com/sunset) --- # Best Sana alternatives for AI knowledge and workflow automation URL: https://www.getdynamiq.ai/post/best-sana-alternatives-for-ai-knowledge-workflow-automation Published: 2025-09-17 Sana, now part of Workday, is an AI platform for knowledge, agents and learning. Compare the best Sana alternatives for agents, RAG, evals and deployment. Sana built an AI-first way to search, organize and learn from company knowledge. It has grown well beyond search: Sana now offers no-code agents, multi-step workflows on a schedule or trigger, enterprise search across 100+ apps, meeting notes, and Sana Learn for training content. In November 2025 Sana became part of Workday, and it continues to work across other systems as well as inside Workday. For teams whose goal is trusted answers and automated knowledge work for every employee, Sana is a strong choice. If your roadmap includes agents inside customer-facing products or back-office operations, voice, evaluations on live traffic, or deployment in your own environment, you will want to compare platforms built for building and running agents. This guide does that as of September 2026, using each vendor's own pages, listed at the end. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Sana](https://sanalabs.com/) | Proprietary | Managed SaaS by default; can deploy on your cloud; single-tenant on Enterprise | No-code agents, workflows and enterprise search across 100+ apps | Analytics dashboard, response feedback, knowledge verification | Company-wide knowledge assistant and learning | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Workflow automation with AI as one part | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Visual workflows and RAG with a plugin marketplace | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | Flowise, which earlier versions of this guide listed, reached end of life on August 31, 2026; see [below](#what-happened-to-flowise). ## What to look for in a Sana alternative 1. **Agentic orchestration.** Moving from answers to execution takes planning, tool use, memory, guardrails and human approvals at the steps that matter. 2. **Knowledge and RAG depth.** Evaluate ingestion, chunking and embeddings, retrieval that respects each user's permissions, source governance and citations, not only the connector list. 3. **Evaluations and observability.** Test sets, scoring with an LLM judge, evaluations on live traffic, and traces with per-step latency and cost keep quality high and spend predictable. 4. **Deployment flexibility.** Managed SaaS is efficient, but many teams need a private VPC, on-prem Kubernetes or an isolated network for sensitive data. ## Top Sana alternatives As a baseline, Sana's strengths are real: semantic search that understands concepts rather than keywords, knowledge verification settings, content authoring in Sana Learn, no-code agents and workflows, SSO and SCIM, and a policy that it does not use customer data to train models. Its pricing page lists custom Enterprise pricing. Here is how the alternatives compare. ### Zapier Zapier is the no-code automation option, connecting 9,000+ apps. Zapier Agents draw on files and app knowledge sources, call other agents and pause for approvals, and chatbots embed as a widget. It runs only in Zapier's cloud, hosted on AWS in the United States, and we found no evaluation datasets or scoring for agents in its documentation. For deep RAG and production telemetry, you will add tooling or pick an agent platform. ### n8n n8n is fair-code workflow automation with AI built in: an AI Agent node with tools and sub-agents, n8n Agents in preview since September 2026, knowledge bases on n8n Cloud, a Guardrails node, human approval before an agent runs a chosen tool, and quick and metric-based evaluations. It suits teams that want Zapier-style automation with self-hosting, under a Sustainable Use License that limits use to internal business, non-commercial or personal purposes. ### Langflow Langflow is an MIT licensed, Python-based builder for agents and RAG applications, with knowledge bases on vector databases such as Chroma, OpenSearch or pgvector, a Guardrails component, human-in-the-loop checkpoints and built-in traces. It runs as a desktop app, in Docker or on Kubernetes. Plan for gaps around built-in evaluations and enterprise SSO, which are not in open-source Langflow. ### Dify Dify is a visual platform for agentic workflows, chat apps and agents, with Knowledge Pipelines that turn files, websites and drives into knowledge bases, and a large plugin marketplace. Dify Enterprise runs in a VPC, on-prem or air-gapped. Its modified Apache License bars multi-tenant hosting without written permission, and we found no built-in evaluation suite in its docs. ### Microsoft Copilot Studio Within Microsoft 365, Copilot Studio builds agents and workflows for Teams, Microsoft 365 Copilot and websites, with multi-agent orchestration, OpenAI and Anthropic models, agent evaluations and a Microsoft Entra Agent ID for each agent. The trade-off is hosting: it runs only as a Microsoft cloud service. If your stack is standardized on Microsoft, it is the quickest path to internal agents. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. ### What happened to Flowise? Flowise, a visual builder for chatbots and agents, announced on July 29, 2026 that it was winding down. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026, with npm packages and Docker images marked deprecated. It is not a fit for new projects. ## The bottom line: picking the right platform Sana is a strong choice for a company-wide AI assistant and for AI-assisted learning, especially for organizations already on Workday. If your roadmap includes agents inside products and operations, voice, RAG with permission-aware retrieval across many sources, evaluations on live traffic, or deployment in your own environment, the alternatives above, and Dynamiq in particular, are built for that job. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Sana: [homepage](https://sanalabs.com/), [Sana and Workday](https://sanalabs.com/workday), [product](https://sanalabs.com/products/sana/), [AI agents](https://sanalabs.com/products/sana/ai-agents), [security](https://sanalabs.com/products/sana/security), [pricing](https://sanalabs.com/products/sana/pricing), [Workday completes acquisition of Sana](https://newsroom.workday.com/2025-11-04-Workday-Completes-Acquisition-of-Sana) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [security and compliance](https://zapier.com/security-compliance) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/) - Langflow: [homepage](https://www.langflow.org/), [docs](https://docs.langflow.org/), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio) - Flowise: [sunset announcement](https://flowiseai.com/sunset) --- # Microsoft Copilot alternatives for enterprise AI teams URL: https://www.getdynamiq.ai/post/microsoft-copilot-alternatives-for-enterprise-ai-teams Published: 2025-08-13 Copilot Studio runs only in Microsoft's cloud. Compare Copilot Studio alternatives in 2026 for self-hosting, model choice, voice agents, RAG and evals. Microsoft Copilot Studio shines when your organization lives in Microsoft 365. It is a low-code studio for agents and workflows that publish to Teams, Microsoft 365 Copilot, websites, mobile apps and other channels, with strong identity through Microsoft Entra (each new agent gets an Entra Agent ID), Power Platform connectors, OpenAI and Anthropic models, agent evaluations and computer use. For internal agents inside the Microsoft ecosystem, it is a strong option. If your roadmap includes agents that run in your own cloud or data center, open models you host yourself, inbound voice agents over SIP or WebRTC, or fine-grained evaluation and observability across many systems, you may want alternatives with more deployment freedom. This guide compares the best Copilot Studio alternatives as of September 2026, using each vendor's own pages, listed at the end, and closes with Dynamiq. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Self-hosted workflow automation with AI as one part | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Visual workflows and RAG with a plugin marketplace | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | Flowise, which earlier versions of this guide listed, reached end of life on August 31, 2026; see [below](#what-happened-to-flowise). ## What to look for in a Copilot Studio alternative ### Deployment you control Copilot Studio is a Microsoft online service, including US government clouds, with no self-hosted option. If data or agents must stay in your own cloud account, data center or an isolated network, that is the first filter. ### Model choice, including models you host Copilot Studio offers OpenAI and Anthropic models, with Mistral and xAI as experimental options hosted outside Microsoft. If you need open models served inside your environment or fine-tuned models, check for an AI gateway and self-hosted serving. ### Open integration surface Make sure you can connect beyond Microsoft 365: databases, vector stores, third-party SaaS, MCP servers and custom tools. ### Evaluations, observability and cost control You will want evaluations built from real traces, scoring on live traffic, traces with per-step latency and cost, and budgets that prevent runaway spend. ## Top Copilot Studio alternatives ### Langflow Langflow is an MIT licensed, Python-based builder for agents and RAG applications that runs as a desktop app, a Python package, a Docker container or on Kubernetes, so it can live entirely inside your environment. It supports many model providers, publishes flows as MCP servers, and recent releases added a Guardrails component, human-in-the-loop checkpoints, built-in traces and OpenTelemetry export. Plan for gaps around built-in evaluations and enterprise SSO, which are not in open-source Langflow. ### n8n n8n is fair-code workflow automation that self-hosts for free or runs on n8n's cloud, with a large library of app integrations, JavaScript and Python code nodes, and substantial AI features: an AI Agent node with tools and sub-agents, n8n Agents in preview since September 2026, a Guardrails node, human approval before an agent runs a chosen tool, and quick and metric-based evaluations. Its Sustainable Use License limits use to internal business, non-commercial or personal purposes. ### Dify Dify is a visual platform for agentic workflows, chat apps and agents, with Knowledge Pipelines for RAG and a marketplace of plugins for models, tools, data sources and MCP. Dify Enterprise runs in your VPC, on-prem or air-gapped, with SSO, SCIM, RBAC and audit logs. Check its modified Apache License, which bars multi-tenant hosting without written permission, and note that we found no built-in evaluation suite in its docs. ### Zapier Zapier remains excellent for quick, no-code automation across 9,000+ apps and now offers Zapier Agents, AI steps in Zaps, chatbots and AI guardrails. It is ideal for marketing ops and multi-app hand-offs, but it runs only in Zapier's cloud, hosted on AWS in the United States, and we found no evaluation datasets or scoring for agents in its documentation. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. For a row-by-row view, see [Dynamiq vs Microsoft Copilot Studio](/compare/microsoft-copilot-studio). ### What happened to Flowise? Flowise, a visual builder for chatbots and agents that earlier versions of this guide covered, announced on July 29, 2026 that it was winding down. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026, with npm packages and Docker images marked deprecated. It is not a fit for new projects. ## When to consider a Copilot Studio alternative - You need agents in your own cloud, data center or an isolated network, not only Microsoft's cloud. - You want open models served inside your environment, or fine-tuned models, behind one gateway. - Your voice agents should answer inbound calls over SIP, WebRTC or WebSocket, with latency measured on every turn. - You require evaluation reports on live traffic, deep tracing and cost per step. - You are standardizing on an open engine to keep your options open. ## Conclusion Copilot Studio is a strong choice for Microsoft-centric enterprises, and its 2026 releases add model choice, evaluations and developer tooling. If you need deployment outside Microsoft's cloud, broader integrations, deeper observability or voice, the alternatives above may fit better. For teams that want the full lifecycle of agentic AI in one place, with engineers who ship the first agent with you, [Dynamiq](/) deserves a serious look at the end of your evaluation. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), [model selection](https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-select-agent-model), [pricing](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio), [real-time voice agents](https://learn.microsoft.com/en-us/microsoft-copilot-studio/voice-realtime-voice-agents), [Entra Agent ID](https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-use-entra-agent-identities), [what's new](https://learn.microsoft.com/en-us/microsoft-copilot-studio/whats-new), [US government](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-licensing-gcc) - Langflow: [homepage](https://www.langflow.org/), [docs](https://docs.langflow.org/), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [security and compliance](https://zapier.com/security-compliance) - Flowise: [sunset announcement](https://flowiseai.com/sunset) --- # Smarter Zapier alternatives to scale AI automation URL: https://www.getdynamiq.ai/post/smarter-zapier-alternatives-to-scale-ai-automation Published: 2025-08-13 Zapier connects 9,000+ apps and runs AI agents in its own cloud. Compare smarter Zapier alternatives in 2026 for self-hosting, voice agents, RAG and evals. Zapier made it easy to connect apps without code, and it remains the go-to for marketing ops and simple SaaS hand-offs. It has also moved into AI: Zapier now describes itself as a way to build and govern AI workflows and agents across 9,000+ apps, with Zapier Agents, AI steps inside Zaps, chatbots, AI guardrails and an MCP server that lets AI clients act in your apps. As teams shift toward agentic AI, retrieval-augmented generation (RAG), evaluation and stricter data controls, a few limits show up. Zapier runs only in its own cloud, hosted on AWS in the United States. We found no evaluation datasets or scoring for agents in its documentation. And pricing counts tasks and agent activities, which adds up for high-volume AI work. This guide compares the main alternatives as of September 2026, using each vendor's own pages, listed at the end. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Self-hosted workflow automation with AI as one part | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Visual workflows and RAG with a plugin marketplace | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | Flowise, which earlier versions of this guide listed, reached end of life on August 31, 2026; see [below](#what-happened-to-flowise). ## What to look for in an AI workflow builder ### Agentic AI and orchestration Look beyond linear tasks: you want multi-step planning, tools, memory, human approvals and evaluators that keep agents on task and catch hallucinations. ### Low-code and code flexibility Drag-and-drop is great until you hit custom logic. Platforms with code nodes or an open SDK on the same engine prevent painful rewrites. ### Observability and cost control LLM calls cost money. Native tracing, per-step latency and cost, and evaluations on live traffic help you debug and avoid runaway bills. ### Enterprise-ready deployment Security teams may require self-hosting, a private VPC or specific regions. Self-hosted and on-prem options keep regulated data where it belongs. ## Top Zapier alternatives ### n8n n8n is the closest match to Zapier's model with more deployment control. It connects a large library of apps and APIs on a node-based canvas, adds JavaScript and Python code nodes, and runs self-hosted for free or on n8n's cloud. Its AI features are now substantial: an AI Agent node with tools and sub-agents, n8n Agents in preview since September 2026, knowledge bases on n8n Cloud, a Guardrails node, human approval before an agent runs a chosen tool, and quick and metric-based evaluations. Note that n8n is fair-code, not open source: its Sustainable Use License limits use to internal business, non-commercial or personal purposes. ### Dify Dify is a visual platform for agentic workflows, chat apps and agents, with Knowledge Pipelines for RAG and a marketplace of plugins for models, tools, data sources and MCP. It offers a free cloud Sandbox, a self-hosted Community Edition, and Dify Enterprise in a VPC, on-prem or air-gapped. Two things to check: its modified Apache License bars running Dify as a multi-tenant service without written permission, and we found no built-in evaluation suite in its docs. ### Langflow Langflow is an MIT licensed, Python-based builder for agents and RAG applications that can publish flows as MCP servers. It runs as a desktop app, a Python package, a Docker container or on Kubernetes, and recent releases added a Guardrails component, human-in-the-loop checkpoints, built-in traces and OpenTelemetry export. It fits developer teams; plan for gaps around built-in evaluations and enterprise SSO, which are not in open-source Langflow. ### Microsoft Copilot Studio Inside the Microsoft 365 universe, Copilot Studio is compelling: agents and workflows for Teams, Microsoft 365 Copilot and websites, multi-agent orchestration, OpenAI and Anthropic models, agent evaluations with a REST API, and a Microsoft Entra Agent ID for each new agent. It is priced in Copilot Credits, and it runs only as a Microsoft cloud service. If your stack is standardized on Microsoft and you accept that constraint, it accelerates internal agents. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. ### What happened to Flowise? Flowise, a visual builder for chatbots and agents that earlier versions of this guide covered, announced on July 29, 2026 that it was winding down. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026, with npm packages and Docker images marked deprecated. It is not a fit for new projects. ## When to consider a Zapier alternative - Your AI or agentic workflows outgrow linear Zaps. - You must store and query proprietary knowledge with permission-aware retrieval. - Compliance requires evaluation reports, tracing and cost per step. - You need self-hosted, on-prem or isolated deployments rather than one vendor's cloud. - Your agents need to answer phone calls. - Task and activity pricing makes high-volume AI work costly. ## Conclusion Zapier remains a strong choice for automating work across a very large app catalog, and its AI features keep improving. AI-native workloads often demand more: agent orchestration, knowledge retrieval, rigorous evaluation, deep observability and deployment in your own environment. n8n, Dify, Langflow and Copilot Studio each cover part of that picture. For teams that want the whole lifecycle in one place, including voice, [Dynamiq](/) is worth a serious look at the end of your evaluation. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [pricing](https://zapier.com/pricing), [security and compliance](https://zapier.com/security-compliance), [MCP](https://zapier.com/mcp), [AI Guardrails](https://help.zapier.com/hc/en-us/articles/43960366238221-How-to-get-started-with-AI-Guardrails-by-Zapier) - n8n: [pricing](https://n8n.io/pricing/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/), [Guardrails node](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.guardrails/) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - Langflow: [homepage](https://www.langflow.org/), [docs](https://docs.langflow.org/), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), [pricing](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio) - Flowise: [sunset announcement](https://flowiseai.com/sunset) --- # LLM guardrails in private equity: how to scale GenAI safely across portfolio companies URL: https://www.getdynamiq.ai/post/llm-guardrails-in-private-equity-how-to-scale-genai-safely-across-portfolio-companies Published: 2025-08-01 How private equity firms set LLM guardrails once and apply them across a portfolio: data boundaries, risk tiers, approvals, audit trails and IC metrics. Private equity firms scale generative AI safely by setting guardrails once, at the fund level, and applying them across every portfolio company. That means hard data boundaries between deals and companies, a risk tier for each use case, human approval for anything investors or counterparties will see, and an audit trail for every run. A shared platform keeps those controls consistent while each company connects its own systems. This guide covers why AI risk looks different in private equity, the guardrail layers that matter, how to tier use cases, a governance checklist and the metrics an investment committee will ask for. ## Why is AI risk different in private equity? A PE firm does not run one AI program. It runs one at the firm and one in every portfolio company, each with its own systems, data and maturity. That creates risks a single company rarely faces: - **Data that must never mix.** Deal documents, data rooms and portfolio financials must stay separated between deals, funds and companies. An assistant that remembers one company's figures in another company's session is a serious breach. - **Material non-public information.** Diligence material and portfolio results are highly sensitive. Where prompts, logs and model context go matters as much as where the documents live. - **Untrusted documents.** Confidential information memorandums and data room files come from sellers and advisers. Text hidden in a document can try to steer an agent that reads it, a form of prompt injection. - **External-facing accuracy.** A wrong figure in an LP report, an investment committee memo or a lender update is a reputational and legal problem, not a typo. - **Different rules per company.** Portfolio companies operate under different privacy laws and sector regulations, from GDPR and the [EU AI Act](/post/how-the-eu-ai-act-will-impact-your-business) to US state privacy laws. - **Short timelines.** Hold periods and value creation plans push teams to move fast, which is exactly when controls get skipped. ## What guardrails does a private equity firm need? Guardrails work in layers. Each layer below addresses a specific risk, with an example from fund and portfolio work: | Layer | Controls | Example | | --------------- | ---------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------- | | Data boundaries | Separate projects, knowledge bases and memory per fund, deal and company; permission-aware retrieval | A diligence assistant for one deal cannot retrieve another deal's data room | | Input screening | Detect PII and prompt injection, including in uploaded documents | A seller's CIM with hidden instructions is flagged before an agent reads it | | Output checks | Citations to source pages, validated formats, numbers checked against source data | Every revenue figure in a memo draft links to the page it came from | | Approvals | Human sign-off before anything leaves the firm, with fields a reviewer can edit | A partner approves the LP update text before it is sent | | Audit trail | Traces of every run, versioned workflows, encrypted logs with retention rules | Compliance can replay what an agent read and wrote for a given memo | | Evaluation | Test sets per use case, scoring of live runs, drift monitoring | Monthly checks that CIM summaries still capture every key term | The action layer matters most once agents do more than draft. An agent that can send an email, update a CRM or file a document needs an approval step on that action, not just a filter on its text. ## How should firms tier AI use cases by risk? Not every use case needs the same controls. An inventory with risk tiers lets teams move fast on low-risk work and apply full controls where they count: | Tier | Typical use cases | Required controls | | ------------------------------------ | ---------------------------------------------------------------------- | ------------------------------------------------------------------ | | Tier 1: external and investor-facing | LP reports, investor communications, lender updates, public statements | Citations, number validation, mandatory approval, full audit trail | | Tier 2: decision support | IC memo drafts, diligence summaries, market and competitor research | Citations, reviewer sign-off, evaluation on test cases | | Tier 3: internal productivity | Meeting notes, first drafts, document search, translations | Data boundaries, logging and spot checks | Assign every new use case a tier before it goes live, and review the tier when its audience or data changes. ## How do you roll guardrails out across portfolio companies? 1. **Build the controls once.** A fund-level platform team defines data boundaries, detectors, approval rules and logging standards as reusable templates. 2. **Give each company its own space.** Separate projects, knowledge bases and credentials per company, so access follows company and role. 3. **Start with shared use cases.** Board pack preparation, contract review and customer support recur across portfolios and reuse the same guardrails. 4. **Let companies connect their own systems.** Each company plugs in its CRM, ERP or ticketing tool; the guardrails stay the same. 5. **Report centrally.** Usage, cost, incidents and evaluation scores roll up to the fund, so the operating team sees adoption and risk across the portfolio. This is also where value creation plans benefit: a playbook proven in one company can be deployed in the next without rebuilding controls. See how firms approach [AI agents in private equity](/post/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy) more broadly. ## What should a responsible AI checklist for PE include? - **Ownership.** A named owner for each use case, and for guardrail tuning, prompt changes and incidents. - **Inventory.** A list of every deployed AI system with its tier, data sources, models and owner. - **Data boundaries.** Documented separation between funds, deals and portfolio companies, tested regularly. - **Log security.** Prompts and outputs often contain deal data, so logs are encrypted, access-controlled and kept only as long as policy allows. - **Audit fields.** For every run: user, data sources, tools called, output, reviewer and decision. - **Model documentation.** For each system: models used, retrieval sources, citation rules and known limitations. - **Incident playbook.** What counts as an incident, who is notified and how fast. - **Validation schedule.** Regular evaluation runs and reviews, especially after model or prompt changes. ## What metrics will the investment committee want? Committees fund what they can measure. Track, for each use case: - **Time saved** on the process it supports, such as hours per CIM screened or per board pack prepared. - **Quality**, measured as evaluation scores and citation accuracy on a fixed test set. - **Incidents and near misses** caught by guardrails or reviewers. - **Cost per task**, including model usage and review time. - **Adoption** across teams and portfolio companies. Put these in the same dashboards and reports the committee already reads. A before-and-after comparison on one process is more persuasive than a portfolio-wide estimate. ## How does Dynamiq support guardrails across a portfolio? Dynamiq gives a fund one platform for building and governing agents, deployable in the firm's own cloud account or data center. Organizations, projects and roles keep each fund team and portfolio company separate, with every request checked against project membership, and knowledge bases use permission-aware retrieval. [Guardrails](/product/guardrails) detect PII and prompt injection before a model reads the text, validators check output formats, and approvals with editable fields gate any tool step, from sending an LP update to writing to a CRM. Every run is traced with inputs, outputs, cost and latency, evaluations score test sets and live traffic, and workflows are versioned with rollback. Dynamiq has SOC 2, HIPAA and GDPR in place, with SSO on the Enterprise plan. ## FAQ ### What are LLM guardrails in private equity? They are the controls that limit what AI systems can read, say and do across a firm and its portfolio companies: data boundaries between deals and companies, input screening, output checks, human approvals and audit trails. ### How do PE firms keep deal data separate when using AI? By giving each fund, deal and portfolio company its own projects, knowledge bases, memory and credentials, and by using retrieval that checks permissions on every query. Test the boundaries regularly, as you would any information barrier. ### Can AI draft LP reports and IC memos? Yes, as a draft. Treat these as the highest risk tier: require citations for every figure, validate numbers against source data, and require a named person's approval before anything is shared. ### What is prompt injection, and why does it matter in diligence? Prompt injection is text crafted to override an AI system's instructions. In diligence, agents read documents from third parties, so a hidden instruction in a data room file could try to change what the agent reports. Screening inputs and limiting what the agent can do contain the risk. ### How do you measure the value of AI guardrails? Track incidents and near misses caught, review time, evaluation scores and adoption. Guardrails that let teams use AI on higher-value work, such as tier 1 and tier 2 processes, show their value in the time saved on those processes. --- # Automating mortgage pre-approval with Dynamiq and Amazon Nova for banks and lenders URL: https://www.getdynamiq.ai/post/automating-mortgage-pre-approval-using-dynamiq-and-amazon-nova-for-financial-and-banking-institutions Published: 2025-07-22 Build a multi-agent mortgage pre-approval workflow on Amazon Nova: four specialist agents, a senior risk analyst, your lending policy and a human sign-off. You can automate most of the mortgage pre-approval review with a team of AI agents that mirrors an underwriting team. Specialists analyze the applicant's finances, credit, income and property in parallel, and a senior risk analyst combines their reports under your lending policy. This tutorial builds that workflow with the Dynamiq SDK and Amazon Nova models on Amazon Bedrock; the result is a recommendation with its reasons, and an underwriter still makes the decision. ## Why automate mortgage pre-approval? Pre-approval is one of the slowest steps in a home purchase. An underwriter pulls bank data, a credit report and income documents, runs the same calculations for every file and writes up a recommendation, often while the buyer waits on it to make an offer. Agents help in three ways: - **Speed.** The analysis starts as soon as the file is complete, and the four specialist reviews run at the same time instead of one after another. - **Consistency.** Every file gets the same checks, calculations and report structure. - **Explainability.** Each agent shows its inputs and formulas, and every run is traced, so the reasons behind a recommendation can be checked. That last point is a legal requirement, not a nice-to-have. In the US, the Equal Credit Opportunity Act and Regulation B require lenders to give applicants specific reasons when they deny credit. A workflow that cannot explain itself cannot be used for credit decisions. ## What does the workflow look like? The workflow mirrors a pre-approval team: 1. **Applicant file.** The loan request, bank account and liability data (for example from Plaid) and a credit report (for example from Experian) are combined into one packet. This tutorial mocks both data sources. 2. **Four specialists, in parallel:** - A **financial analyst** calculates the debt-to-income ratio (DTI). - A **credit analyst** reviews the score, payment history, inquiries and utilization. - An **income verifier** assesses employment stability and verified income. - A **property analyst** checks the value and calculates the loan-to-value ratio (LTV). 3. **A senior risk analyst** reads the four reports, applies your lending policy and writes the recommendation. 4. **An underwriter** reviews the recommendation and decides. ## Step 1: Connect to Amazon Nova on Bedrock Install the SDK with `pip install dynamiq`, and make sure your AWS credentials can call Amazon Bedrock in a region where your chosen Nova model is enabled. You also need an `E2B_API_KEY`, because the specialists run their arithmetic in a code sandbox instead of doing it in their heads. ```python from dynamiq.connections import AWS as AWSConnection from dynamiq.nodes.llms import Bedrock llm = Bedrock( connection=AWSConnection(), # reads AWS credentials and region from the environment model="us.amazon.nova-premier-v1:0", # any Amazon Nova model enabled in your account temperature=0, max_tokens=4000, ) ``` The same workflow runs on any of the 29 model providers the SDK supports; change the LLM node and nothing else. ## Step 2: Assemble the applicant file In production, this data comes from your loan origination system and data providers. Here it is mocked: ```python import json loan_request = { "loan_amount": 240000, "property_value": 300000, "down_payment_percentage": 20, "interest_rate": 6.5, "term_years": 30, "property_type": "single-family residence", "occupancy": "primary residence", "annual_income": 96000, "employment": {"employer": "Example Manufacturing Co.", "years_in_role": 4}, } bank_data = { # shaped like a bank data aggregator's response "accounts": {"checking": 9200.0, "savings": 62500.0}, "liabilities": { "credit_cards": {"total_balance": 1800.0, "minimum_payment": 60.0}, "student_loans": {"total_balance": 12800.0, "monthly_payment": 275.0}, }, } credit_report = { # shaped like a credit bureau's response "credit_score": 688, "credit_history_length_years": 5, "payment_history": {"on_time": "94%", "late_30_days": 2, "late_60_days": 1}, "revolving_utilization": "36%", "hard_inquiries_last_6_months": 3, "verification": {"identity": True, "employment": False, "income": True}, } application_packet = ( f"Loan request:\n{json.dumps(loan_request, indent=2)}\n\n" f"Bank accounts and liabilities:\n{json.dumps(bank_data, indent=2)}\n\n" f"Credit profile:\n{json.dumps(credit_report, indent=2)}" ) ``` ## Step 3: Create the four specialist agents Each specialist gets a focused role, a code interpreter for calculations and the same reporting rules, so the senior analyst receives reports in a consistent shape. ```python from dynamiq.connections import E2B as E2BConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.tools.e2b_sandbox import E2BInterpreterTool from dynamiq.nodes.types import Behavior, InferenceMode REPORTING_RULES = ( " Show each input and formula before the result, and run calculations in the code interpreter." " If data is missing, write 'Data not provided' and explain the effect on confidence." " End with a risk level (Low, Moderate or High) and a recommendation" " (Approve, Approve with conditions, or Decline) with at most two sentences of reasons." ) ROLES = { "Financial analyst": ( "You are a mortgage underwriting financial analyst. List every monthly debt obligation, " "include the proposed mortgage payment, confirm gross monthly income, and calculate the " "debt-to-income ratio (DTI = total monthly debt / gross monthly income x 100)." ), "Credit analyst": ( "You are a credit risk underwriter. Assess the credit score on the 300 to 850 scale, " "payment history, recent hard inquiries and revolving utilization." ), "Income verifier": ( "You are an income verification specialist. Assess employment stability, which income " "sources are verified or unverified, and any inconsistencies an underwriter should resolve." ), "Property analyst": ( "You are a property valuation analyst. Check the property value, calculate the loan-to-value " "ratio (LTV = loan amount / property value x 100), state whether mortgage insurance applies, " "and assess collateral risk." ), } specialists = [ Agent( name=name, role=role + REPORTING_RULES, llm=llm, tools=[E2BInterpreterTool(connection=E2BConnection())], # reads E2B_API_KEY inference_mode=InferenceMode.XML, max_loops=8, behaviour_on_max_loops=Behavior.RETURN, ) for name, role in ROLES.items() ] ``` ## Step 4: Add your lending policy and the senior risk analyst The senior analyst runs after all four specialists and receives their reports together with your institution's policy. The policy below is illustrative; replace it with your own credit policy. ```python LENDING_POLICY = """Illustrative pre-approval policy: - Maximum DTI 43%; 36% or lower preferred. - Minimum credit score 640. - Maximum LTV 95%; mortgage insurance required above 80%. - If any major factor is weak, at least two strong compensating factors must exist: a credit score above 760, a down payment above 20%, reserves above 6 months, or an LTV below 75%.""" SENIOR_ANALYST_ROLE = ( "You are the senior risk analyst and write the recommendation an underwriter will review. " "Rely on the specialist reports and the lending policy; do not re-analyze raw data. " "Write a report with: an overall risk score out of 100 with a label; a recommendation " "(Recommend pre-approval, Recommend with conditions, or Recommend decline); the specific " "reasons, citing DTI, credit, income and LTV; the maximum loan amount you recommend; " "and the conditions an underwriter must verify." ) def compile_reports(inputs: dict, outputs: dict) -> str: reports = "\n\n".join(f"## {agent.name}\n{outputs[agent.id]['content']}" for agent in specialists) return f"{reports}\n\n## Lending policy\n{LENDING_POLICY}" risk_analyst = ( Agent( name="Senior risk analyst", role=SENIOR_ANALYST_ROLE, llm=llm, max_loops=4, behaviour_on_max_loops=Behavior.RETURN, ) .inputs(input=compile_reports) .depends_on(specialists) ) ``` ## Step 5: Run the workflow The four specialists share no dependencies, so the flow runs them in parallel; the senior analyst starts when all four have finished. ```python from dynamiq import Workflow from dynamiq.flows import Flow workflow = Workflow(flow=Flow(nodes=[*specialists, risk_analyst])) result = workflow.run(input_data={"input": application_packet}) print(result.output[risk_analyst.id]["output"]["content"]) ``` This applicant is a borderline case by design. The DTI is comfortable, but the LTV sits at the 80% threshold, the credit score clears the policy minimum without being strong, and employment is not yet verified. A good report reflects that with a moderate risk score and conditions for the underwriter to clear, such as employment verification and recent pay stubs. ## What does it take to run this at a lender? The workflow above is the analytical core. A production deployment adds controls that let risk, compliance and model governance teams sign off: - **A human decision.** On the Dynamiq platform, an [approval](/product/guardrails) pauses the workflow before a recommendation leaves the system, and the underwriter can edit its fields before it proceeds. - **A complete record.** Every run is traced, with each agent's inputs, outputs, tool calls, cost and latency, which gives you the reasons behind every recommendation; see [observability](/product/observability). - **Measured consistency.** Build a dataset from past applications with known outcomes, score the workflow against it with evaluations, and attach real-time evals to the live deployment. - **Fair lending checks.** Keep protected characteristics out of the agents' inputs, test outcomes across applicant groups, and review the prompts and policy with your compliance team. - **Your data, your perimeter.** Detect PII before it reaches a model, and run Dynamiq in your own AWS account; it is available on AWS Marketplace and can be self-hosted, as described on [how we deploy](/deployment). Plaid is among the platform's app integrations, and credit bureau or loan origination APIs connect through an HTTP tool, so the mocked data sources become live ones without changing the agents. For the broader pattern of reading loan files, claims and contracts against your rules, see [document review and underwriting](/use-cases/document-review). ## FAQ ### Can AI approve a mortgage on its own? It shouldn't. AI agents can analyze an application and recommend a decision with reasons, but a qualified underwriter should make the credit decision. Regulations such as the Equal Credit Opportunity Act also require lenders to explain denials with specific reasons. ### Why use several agents instead of one? Each specialist has a narrow role and a short prompt, which makes its analysis easier to test and audit. Running them in parallel is also faster, and a failure in one review is visible in its own report instead of buried in a single long answer. ### Which models can I use for mortgage analysis? This tutorial uses Amazon Nova through Amazon Bedrock, which keeps inference inside AWS. The Dynamiq SDK supports 29 model providers, so you can use the model your institution has approved and compare candidates with evaluations before you switch. ### How do you explain an AI-assisted credit recommendation? Require every agent to show its inputs, formulas and reasons, keep the lending policy explicit in the workflow, and keep the trace of each run. Together they show which data and which rule led to the recommendation. ### Can this run inside our own AWS account? Yes. Dynamiq can be self-hosted in your AWS account on Kubernetes and is available on AWS Marketplace, and Amazon Bedrock keeps model calls within AWS. --- # How to create an AI agent from scratch: a step-by-step guide URL: https://www.getdynamiq.ai/post/how-to-create-an-ai-agent-from-scratch-step-by-step-guide Published: 2025-07-17 Build an AI agent step by step: define the job, choose a model, add tools, knowledge, memory and guardrails, then test and deploy. With working Python code. To create an AI agent from scratch, you connect a language model to tools and data, give it clear instructions, and run it in a loop where it reasons, acts and checks the result until the task is done. What makes an agent reliable is everything around that loop: a narrow job, well-described tools, grounded knowledge, memory, guardrails and tests on real cases. This guide explains the components, walks through eight build steps, and ends with a working Python example of a bank card support agent that looks up data, remembers the conversation and asks a person before it blocks a card. ## What is an AI agent? An AI agent is software that uses a language model to decide what to do next, takes actions through tools such as APIs and databases, observes the results and repeats until it reaches its goal. The model supplies judgment; the tools let it act; the loop lets it recover from a wrong step. It helps to compare agents with the two things they are often confused with: | | Chatbot | Workflow automation | AI agent | | ---------------------- | -------------------- | ---------------------------- | -------------------------------- | | How it decides | Answers each message | Follows fixed rules | Plans steps toward a goal | | Can it act in systems? | Rarely | Yes, along a fixed path | Yes, choosing which tool to use | | Handles the unexpected | Poorly | Fails or stops | Adapts, or escalates to a person | | Best for | Simple questions | Stable, repetitive processes | Varied tasks that need judgment | Most production systems combine them: a workflow with fixed checkpoints, and agents doing the steps that need judgment. Our guide to [LLM agents](/post/llm-agents-explained-complete-guide-in-2025) covers the concepts in more depth. ## What are the core components of an AI agent? | Component | What it does | Example in a card support agent | | ------------------- | ----------------------------------------------- | --------------------------------------------- | | Model | Reasons and chooses the next step | A hosted or self-hosted chat model | | Instructions | Set the scope, rules, tone and when to escalate | "Never ask for a full card number or PIN" | | Tools | Let the agent look things up and take actions | Card status lookup, block card | | Knowledge | Documents the agent retrieves and cites | Fee schedule and card policy | | Memory | Keeps context within and across conversations | What the customer said two messages ago | | Guardrails | Check inputs, outputs and actions | Detect personal data, approve before blocking | | Runtime and tracing | Runs the loop and records every step | A trace of each lookup, with cost and latency | ## Should you build from scratch, use a framework, or use a platform? "From scratch" can mean three different things, and the right one depends on your team: | Approach | What you write | Best for | Watch out for | | --------------- | --------------------------------------------------- | ------------------------------------------- | --------------------------------------------- | | Raw model API | The loop, tool calling, retries, memory and logging | Learning how agents work | Everything production needs is yours to build | | Open-source SDK | Agent configuration, tools and business logic | Engineering teams that want control in code | You still host, secure and monitor it | | Platform | Configuration on a canvas, code where needed | Mixed teams and regulated production | Check self-hosting options and lock-in | Dynamiq covers the last two with one engine: the open-source [Python SDK](/open-source) (Apache-2.0) and a visual builder that run the same agents, so a team can start in code and hand over to a canvas, or the other way around. ## How do you build an AI agent? Eight steps ### Step 1: Define one job and how you will measure it Start with a single, narrow task that happens often and has a clear definition of done, such as "answer card questions and block cards on request" rather than "help customers". Write down what success looks like before you build: resolution rate, accuracy on a set of real cases, time to answer, and which requests must go to a person. ### Step 2: Choose the model Pick the smallest model that handles your task reliably, and test two or three candidates on the same cases. Consider where the model runs: hosted models are easiest to start with, while open-weight models such as Llama, Mistral or Qwen can run inside your own environment when data cannot leave it. Keep the model swappable; it will change more often than anything else in your agent. ### Step 3: Write the instructions The instructions, often called the role or system prompt, define the agent's scope, rules and style. Be specific about what it must never do and when it must hand off. Vague instructions produce vague agents. ### Step 4: Give it tools Tools are how an agent acts: API calls, database queries, search, code execution, or other agents. Each tool needs a clear name, a description the model can act on and a typed input schema. Start with read-only tools and add actions once you trust the agent. Many tools can now be connected through the [Model Context Protocol](/glossary/model-context-protocol), an open standard for exposing tools to agents. ### Step 5: Ground it in your knowledge For questions about your products, policies or documents, give the agent retrieval over a knowledge base instead of relying on what the model learned in training. Retrieval-augmented generation (RAG) returns relevant passages the agent can quote and cite. In regulated settings, retrieval should respect permissions, so the agent only reads what the person it serves is allowed to see. See how [knowledge bases](/product/knowledge-rag) handle parsing, retrieval and permissions. ### Step 6: Add memory Session memory keeps the conversation coherent, so a customer does not repeat themselves. Long-term memory stores durable facts about a user, such as a preferred language, and recalls them in later sessions. Scope both to the user and session, and decide what should never be stored. ### Step 7: Add guardrails and approvals Screen inputs for personal data and prompt injection before the model sees them, validate outputs against the format and policy you expect, and require a person's approval before any step that is hard to undo, such as moving money, changing an account or sending a letter. Our [guardrails](/product/guardrails) page covers the detectors and approval gates in detail. ### Step 8: Test, deploy and monitor Build a test set from real cases, including edge cases and attempts to misuse the agent, and run it on every change to instructions, tools or models. After launch, trace every run, watch cost and latency, and score a sample of live conversations so quality problems surface before customers report them. Deploy with versions, so you can roll back a bad change in one step. ## Build it in Python: a working example This example uses the open-source Dynamiq SDK (`pip install dynamiq`) and an OpenAI model. It builds a card support agent with two tools, conversation memory and an approval step before a card is blocked. The card data is a stand-in for your core banking API. ```python from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.flows import Flow from dynamiq.memory import Memory from dynamiq.memory.backends import InMemory from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.tools.function_tool import function_tool from dynamiq.nodes.types import InferenceMode from dynamiq.types.feedback import ApprovalConfig, FeedbackMethod # A stand-in for your core banking API CARDS = {"4821": {"status": "active", "last_transaction": "Online store, 214.90 USD"}} @function_tool def get_card_status(card_last4: str, **kwargs) -> dict: """Look up a card's status and last transaction by its last four digits.""" return CARDS.get(card_last4, {"error": "Card not found"}) @function_tool def block_card(card_last4: str, reason: str, **kwargs) -> str: """Block a card. Use only after the customer has confirmed they want it blocked.""" CARDS[card_last4]["status"] = "blocked" return f"Card ending {card_last4} is now blocked." # Every block waits for a person to approve it block_card_tool = block_card( approval=ApprovalConfig( enabled=True, feedback_method=FeedbackMethod.CONSOLE, msg_template=( "Block card ending {{input_data.card_last4}} " "(reason: {{input_data.reason}})? " "Press Enter to approve, or type feedback to cancel." ), ) ) llm = OpenAI(connection=OpenAIConnection(), model="gpt-4.1", temperature=0.1) agent = Agent( name="card-support", id="card-support", llm=llm, role=( "You help bank customers with card questions.\n" "- Look up the card with get_card_status before you answer.\n" "- Block a card only when the customer asks for it and confirms.\n" "- Never ask for a full card number, PIN or password.\n" "- For disputes and anything else, say a colleague will follow up." ), tools=[get_card_status(), block_card_tool], memory=Memory(backend=InMemory()), inference_mode=InferenceMode.XML, max_loops=8, ) workflow = Workflow(flow=Flow(nodes=[agent])) session = {"user_id": "customer-17", "session_id": "chat-001"} for message in [ "I don't recognize the last payment on my card ending 4821.", "That wasn't me. Please block the card.", ]: result = workflow.run(input_data={"input": message, **session}) print(result.output[agent.id]["output"]["content"]) ``` What each part does: - `@function_tool` turns a Python function into a tool. The type hints become the input schema and the docstring becomes the description the model reads. - `ApprovalConfig` pauses the `block_card` tool until a person approves in the console. If they type feedback instead, the step is canceled and the agent sees the feedback. - `Memory` with `user_id` and `session_id` lets the second message refer to "the card" without repeating the number. - `InferenceMode.XML` makes the model format its reasoning and tool calls as XML, which works reliably across providers. To swap the model, change the LLM class, for example to `Anthropic` or `TogetherAI`; the SDK supports 29 model providers. The [SDK documentation](https://docs.getdynamiq.ai/docs/sdk) covers knowledge retrieval, long-term memory, sandboxes and multi-agent orchestration. ## How do you build the same agent without code? In Dynamiq's [Agent Builder](/product/agents), the same agent is a set of nodes on a visual canvas: an agent node with its model, instructions, tools and memory, wired between an input and an output. You can describe the workflow in a prompt and get a first draft to edit, add Choice nodes for rules, and turn on approval for any tool step, with fields a reviewer can edit before it runs. When it works, deploy it as an App with its own API, a chat widget or a schedule, with versions and rollback. Guardrails that detect PII and prompt injection, evaluations and traces with cost and latency are part of the platform, which runs in Dynamiq Cloud or self-hosted in your own cloud or data center. If you want an agent that works today without designing one, [AI Coworker](/product/chat) is a ready-made agent with its own cloud sandbox, connectors to your apps and scheduled tasks, in the browser, Slack, Microsoft Teams and Telegram. ## Where do AI agents deliver value first? - **Customer support:** resolving routine requests and handing complex ones to a person with context. An Asian neo-bank [automated about 85% of support inquiries](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) this way and went live in 30 days. - **Back-office operations:** reconciliations, exception handling and report drafting, with approvals before anything posts. - **Document review:** extracting and checking fields in loan files, claims and contracts. - **Compliance:** preparing KYC and AML cases for analysts to decide. - **IT and HR help desks:** answering policy questions and handling routine requests such as access and onboarding tasks. ## FAQ ### How do I create an AI agent from scratch? Define one narrow job and how to measure it, choose a model, write instructions, add tools for the actions and lookups the job needs, ground the agent in your knowledge, add memory, put guardrails and approvals on risky steps, and test it on real cases before you deploy and monitor it. ### Do I need to train a model to build an AI agent? No. Most agents use a pre-trained model and get their behavior from instructions, tools and retrieved knowledge. Fine-tuning is optional and usually reserved for narrow needs such as a specific output format or domain vocabulary. ### What programming language is best for building AI agents? Python is the most common choice because most agent SDKs, model clients and data libraries support it first. The Dynamiq SDK is a Python library, and agents built with it also run on the visual platform. ### How long does it take to build an AI agent? A prototype comes together quickly. Most of the time goes into connecting real systems, testing on real cases and passing security review. With an experienced team and a platform, a production agent can go live in weeks; the neo-bank above went live in 30 days. ### Can I build an AI agent without coding? Yes. Visual builders let you configure the model, instructions, tools, memory and approvals on a canvas, and generate a first draft from a prompt. Keep a code path available for custom tools and integrations. --- # Best n8n alternatives in 2026: free and paid options URL: https://www.getdynamiq.ai/post/best-n8n-alternatives-for-ai-workflow-automation Published: 2025-07-14 Compare the best n8n alternatives in 2026, free and paid, from Langflow and Dify to Zapier and Dynamiq, on license, deployment, agents and evaluations. n8n made its name with visual workflow automation: a node-based canvas that connects apps and APIs, with JavaScript or Python whenever you need it. In 2026 it leans hard into AI, with AI Agent nodes, a Guardrails node, human approval before an agent runs a tool, evaluations and, since September 2026, standalone n8n Agents in preview. It is still a workflow tool first, and it is source-available under its fair-code Sustainable Use License rather than open source. Teams look for an n8n alternative for a few specific reasons: an open-source license without use restrictions, agents that answer phone calls, evaluations on live production traffic, or engineers who ship the first agent inside their own environment. This guide compares the main options as of September 2026, using each vendor's own pages, which are listed at the end. ## Quick comparison | Platform | License | Deployment | Agents and voice | Evals and observability | Best fit | | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | [n8n](https://n8n.io/) | Sustainable Use License (fair-code, source-available) | n8n Cloud, or self-hosted Community, Business and Enterprise (air-gapped on Enterprise) | AI Agent nodes and n8n Agents in preview; no voice agent product | Quick and metric-based evaluations, Insights, OpenTelemetry tracing | Workflow automation across many apps, with AI as one part | | [Dynamiq](/) | Apache-2.0 SDK; the platform is not open source | Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, OpenShift, any Kubernetes cluster or on-prem | Chat, workflow and inbound voice agents on one engine | Evals, including real-time evaluations on live deployments; traces with cost and latency per node | Regulated teams that want governed agents in their own environment | | [Zapier](https://zapier.com/) | Proprietary | Zapier's cloud only, hosted on AWS in the US | Zapier Agents, AI steps in Zaps and chatbots across 9,000+ apps | Agent drafts and versions; analytics and an Observability API on Enterprise | No-code automation for business teams | | [Dify](https://dify.ai/) | Modified Apache 2.0 (source-available; no multi-tenant hosting without permission) | Dify Cloud, self-hosted Community Edition, or Enterprise in a VPC, on-prem or air-gapped | Agentic workflows, chat apps and agents; no voice agent product | Logs, annotations and tracing integrations; no built-in eval suite described | Fast prototyping of LLM apps and RAG | | [Langflow](https://www.langflow.org/) | MIT | Desktop app, Python package, Docker or Kubernetes | Agents, multi-agent flows, MCP server and client; voice mode deprecated | Built-in traces and OpenTelemetry; no built-in evals documented | Python developers building agents, RAG apps and MCP servers | | [Copilot Studio](https://www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio) | Proprietary service with MIT-licensed tooling | Microsoft cloud only, including US government clouds | Chat and autonomous agents, workflows, computer use; voice through Dynamics 365 Contact Center | Agent evaluations and analytics | Organizations standardized on Microsoft 365 | Flowise, which earlier versions of this guide listed, reached end of life on August 31, 2026 and is no longer an option for new projects. More on that [below](#what-happened-to-flowise). ## What to look for in an n8n alternative ### Agents, not only nodes Look for planning, tool use, memory and multi-agent patterns as first-class features, plus guardrails and human approvals at the steps that matter. A linear chain of nodes with an LLM step in the middle is not the same as an agent that decides which tools to call. ### Low-code and code on one engine Drag-and-drop is fast until you hit an edge case. Platforms where the visual builder and an SDK share one engine let developers drop into code without rewriting what business users built. ### Evaluations and observability in production LLM calls cost money and fail in new ways. You want datasets built from real traces, scoring with an LLM judge, evaluations on live traffic, and traces with cost and latency for every step. ### Deployment and license Security teams may require your own cloud, on-prem Kubernetes or an isolated network. Read the license too: fair-code and source-available licenses can restrict how you run or resell the software. ## Top n8n alternatives ### Zapier Zapier is the no-code option. It describes itself as a way to build and govern AI workflows and agents across 9,000+ apps. Zapier Agents draw on files and app sources such as Google Drive, Notion and Confluence, can call other agents and pause for approval steps, and are moving into AI by Zapier inside the Zap editor. Zapier also offers embeddable chatbots, Tables, Forms and an MCP server that lets AI clients act in those apps. Team and Enterprise plans add SSO, Enterprise adds SCIM, analytics and an Observability API, and Zapier maintains SOC 2 and SOC 3 reports. The limits for AI-heavy teams: Zapier runs only in its own cloud, hosted on AWS in the United States, and we found no evaluation datasets or scoring for agents in its documentation. Choose Zapier for fast automation across a very large app catalog; look elsewhere if agents must run inside your own environment. ### Dify Dify calls itself a platform for production-ready agentic workflows. Its visual builder covers workflows, chat apps, agents and Knowledge Pipelines for RAG, with a marketplace of plugins for models, tools, data sources and MCP. Apps publish as web apps, API endpoints, embeds or MCP servers, and Dify Enterprise runs in your VPC, on-prem or air-gapped, with SSO, SCIM, RBAC and audit logs. Two things to check: the license is a modified Apache 2.0 that bars running Dify as a multi-tenant service without written permission, and we found no built-in evaluation suite in its docs. Choose Dify for fast LLM app and RAG prototyping with a large community. ### Microsoft Copilot Studio Copilot Studio is Microsoft's low-code studio for agents and workflows that publish to Teams, Microsoft 365 Copilot, websites and other channels. It supports multi-agent orchestration, computer use, OpenAI and Anthropic models, agent evaluations (generally available since March 2026) and a Microsoft Entra Agent ID for each new agent. Pricing is in Copilot Credits rather than seats: $200 for 25,000 credits a month, or $0.01 per credit pay-as-you-go. The trade-off is hosting. Copilot Studio runs only as a Microsoft cloud service, including US government clouds, with no self-hosted option. Choose it if your organization is standardized on Microsoft 365. ### Langflow Langflow is an MIT licensed, Python-based builder for agents and RAG applications that can also publish flows as MCP servers. It runs as a desktop app, a Python package, a Docker container or a headless runtime on Kubernetes, and recent releases added a Guardrails component, human-in-the-loop checkpoints, built-in traces and OpenTelemetry export. Gaps to plan for: no built-in evaluation suite is documented, its voice mode is deprecated, and full SSO and RBAC enforcement are plugins that open-source Langflow does not include. Choose it for developer-led projects that stay close to Python. ### n8n, when it still fits n8n remains a strong choice when AI is one part of broader automation. It has a large integration and template library, JavaScript and Python code nodes, an AI Agent node with tools and sub-agents, a Guardrails node, human approval before an agent runs a chosen tool, quick and metric-based evaluations, and OpenTelemetry tracing. Self-hosting is free under the Sustainable Use License, which limits use to internal business, non-commercial or personal purposes; SSO and Git version control come with the Business and Enterprise plans. ### Dynamiq Dynamiq is a platform for chat, voice and workflow agents, with guardrails, evals and observability on the same engine, deployed in Dynamiq Cloud or inside your own environment. The agent engine is an open-source Python SDK under Apache-2.0; the platform around it is not open source. - [**Agent Builder**](/product/agents): a visual canvas and the open-source Python SDK on one engine, with Choice and rules nodes, approvals with editable fields on any tool step, workflows generated from a prompt, and versions with rollback. - [**AI Coworker**](/product/chat): an agent with its own cloud sandbox computer (Linux shell, files, browser automation and live web previews), connectors, a skills library, scheduled tasks and subagents, in the browser and in Slack, Microsoft Teams and Telegram. - [**Voice Agents**](/product/voice-agents): inbound calls over WebRTC, SIP and WebSocket, with per-turn latency measurement and simulations scored by an LLM judge. - **Agentic memory**: session memory scoped by user and session, with recent, relevant or both retrieval, and long-term memory where the agent itself decides when to save and recall durable, user-scoped facts through remember and recall tools, updating an existing fact rather than saving a near duplicate. - [**Knowledge**](/product/knowledge-rag): ingestion pipelines, permission-aware retrieval filters, knowledge graphs, and document parsing and extraction. - [**Integrations**](/product/integrations): 2,100+ app integrations and 11,800+ actions, plus MCP servers and per-user connections. - [**Models and AI Gateway**](/product/models): 29 model providers in the SDK and the most-used models behind one OpenAI-compatible AI Gateway endpoint, plus open models served on vLLM and LoRA fine-tuning. - [**Guardrails and approvals**](/product/guardrails): detect PII and prompt injection, with policy detection, validators and execution approvals. - [**Evals**](/product/evaluations) and [**observability**](/product/observability): datasets built from traces, LLM-as-a-judge, predefined and code metrics, and real-time evaluations on live deployments; traces with cost and latency per node, monitoring, and an AI agent that surfaces issues from production traces. - [**Deployment**](/product/deployments): Dynamiq Cloud, or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift or any Kubernetes cluster, or on-prem, with self-hosting on Enterprise. Fully isolated or air-gapped setups are delivered with our engineers. Available on AWS Marketplace, Microsoft Azure Marketplace and IBM Marketplace. - **Security and compliance**: SOC 2, HIPAA and GDPR. Guardrails and approvals come on every plan, and SSO (OIDC with Okta and Microsoft Entra ID) on Enterprise. - [**Pricing**](/pricing): free to start ($0), and Enterprise with custom pricing. - **Forward-deployed engineers**: a bootcamp in your environment, your first production agent shipped with your team, then a handover. For a row-by-row view, see [Dynamiq vs n8n](/compare/n8n). ### What happened to Flowise? Flowise, a visual builder for chatbots and agents that earlier versions of this guide recommended, announced on July 29, 2026 that it was winding down and froze development that day. Its GitHub repository was archived on August 13, and it reached end of life on August 31, 2026. The Apache 2.0 source remains available, but npm packages and Docker images are marked deprecated, so it is not a fit for new projects. See [Dynamiq vs Flowise](/compare/flowise) for migration notes. ## When to consider an n8n alternative - Your agents need to answer phone calls, not only chat. - You need an open-source license without limits on commercial use, or hosting rights without a separate agreement. - Compliance asks for evaluations on live traffic and traces with cost and latency for every step. - You want engineers to ship the first production agent inside your environment. - AI agents, not app-to-app automation, are the main workload. ## Conclusion n8n is excellent for automating work across many apps, and its 2026 releases make it a credible place to run AI steps too. Zapier, Dify, Langflow and Copilot Studio each cover part of the picture. If your roadmap centers on governed agents, including voice, evaluated in production and deployed in your own environment, compare [Dynamiq](/) side by side and read the [open-source SDK on GitHub](https://github.com/dynamiq-ai/dynamiq). ## FAQ ### Is n8n free? The Community Edition is free to self-host under n8n's Sustainable Use License, a fair-code license that allows internal business, non-commercial and personal use. n8n Cloud offers a free trial rather than a free plan, with paid Starter, Pro, Business and Enterprise plans, per n8n's pricing page. ### What is the best free alternative to n8n? For a free, self-hosted builder for LLM apps and agents, Langflow (MIT) and Dify's Community Edition (free, source-available) are the closest options. Dynamiq offers a free plan to start, plus custom Enterprise pricing, and its Apache-2.0 SDK is free to use in your own code. ### Is there an open-source alternative to n8n? n8n is source-available rather than open source in the strict sense, because its license limits commercial use. Langflow (MIT) and Dynamiq's SDK (Apache-2.0) are open source without those limits. Dify uses a modified Apache license that needs permission for multi-tenant hosting. ### How does Dynamiq compare with n8n? See the [Dynamiq vs n8n comparison](/compare/n8n) for license, deployment, agents, evaluations and how to migrate. ## Sources Vendor facts were checked on September 29, 2026 against these pages: - n8n: [pricing](https://n8n.io/pricing/), [security](https://n8n.io/security/), [license](https://github.com/n8n-io/n8n/blob/master/LICENSE.md), [agents](https://docs.n8n.io/build/build-and-manage-agents/), [Guardrails node](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.guardrails/), [evaluations](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/) - Zapier: [homepage](https://zapier.com/), [Agents](https://zapier.com/agents), [pricing](https://zapier.com/pricing), [security and compliance](https://zapier.com/security-compliance) - Dify: [homepage](https://dify.ai/), [Enterprise](https://dify.ai/enterprise), [license](https://github.com/langgenius/dify/blob/main/LICENSE) - Microsoft Copilot Studio: [overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), [pricing](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio), [what's new](https://learn.microsoft.com/en-us/microsoft-copilot-studio/whats-new) - Langflow: [homepage](https://www.langflow.org/), [docs](https://docs.langflow.org/), [license](https://github.com/langflow-ai/langflow/blob/main/LICENSE) - Flowise: [sunset announcement](https://flowiseai.com/sunset), [archived repository](https://github.com/FlowiseAI/Flowise) --- # Enterprise AI agents: benefits, use cases, implementation guide URL: https://www.getdynamiq.ai/post/enterprise-ai-agents-benefits-use-cases-implementation-guide Published: 2025-07-09 What enterprise AI agents are, the main types, what makes an agent enterprise-ready, where agents deliver value, and how to implement them without stalling. Enterprise AI agents are software systems that pursue a business goal with some autonomy: they reason with a language model, use your tools and data, and act within the permissions and approvals you set. Unlike chatbots and scripts, they complete multi-step work across systems, such as resolving a support case or reconciling an invoice. What makes an agent enterprise-ready is less the model than the controls around it: integration, security, observability and a person on every decision that matters. ## What are AI agents? An [AI agent](/glossary/ai-agent) is a system that carries out tasks with a degree of autonomy. At its core is a decision-making loop, usually driven by a large language model: the agent interprets the goal, plans steps, calls tools, observes the results and decides what to do next until the task is done. In an enterprise, that loop runs against real systems. Agents query databases, call APIs, read documents, send messages and update records, recover from errors along the way, and keep short-term and long-term memory of the work. ## What types of AI agents are there? Classic AI research groups agents by how they perceive, plan and decide. The categories still help you pick the right design for a business task: | Type | How it decides | Enterprise example | | ------------- | ---------------------------------------------- | ----------------------------------------------------- | | Reflex | Maps a condition to an action ("if X, do Y") | Threshold alerts in monitoring | | Goal-based | Chooses actions that move it toward a goal | Contract approval that varies by value and risk | | Utility-based | Weighs outcomes and picks the best trade-off | Pricing that balances margin, conversion and churn | | Learning | Improves from feedback on its results | Alert tuning based on analysts' dispositions | | Hierarchical | Splits work between planners and executors | A planner agent delegating to specialist agents | | Multi-agent | Several agents coordinate toward a shared goal | Research, validation and writing agents on one report | ### Reflex agents Reflex agents respond to specific inputs with pre-defined actions. Simple reflex agents act only on the current input, while model-based reflex agents keep limited memory to make better-informed decisions. They suit monitoring, threshold-triggered alerts and static rule sets. ![Reflex agent diagram: if condition X, then action Y, with simple reflex agents that keep no memory and model-based reflex agents that keep limited memory](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/2-2-reflex-agents-fast-rule-based-responses-for-mo.webp "Reflex agents map conditions to actions.") ### Goal-based agents Goal-based agents evaluate sequences of actions by whether they move closer to a target state. They fit tasks with branching logic, such as a contract approval process where the steps vary by value, department or risk. ![Goal-based agent diagram: memory, tools and goals feed the agent, which acts on its environment and takes in observations](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/3-3-goal-based-agents-choosing-actions-to-reach-ta.webp "Goal-based agents use memory, tools and goals to choose actions, then observe the results.") ### Learning agents Learning agents improve with feedback on their outcomes. In the classic model, a critic scores results against a performance standard and a learning element adjusts behavior. With today's language-model agents, the practical version is an evaluation loop: reviewers' corrections become test cases, and teams update instructions, tools or fine-tuned models. ![Learning agent diagram with a critic, learning element, performance element and problem generator connected to the environment through sensors and effectors](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/4-4-learning-agents-improving-decisions-through-ex.webp "The classic learning agent model, adapted from Russell and Norvig.") ### Utility-based agents Utility-based agents compare possible outcomes and choose the one with the highest expected value, which suits decisions with trade-offs, such as a retail pricing agent balancing margin, conversion and churn risk. ### Hierarchical agents Hierarchical agents split responsibility into layers: a top-level planner sets goals and strategy, mid-level agents apply rules and decision paths, and low-level agents take direct actions. ![Hierarchy diagram: a top-level planner sets goals, mid-level agents apply rules and decision paths, and low-level agents take direct actions](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/6-6-hierarchical-agents-layered-task-management-fo.webp "Hierarchical agents split work between a planner, mid-level agents and executors.") ### Multi-agent systems Multi-agent systems combine several specialized agents that share data and coordinate toward a common objective, usually through an orchestration layer that assigns tasks and combines results. They are the most complex to build and often the most accurate and resilient; our guide to [multi-agent systems](/post/multi-agent-ai-systems-definition-benefits-limitations-how-to-build) goes deeper. ![Multi-agent diagram: a user talks to one AI agent that works with two other agents, each with its own memory and tools](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/7-7-collaborative-agents-multi-agent-systems-worki.webp "Collaborating agents coordinate toward a shared goal, each with its own memory and tools.") ## What makes an AI agent enterprise-ready? ![Diagram of five traits of enterprise-ready AI agents: operational fit, observability, controlled autonomy, system integration, and security and compliance](/blog/enterprise-ai-agents-benefits-use-cases-implementation-guide/8-8-what-makes-ai-agents-enterprise-ready-1.webp "Five traits that make AI agents ready for enterprise use.") Enterprises need stricter controls than a startup experimenting with a chatbot. An agent is ready for enterprise work when it has: - **Operational fit.** It handles a defined, repetitive task with measured accuracy. - **System integration.** It reads and writes in the systems where the work happens, through APIs and connectors. - **Security and compliance.** It follows your permission model and runs on a platform that meets your security and privacy obligations, such as GDPR and HIPAA. - **Controlled autonomy.** It stays within defined boundaries and cannot call tools outside its authority. - **Observability.** Every decision can be traced, explained and reviewed. ## What can enterprise AI agents do? - **Break down goals** from events or plain-language requests into executable steps. - **Retrieve and combine data** from several sources, pulling only what the task needs. - **Use tools,** such as APIs, databases, business applications and other agents, and choose among them based on the task. - **Remember context,** such as a customer's earlier issues or the steps already taken on a case. - **Run multi-step workflows** across systems, retrying on failure and following conditional paths. - **Monitor and flag,** watching operational data for anomalies such as a missed delivery or a budget overrun. - **Improve through feedback,** as reviewers' corrections feed evaluations and updates. ## How should people and agents work together? Enterprise agents rarely run unsupervised. Two interaction patterns cover most deployments: 1. **Autonomous agents** execute well-defined, repeatable workflows with minimal input. People supervise results, review traces and handle the cases the agent escalates. Examples include invoice validation, anomaly flagging and routine support requests. 2. **Assistive agents (copilots)** handle complex work but present results for a person's approval before any irreversible action. Examples include credit memos, legal review and strategic analysis. Many workflows combine both: the agent works autonomously on routine steps and pauses for approval on the ones with legal, financial or reputational impact. ## What powers enterprise AI agents? - **Large language models** interpret instructions, plan, reason and generate text. - **Tools and integrations** connect agents to business applications, databases and APIs, including through the [Model Context Protocol](/glossary/model-context-protocol). - **Knowledge and retrieval** ground agents in your documents and policies. - **Memory** keeps context within a task and across sessions. - **Orchestration** coordinates multiple agents and steps. - **Guardrails and approvals** detect risky inputs and pause high-impact actions for people. - **Observability and evaluation** record every run and measure quality. - **Infrastructure** in the cloud, on-premises or hybrid, close to the data the agents use. ## Why invest in AI agents for enterprise workflows? - **Operational efficiency.** Agents take on high-volume, repetitive work in support, back-office operations and approvals, around the clock. - **Lower cost at scale.** Volume grows without a matching increase in headcount for routine work. - **Consistent service.** Every customer and every case gets the same checks and the same quality of answer. - **Faster decisions.** Agents monitor live data and bring the relevant evidence into decision workflows. - **Resilience.** Teams absorb spikes in demand without emergency hiring. ## Where are enterprises using AI agents? - **Customer service.** Agents resolve routine chat and inbound call requests using account data and policies, and hand the rest to people, passing chats on with full context and transferring calls; see [customer service agents](/use-cases/customer-service). - **Financial services operations.** Agents support KYC checks, transaction disputes, fraud alert triage and insurance claims; see [financial services](/industries/financial-services). - **Back-office work.** Agents reconcile records, prepare reports and move approvals across finance systems; see [back-office agents](/use-cases/back-office). - **Document review.** Agents extract and check fields in loan files, claims and contracts; see [document review agents](/use-cases/document-review). - **IT operations.** Agents watch telemetry, open and enrich incidents, and assist engineers with configuration changes and testing. - **Healthcare administration.** Agents handle intake, referral documents and scheduling, with clinicians reviewing anything flagged; see [healthcare](/industries/healthcare). ## What are the pitfalls to avoid? ### No value-first strategy Launching agents without a defined business outcome produces expensive pilots with little impact. Pick the KPI first, such as ticket resolution time, then work backward to the agent's outputs, data needs, approval rules and technical requirements. ### Integration gaps Legacy systems without consistent APIs keep agents from the data they need. Invest in connectors, middleware or MCP servers, and start with processes whose systems are already reachable. ### Poor data Agents grounded in outdated or inconsistent documents give wrong answers with confidence. Assign owners to knowledge sources, keep them current and require agents to cite them. ### Opaque decisions When an agent's reasoning can't be reviewed, neither compliance teams nor users will trust it. Choose platforms that trace every step and make agents state their reasons. ### Security and privacy Agents that control enterprise tools and handle personal data widen the attack surface, including through prompt injection hidden in documents or web pages. Enforce least-privilege access for every agent, detect PII and prompt injection, keep sensitive data in your own perimeter and prepare an incident playbook for agent failures. ## What are the best practices for implementing AI agents? - Map processes or customer journeys to find where agents can make a measurable difference. - Automate clearly defined tasks first, and scale only after agents perform consistently. - Set strict boundaries on each agent's actions and data access. - Measure business KPIs, not vanity metrics. - Build several narrow agents rather than one general agent. - Set up version control, access logs and monitoring from the start. - Give users a way to approve, reject or flag agent actions, and turn that feedback into tests. - Design agents as modules that can be reused in other processes and multi-agent systems. - Train employees to write good requests, review agent output and spot red flags. ## How do you build and deploy enterprise AI agents with Dynamiq? Dynamiq offers three ways to put agents to work, on one governed platform: - **[AI Coworker](/product/chat)** is a ready-to-use agent with its own cloud computer, connectors, skills and scheduled tasks, on the web and in Slack, Microsoft Teams and Telegram. - **[Agent Builder](/product/agents)** lets teams design custom agents and workflows on a visual canvas or in the open-source Python SDK, with approvals, versions and rollback. - **[Voice Agents](/product/voice-agents)** answer inbound calls, use tools and hand off to people. Around them, the platform provides 2,100+ app integrations, knowledge bases with permission-aware retrieval, guardrails that detect PII and prompt injection, evaluations including real-time evals on live deployments, and traces of every run. It runs in Dynamiq's cloud or in your own infrastructure, with SOC 2, HIPAA and GDPR in place, and our engineers can take the first agents live with you; see [how we deploy](/deployment). To compare platforms on your own criteria, use the [agent platform evaluation kit](/resources/agent-platform-evaluation-kit). ## FAQ ### What is an enterprise AI agent? An enterprise AI agent is a software system that pursues a business goal with some autonomy, using a language model to reason and your tools and data to act, within the permissions, approvals and monitoring an enterprise requires. ### How are AI agents different from chatbots and RPA? Chatbots answer questions in a conversation, and RPA follows fixed scripts. AI agents complete multi-step tasks across systems, decide which tools to use, adapt to varied inputs and escalate to people when needed. ### What are the main types of AI agents? The classic types are reflex, goal-based, utility-based, learning, hierarchical and multi-agent systems. Most enterprise deployments today are goal-based agents built on language models, often coordinated in multi-agent or hierarchical designs. ### How do you make AI agents secure and compliant? Give each agent least-privilege access, detect PII and prompt injection, keep sensitive data in infrastructure you control, require approvals for high-impact actions, and trace every run so auditors can see what happened. ### How long does it take to deploy an enterprise AI agent? It depends on the process, data access and approvals, more than on the agent itself. A focused first agent can move quickly: [an Asian neo-bank automated about 85% of support inquiries and went live in 30 days](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai). --- # Multi-agent AI systems: definition, benefits, limitations and how to build one URL: https://www.getdynamiq.ai/post/multi-agent-ai-systems-definition-benefits-limitations-how-to-build Published: 2025-06-30 What a multi-agent AI system is, the main coordination patterns, benefits and limits versus a single agent, regulated-industry examples, and how to build one. A multi-agent AI system is a group of AI agents, each with its own role, tools and instructions, coordinated to complete a task that is too broad or too long for a single agent. A manager agent or an explicit workflow graph decides which agent works on what, passes results between them and assembles the answer, with people approving the steps that matter. Multi-agent systems add specialization and parallelism, but they also multiply cost and coordination problems, so the best practice is to start with one agent and split only when you have a clear reason. This guide explains how multi-agent systems work, the main coordination patterns, their benefits and limits, examples from regulated industries, and how to build one. ## What is a multi-agent AI system? An AI agent uses a language model to plan, call tools and act until a goal is met. A multi-agent system connects several such agents. Each agent has a narrow job, for example retrieving registry data, screening adverse media or drafting a summary, and a coordination layer turns their work into one result. The idea comes from decades of research on distributed agents, but today's multi-agent systems are built from LLM agents. What makes them useful in enterprises is not agents chatting freely; it is controlled delegation, where each agent sees only what it needs and every hand-off is recorded. ## How do multi-agent systems work? A typical run looks like this: 1. **Decompose.** A manager agent, or a predefined graph, breaks the request into subtasks. 2. **Assign.** Each subtask goes to the agent with the right tools, data access and instructions. 3. **Execute.** Agents work, in sequence or in parallel, each in its own reasoning loop. 4. **Hand off.** Results pass to the next agent or back to the manager through a shared context. 5. **Check.** Validators, evaluators or people review intermediate and final results. 6. **Assemble.** The manager combines the results into the final answer or action. 7. **Recover.** Failed steps are retried, rerouted or escalated to a person. ## What are the main coordination patterns? | Pattern | How it works | Use it when | | ----------------------- | ------------------------------------------------------------------------------ | -------------------------------------------------------------------------- | | Manager with sub-agents | A manager agent calls specialist agents like tools and assembles their results | Subtasks vary by request and a model should decide the order | | Graph orchestration | You define states, edges and conditions; agents run inside states | The process can be drawn as boxes and arrows, with loops and exit criteria | | Sequential pipeline | Agents run in a fixed order, each consuming the last one's output | The steps never change, such as extract, then check, then summarize | | Parallel fan-out | The same agent or step runs once per item, then results are merged | Many similar items, such as screening each counterparty in a list | | Peer-to-peer network | Agents message each other without a central coordinator | Research settings; rarely the right choice for regulated production | Most production systems combine patterns: a graph for the overall process, a manager for one judgment-heavy step and a fan-out for bulk work. Deterministic routing, where rules rather than a model decide the next step, should handle every decision that has objective criteria. See our [glossary entry on agent orchestration](/glossary/agent-orchestration) for the core terms. ## When should you use multiple agents instead of one? Start with a single agent with good tools. It is simpler to build, test and debug, and a strong model can plan surprisingly long tasks on its own. Split into multiple agents when you hit one of these boundaries: - **Different data access.** One step needs customer data and another must never see it. - **Different tools and prompts.** A single agent's instructions have become long and contradictory. - **Different models.** A cheap model can handle routing or extraction while a stronger one reasons. - **Parallel work.** Independent subtasks can run at the same time to cut total latency. - **Independent review.** A separate agent, or a person, should check the work of another. | | Single agent | Multi-agent system | Rule-based automation | | ---------------------- | -------------------------- | ------------------------------------------------- | ------------------------- | | Best for | One coherent task | Broad work with distinct roles or data boundaries | Stable, repetitive steps | | Build and debug effort | Lowest | Highest | Low, until rules multiply | | Cost per task | One reasoning loop | Several loops plus coordination | Lowest | | Handles variation | Well | Well | Poorly | | Permissions | One set for the whole task | Scoped per agent | Fixed per step | ## What are the benefits of multi-agent systems? - **Specialization.** Each agent gets focused instructions, tools and examples, which makes its behavior easier to test and improve. - **Least privilege.** Agents only get the tools and data their role needs, which limits the damage from a mistake or a manipulated input. - **Parallelism.** Independent subtasks run at the same time, which shortens long jobs such as research across many sources. - **Cost control.** Simple roles run on small, cheap models; only the hardest step uses the most capable one. - **Modularity.** You can upgrade, test or roll back one agent without rebuilding the rest, and reuse proven agents in new workflows. - **Built-in review.** A reviewer agent or an approval step between agents catches errors before they reach a customer or a system of record. ## What are the limitations and risks? - **Cost and latency add up.** Every agent runs its own loop, and a manager adds calls on every hand-off. - **Errors propagate.** A wrong result from one agent becomes another agent's input. Validate hand-offs, not just final answers. - **Coordination failures.** Agents can repeat work, loop or stop early. Loop caps, explicit exit criteria and deterministic edges prevent most of it. - **Harder debugging.** Without a trace across all agents, finding which step went wrong is slow. - **Agent-to-agent prompt injection.** Text produced by one agent, or read from a document, can steer another. Treat agent outputs as untrusted input. - **Interoperability.** Agents built on different stacks need shared formats. The [Model Context Protocol](/glossary/model-context-protocol) standardizes access to tools, and the Agent2Agent protocol targets communication between agents. ## What do multi-agent systems look like in regulated industries? | Use case | Agents | Human checkpoint | | ---------------------------- | ----------------------------------------------------- | ------------------------------------ | | Periodic KYC review | Registry analyst, adverse media screener, case writer | An analyst approves the risk rating | | Loan file review | Document extractor, policy checker, exceptions writer | An underwriter decides on exceptions | | Customer service | Router, account agent, policy agent | Hand-off to a person for disputes | | Regulatory change monitoring | Source monitor, impact analyst, summary writer | Compliance signs off on actions | | Research report | Researcher, data analyst, reviewer | An editor approves publication | In each case, the agents prepare and the person decides. That split is what makes multi-agent systems acceptable in banking, insurance and healthcare, and it matches how [compliance teams](/use-cases/compliance-financial-crime) already work. ## How do you build a multi-agent system? 1. **Map the process.** Write the steps, the data each step needs and where a person must decide. 2. **Build and test one agent first.** Add agents only at the boundaries listed above. 3. **Choose the pattern.** Use a graph where the process is known, a manager where the order varies. 4. **Scope each agent.** Give it a clear description, its own instructions, and only the tools and data its role needs. 5. **Define hand-offs.** Use structured outputs between agents and validate them. 6. **Add checkpoints.** Put approvals before consequential actions and a reviewer where quality matters. 7. **Trace and evaluate.** Record every agent's steps, test the whole system on real cases and monitor live runs. Here is a manager-led example with the open-source Dynamiq SDK. A manager coordinates a registry analyst and a case writer for a KYC review; each specialist is an `Agent` passed to the manager as a tool: ```python from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.flows import Flow from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.tools.function_tool import function_tool from dynamiq.nodes.types import Behavior, InferenceMode llm = OpenAI(connection=OpenAIConnection(), model="gpt-4.1", temperature=0.1) @function_tool def get_company_record(company_id: str, **kwargs) -> dict: """Return the registry record for a company: legal name, directors and owners.""" return { # a stand-in for your registry data provider "legal_name": "Harbor Trading Ltd", "directors": ["A. Novak"], "owners": [{"name": "A. Novak", "share": 0.8}, {"name": "Delta Holdings", "share": 0.2}], } registry_analyst = Agent( name="Registry Analyst", description="Looks up company ownership and directors. Call with {'input': ''}.", role="Retrieve registry data and summarize ownership and control. Cite the record.", llm=llm, tools=[get_company_record()], inference_mode=InferenceMode.XML, max_loops=5, behaviour_on_max_loops=Behavior.RETURN, ) case_writer = Agent( name="Case Writer", description="Writes the KYC review summary from findings. Call with {'input': ''}.", role="Write a concise KYC review summary: findings, a suggested risk rating and open questions.", llm=llm, inference_mode=InferenceMode.XML, max_loops=3, behaviour_on_max_loops=Behavior.RETURN, ) manager = Agent( name="KYC Review Manager", role=( "Coordinate a periodic KYC review. Delegate registry research first, then the summary.\n" "Always call tools with {'input': ''} payloads. Never make the final risk decision." ), llm=llm, tools=[registry_analyst, case_writer], inference_mode=InferenceMode.XML, max_loops=8, behaviour_on_max_loops=Behavior.RETURN, ) workflow = Workflow(flow=Flow(nodes=[manager])) result = workflow.run(input_data={"input": "Prepare the periodic KYC review for company C-1042."}) print(result.output[manager.id]["output"]["content"]) ``` The `description` on each specialist tells the manager when to call it. When the process is fixed, the SDK's `GraphOrchestrator` lets you define states, edges and conditional edges explicitly instead of leaving the order to the manager; the [orchestrators documentation](https://docs.getdynamiq.ai/docs/sdk/agents/orchestrators) covers both patterns. ## How does Dynamiq support multi-agent systems? In [Agent Builder](/product/agents), the same patterns are nodes on a visual canvas that run on the same engine as the SDK. An Agent node can call other agents as sub-agents. The Graph Agent Orchestrator runs agents inside explicit states with conditional edges and a manager model for judgment-based routing. Choice nodes route deterministically on rules, and Map nodes fan a step out over a list of items. Approvals can pause any tool step, and an orchestrator's task plan can require a person's approval before any task runs. Every agent's steps appear in one trace with cost and latency, evaluations score the whole system, and workflows are versioned with rollback. For ad hoc research, [AI Coworker](/product/chat) spawns subagents in parallel for fan-out work and folds their results back in. ## FAQ ### What is a multi-agent system in AI? It is a system of several AI agents, each with its own role, tools and instructions, coordinated to complete a task together. Coordination comes from a manager agent, an explicit workflow graph or a combination of both. ### What is the difference between a single-agent and a multi-agent system? A single agent handles the whole task in one reasoning loop with one set of tools and permissions. A multi-agent system splits the task among specialized agents, which allows narrower permissions and parallel work but adds cost and coordination complexity. ### What is agent orchestration? Agent orchestration is the logic that decides which agent works on which subtask, in what order, with what inputs, and what happens when a step fails. It can be a manager agent, a graph of states and edges, deterministic rules or a mix. ### Are multi-agent systems more accurate than single agents? Not automatically. Specialized agents with focused instructions and an independent reviewer often improve quality on broad tasks, but extra hand-offs also create new places for errors. Measure both designs on the same test cases before choosing. ### How do agents from different vendors work together? Through shared protocols. The Model Context Protocol standardizes how agents reach tools and data, and the Agent2Agent protocol defines how agents built on different platforms exchange tasks and results. Within one platform, structured hand-offs between agents do the same job. --- # Agentic workflows explained: benefits, use cases, best practices URL: https://www.getdynamiq.ai/post/agentic-workflows-explained-benefits-use-cases-best-practices Published: 2025-06-26 What agentic workflows are, how they differ from automation and autonomous agents, where they pay off in regulated work, and how to build them safely. An agentic workflow is a business process in which AI agents decide how to complete the steps, calling tools, reading data and checking their own output, within limits you set. Unlike a fixed automation script, it adapts when the inputs vary; unlike a fully autonomous agent, it runs inside a defined process with approvals and records. That combination makes agentic workflows the practical way to apply AI agents to regulated work, from banking back-office operations to customer support. ## What is an agentic workflow? An [agentic workflow](/glossary/agentic-workflow) combines a defined process with AI agents that have some freedom inside it. The process sets the goal, the available tools, the rules and the points where a person must approve. The agents decide how to get each step done. ![Diagram of four agentic AI capabilities around a central circle: make decisions, plan execution, retain memory and invoke other apps](/blog/agentic-workflows-explained-benefits-use-cases-best-practices/3-3-core-capabilities-for-autonomous-execution-and.webp "Four capabilities that let an agentic workflow act on its own: decide, plan, remember and use other applications.") It helps to place agentic workflows between the two approaches they are often confused with: | | Rule-based automation (RPA) | Agentic workflow | Autonomous agent | | -------------------- | ---------------------------------- | ----------------------------------------------- | ------------------------------- | | Who decides the path | A fixed script | Agents, within a defined process | The agent, with few constraints | | Handles variation | Poorly: breaks on unexpected input | Well: reasons about new cases | Well, but less predictably | | Controls | Built into the script | Guardrails, approvals, limits and traces | Mostly the agent's own judgment | | Best for | Stable, repetitive steps | Varied, multi-system work that must be governed | Open-ended exploration | ## Why are companies adopting agentic workflows now? Models have become reliable enough at tool use and multi-step reasoning to handle real process work, and analysts expect adoption to grow quickly. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously. Customer service is expected to move fastest. In a 2025 survey of 7,950 business and technology decision-makers, [Cisco found](https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2025/m05/agentic-ai-poised-to-handle-68-of-customer-service-and-support-interactions-by-2028.html) that 68% of customer service and support interactions with technology vendors are expected to be handled by agentic AI by 2028. ## How does an agentic workflow work? Most agentic workflows move through the same stages. Take a card dispute at a bank: 1. **Trigger.** A customer files a dispute through the app. 2. **Interpret.** An agent reads the customer's message and the transaction, and works out what the customer is claiming. 3. **Plan.** It checks the dispute policy and decides which steps apply, such as whether the case qualifies for a provisional credit. 4. **Act.** It pulls the transaction record, the merchant details and the customer's history through tools, and drafts the case file. 5. **Check.** It verifies its work against the policy rules and retries or corrects when something doesn't fit. 6. **Escalate or complete.** Low-value, clear-cut cases proceed; anything above a threshold goes to an analyst for approval, with the case file attached. Every step is recorded, so the bank can show exactly what the workflow saw and did. ## What are the key components of an agentic workflow? - **A reasoning model** that interprets inputs, plans and decides which tool to call. - **Tools and integrations** that let agents read and write in your systems, such as APIs, databases and [app integrations](/product/integrations). - **Knowledge** from your documents, retrieved when needed, so agents follow your current policies. - **Memory** of the conversation and of durable facts about a user; our guide to [agents with memory](/post/building-a-conversational-ai-agent-with-dynamiq) shows how it works. - **Orchestration** that coordinates several agents, either in a fixed order or under a manager agent; see [orchestration patterns](/post/agent-orchestration-patterns-in-multi-agent-systems-linear-and-adaptive-approaches-with-dynamiq). - **Guardrails and approvals** that detect risky input and pause high-impact steps for a person. - **Observability and evaluation** that record every run and measure quality over time. Rule-based automation still has a place. Stable, repetitive steps, such as moving a file between two systems, are cheaper and more predictable as plain automation, and an agentic workflow can call them as tools. ## Which patterns do agentic workflows use? - **Planning.** The agent breaks an open-ended goal into ordered subtasks and adds steps when it discovers new requirements, such as a permission check. - **Tool use.** The agent picks the tool that fits each subtask and switches to a fallback when one fails. - **Reflection.** The agent checks its output against the goal and the rules before moving on, and revises when it finds a gap. - **Routing.** The workflow sends each case down the path it needs, for example straight-through processing for simple cases and review for complex ones. - **Human in the loop.** People approve, reject or edit at defined points, and the workflow escalates when a person is needed; see [human-in-the-loop](/glossary/human-in-the-loop). ## Where do agentic workflows pay off? The best candidates are processes that span several systems, vary from case to case and happen often enough to matter: - **Back-office operations,** such as reconciliations, report preparation and approvals across finance systems; see [back-office agents](/use-cases/back-office). - **Compliance and financial crime,** such as triaging KYC and AML alerts and drafting the analyst's writeup; see [compliance agents](/use-cases/compliance-financial-crime). - **Customer support** across chat and inbound calls, resolving routine requests and handing the rest to people, passing chats on with full context and transferring calls; see [customer service agents](/use-cases/customer-service). - **Document review and underwriting,** extracting fields from loan files, claims and contracts and checking them against policy; see [document review agents](/use-cases/document-review). - **Research and knowledge work,** answering questions from internal documents with citations; see [research assistants](/use-cases/research-knowledge). ## What are the benefits of agentic workflows? - **Throughput.** Work that used to wait in a queue for a person moves as soon as it arrives, around the clock. - **Consistency.** Every case gets the same checks, in the same order, against the same policy. - **Faster decisions.** Analysts start from a prepared case file instead of gathering data themselves. - **Handles variation.** Agents cope with the unusual cases that break rule-based automation. - **Scales with demand.** Volume spikes don't require hiring and training temporary staff. ## What are the limitations, and how do you manage them? | Limitation | What goes wrong | How to manage it | | ---------------------- | ------------------------------------------------------------------ | --------------------------------------------------------------------------------- | | Unpredictable behavior | The same input can produce different paths or answers | Fix the outer process, keep approvals for high-impact steps, set loop limits | | Data quality | Agents reason from outdated or wrong documents | Keep knowledge sources current and owned, and cite sources in outputs | | Limited judgment | Agents struggle with ethical calls, ambiguity and rare situations | Keep such cases with people, and route them there automatically | | Explainability | Reviewers can't see why an agent reached a conclusion | Trace every run and require agents to state their reasons | | Cost and latency | Many model calls per task add up | Use smaller models for simple steps and plain automation where it suffices | | Security and privacy | Agents touch personal data and can be targeted by prompt injection | Least-privilege tools, PII and prompt-injection detection, data in your perimeter | | Skills gap | Few people have built production agents | Start with one process, use a visual builder, and bring in experienced engineers | ## What are the best practices for building agentic workflows? - **Define outcome goals first.** Pick measurable targets, such as cases resolved without handoff or time to decision, before you build. - **Keep agents narrow.** Several focused agents with clear inputs and outputs are easier to test than one general agent. - **Grant least privilege.** Give each agent only the tools and data it needs, and keep irreversible actions behind approvals. - **Keep humans on high-impact steps.** Define exactly which decisions need a person, and give reviewers the ability to edit before approving. - **Record everything.** Traces of every run are your audit trail, your debugging tool and your evaluation data. - **Test with real edge cases.** Build an evaluation set from real cases, including the messy ones, and turn every production failure into a new test. - **Roll out in phases.** Start with a contained slice of volume, compare results with the manual process, then expand. ## How do you build an agentic workflow step by step? 1. **Pick the process.** Look for high volume, several systems, measurable outcomes and clear rules about what needs a person. 2. **Map the steps.** Write down each step, decision point, system and exception, as the people doing the work describe them. 3. **Choose models and where data lives.** Match models to steps, and decide whether processing must stay in your own infrastructure. 4. **Connect tools and knowledge** with least privilege and credentials the model never sees. 5. **Add guardrails and approvals** at the points your risk and compliance teams define. 6. **Evaluate on real cases** before launch, with metrics your business owners agree on. 7. **Deploy in phases and monitor** quality, cost and escalation rates in production. Fine-tuning is rarely the first step. Most workflows reach their targets with good instructions, retrieval from current documents and the right tools; fine-tune only when evaluation shows a gap those can't close. ## How does Dynamiq help you build agentic workflows? Dynamiq is built for agentic workflows in regulated industries: - **Agent Builder** gives business teams a visual canvas and engineers an open-source Python SDK on the same engine, with Choice and rules nodes, prompt-to-workflow generation, and versions with rollback; see [Agent Builder](/product/agents). - **Approvals with editable fields** pause any tool step for a person, and guardrails detect PII and prompt injection; see [guardrails and approvals](/product/guardrails). - **2,100+ app integrations** and MCP connect agents to the systems your process already uses. - **Evaluations** score quality on test sets and in real time on live deployments; see [Evals](/product/evaluations). - **Observability** traces every run with cost and latency, and an AI agent surfaces issues from those traces; see [observability](/product/observability). The platform runs in Dynamiq's cloud or self-hosted in your own infrastructure, with SOC 2, HIPAA and GDPR in place, and our engineers can take the first workflows live with you; see [how we deploy](/deployment). ## FAQ ### What is the difference between an agentic workflow and an AI agent? An AI agent is a single system that reasons and uses tools to reach a goal. An agentic workflow is a business process that uses one or more agents inside defined steps, rules and approvals. The workflow gives the agents structure; the agents give the workflow flexibility. ### How are agentic workflows different from RPA? RPA follows fixed scripts and breaks when inputs change. Agentic workflows use AI agents that interpret varied inputs and decide how to complete each step, while approvals and traces keep them controlled. Many teams use both, with agents calling RPA bots for stable, repetitive steps. ### What are examples of agentic workflows? Examples include card dispute handling, KYC and AML alert triage, invoice reconciliation, loan file review, customer support that resolves requests across systems, and research assistants that answer questions from internal documents with citations. ### Are agentic workflows safe for regulated industries? They can be, when they are designed for it: least-privilege tools, detection of PII and prompt injection, approvals on high-impact steps, data processed in infrastructure you control, and a trace of every run that auditors can follow. ### How do you measure whether an agentic workflow is working? Track business outcomes, such as cases completed without handoff, time to decision and cost per case, alongside quality metrics from evaluations, such as accuracy against a reviewed test set, and operational signals such as escalation and error rates. --- # LLM agents explained: a complete guide URL: https://www.getdynamiq.ai/post/llm-agents-explained-complete-guide-in-2025 Published: 2025-06-19 What LLM agents are, how they work, their components and types, where enterprises use them, the main challenges, and the frameworks used to build them. An LLM agent is a system that uses a large language model to decide what to do next, calls tools such as APIs, databases or search to act, reads the results and repeats until it reaches a goal. A chatbot answers one message at a time; an agent plans multi-step work, uses your systems and data, and keeps track of context. In enterprises, LLM agents now resolve customer requests, prepare back-office cases, research questions across documents and review files, with guardrails and human approvals around the steps that carry risk. This guide covers how LLM agents work, their components and types, where they are used, the challenges, and the frameworks for building them. ## What is an LLM agent? An LLM agent combines three things: a language model that reasons in natural language, tools that let it act on the world, and a loop that feeds each result back to the model until the task is done. The model supplies judgment, such as understanding a request, choosing a tool or deciding an answer is good enough; the tools supply reach. The difference from a plain LLM is action. Ask a model about a customer's order and it can only guess. Give an agent an order lookup tool and it checks the order, notices a delayed shipment, looks up the carrier status and tells the customer what happened. ## How does an LLM agent work? Most LLM agents follow the same loop, often called reason, act, observe: 1. **Receive a goal.** A user message, an event or a scheduled trigger starts the run. 2. **Gather context.** The agent loads its instructions, relevant memory and, if needed, retrieved documents. 3. **Reason.** The model decides the next step: answer directly, call a tool or ask a clarifying question. 4. **Act.** The agent calls the chosen tool with structured arguments, for example an API request or a database query. 5. **Observe.** It reads the tool's result or error and adds it to its context. 6. **Repeat or finish.** It continues until the goal is met, a step limit is reached, or a person must decide. Modern models support tool calling natively: they return structured requests to call a function, which the runtime executes. Many tools are now exposed through the [Model Context Protocol](/glossary/model-context-protocol), an open standard for connecting agents to tools and data sources. ## What are the key components of an LLM agent? | Component | Role | | ------------ | ---------------------------------------------------------------------------- | | Model | Understands requests, plans and chooses actions | | Instructions | Define scope, rules, tone and when to hand off to a person | | Tools | APIs, databases, search, code execution and other agents the agent can call | | Knowledge | Documents retrieved at run time so answers are grounded and cited | | Memory | Session context, plus long-term facts about a user across sessions | | Planning | Breaking a large goal into steps, sometimes with a written plan or todo list | | Guardrails | Checks on inputs, outputs and actions, including human approval | | Tracing | A record of every step, with inputs, outputs, cost and latency | ## What types of LLM agents are there? | Type | What it does | Example | | ------------------------------ | -------------------------------------------------------------- | ----------------------------------------------------------- | | Task agent | Completes a defined task end to end | Reconciling two exports and listing mismatches | | Conversational agent | Works through a task in dialogue, asking questions as needed | A support agent that checks a card and blocks it on request | | Knowledge (RAG) agent | Answers from retrieved documents, with citations | A policy assistant for relationship managers | | Coding and data agent | Writes and runs code in a sandbox | Analyzing a CSV and returning a chart | | Browser and computer-use agent | Operates web pages or desktop apps like a person | Filling a supplier portal form | | Voice agent | Handles phone or web calls in real time | Answering inbound calls about appointments | | Multi-agent system | Several specialized agents coordinated by a manager or a graph | Research, drafting and review agents producing a report | | Embedded agent | Runs inside an existing product or workflow | An assistant inside a CRM or ticketing tool | Most production systems mix types. A support workflow might pair a conversational agent with a knowledge agent and a fixed approval step. Our guide to [multi-agent systems](/post/multi-agent-ai-systems-definition-benefits-limitations-how-to-build) covers coordination patterns. ## What are LLM agents used for? - **Customer service.** Resolving routine requests and handing complex cases to a person with the full context. An Asian neo-bank [automated about 85% of support inquiries](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) with a single agent that uses tools, memory and a knowledge base. - **Back-office operations.** Matching, reconciling and preparing exceptions for approval in finance and operations teams. - **Document review.** Extracting and checking fields in loan files, claims and contracts, and flagging what does not match policy. - **Compliance.** Gathering evidence and drafting KYC and AML case summaries for analysts to decide. - **Research and knowledge.** Answering questions across internal documents with citations and permission-aware retrieval. - **Software engineering and data analysis.** Writing code, tests and analyses that people review. ## What are the main challenges with LLM agents? - **Reliability.** Agents can misread a request, pick the wrong tool or stop too early. Narrow scopes, clear tool descriptions and step limits help. - **Security.** Prompt injection hidden in emails, documents or web pages can try to redirect an agent, and an agent with broad permissions can do real damage. Least-privilege tools, input detectors and approvals on consequential actions contain the risk. - **Evaluation.** Answers vary between runs, so quality must be measured on test sets and on samples of live traffic, not judged from a demo. - **Cost and latency.** Each loop is a model call. Routing simple steps to smaller models and capping loops keeps cost per task predictable. - **Memory.** Deciding what to remember, for how long and for whom is a design and privacy decision, not just a storage choice. - **Governance.** Regulated teams need to show what an agent read, did and who approved it, which requires traces and versioned configurations. ## Which frameworks are used to build LLM agents? Frameworks handle the loop, tool calling, memory and orchestration so teams write business logic instead. Widely used options include: | Framework | Maintained by | Known for | | ------------------------- | ------------- | ----------------------------------------------------------------------------------------------- | | LangGraph | LangChain | Stateful, graph-based agent workflows | | CrewAI | CrewAI | Role-based teams of agents | | Microsoft Agent Framework | Microsoft | The successor to AutoGen and Semantic Kernel | | OpenAI Agents SDK | OpenAI | Lightweight agents with handoffs and guardrails | | Agent Development Kit | Google | Agents built for the Google Cloud ecosystem | | LlamaIndex | LlamaIndex | Agents over documents and data | | Dynamiq SDK | Dynamiq | Agents, RAG and orchestration in Python (Apache-2.0), with a visual platform on the same engine | A framework gives you code. Production also needs deployment, access control, guardrails, evaluations and tracing, which is where a platform comes in. Our [comparisons](/compare) show how Dynamiq differs from common frameworks and platforms. ## How do you build an LLM agent? 1. Define one narrow job and how you will measure success. 2. Choose a model and keep it swappable. 3. Write specific instructions, including what the agent must never do. 4. Give it the tools and knowledge the job needs, read-only first. 5. Add memory scoped to the user and session. 6. Add guardrails and require approval before irreversible actions. 7. Test on real cases, then deploy with tracing and ongoing evaluation. Our [step-by-step guide to creating an AI agent](/post/how-to-create-an-ai-agent-from-scratch-step-by-step-guide) walks through each step with working Python code. ## How does Dynamiq help teams build LLM agents? [Agent Builder](/product/agents) lets business teams design agents on a visual canvas and engineers build them with the open-source Python SDK, both on one engine. Agents get tools from a large integrations catalog, knowledge bases with permission-aware retrieval, session memory scoped by user and session, and approvals with editable fields on any tool step. Guardrails detect PII and prompt injection, evaluations run on test sets and live deployments, and every run is traced with cost and latency. For work that does not need a custom workflow, AI Coworker is a ready-made agent with its own cloud sandbox, connectors, skills, scheduled tasks and subagents, available in the browser, Slack, Microsoft Teams and Telegram. The platform runs in Dynamiq Cloud or self-hosted in your own cloud or data center. ## FAQ ### What is the difference between an LLM and an LLM agent? An LLM generates text from a prompt. An LLM agent wraps the model in a loop with tools, memory and instructions, so it can take actions, check the results and keep working until a task is done. ### What is the difference between an LLM agent and a chatbot? A chatbot responds to each message, usually from a script or a model's knowledge. An LLM agent plans multi-step work, calls tools and APIs to get real data or take actions, and can hand off to a person when a decision needs one. ### Are LLM agents fully autonomous? They can be, but most enterprise agents are not. Teams set the level of autonomy per step: the agent drafts and looks things up freely, while actions such as payments, account changes or outbound messages wait for a person's approval. ### What is agent memory? Agent memory is the context an agent keeps. Session memory holds the current conversation; long-term memory stores durable facts about a user, such as preferences, and recalls them in later sessions. Both should be scoped to the right user and governed like any other personal data. ### What is the Model Context Protocol? The Model Context Protocol (MCP) is an open standard for connecting AI agents to tools and data sources. A system that exposes an MCP server can be used by any MCP-compatible agent without a custom integration for each one. --- # AI-driven due diligence through to exit: how private equity firms use GenAI URL: https://www.getdynamiq.ai/post/ai-driven-due-diligence-through-to-exit-how-private-equity-firms-are-using-genai-in-2025 Published: 2025-06-17 How private equity firms use GenAI from due diligence to exit: red flag scanning, target screening, investment memos, KPIs to track and pitfalls to avoid. Private equity firms use generative AI (GenAI) at every stage of the deal lifecycle: to scan data rooms for red flags during due diligence, screen targets, draft investment memos, monitor portfolio companies and prepare exit materials. The firms that trust the results make every output traceable to its source, keep people accountable for what enters an investment decision, and keep deal data inside infrastructure they control. This guide covers the use cases, the metrics that prove value, the common roadblocks and a roadmap for building AI maturity inside a fund. ## Where does GenAI fit in the private equity lifecycle? ### Due diligence and red flag scanning Deal windows are short, and diligence teams face more material than they can read: contracts, litigation records, regulatory disclosures, media coverage and social posts. GenAI reads all of it, extracts the clauses and facts that matter, and flags red flags for a person to check, such as change-of-control clauses, pending litigation or undisclosed related parties. Connected to sanctions lists and corporate relationship data, agents can also surface links between a target's founders or counterparties and sanctioned individuals or politically exposed persons, the same kind of screening banks run in [KYC and AML work](/glossary/kyc-and-aml-automation). ### Technical and cyber due diligence For technology targets, GenAI helps assess the quality of the codebase, architecture risks, dependence on third-party software, the security posture and data privacy compliance. It speeds up expert review rather than replacing it. ### Target discovery GenAI screening combines commercial databases such as PitchBook, Crunchbase and Apollo with company websites, hiring pages, job listings and social networks, which gives a broader view than traditional filters. ![Diagram of data sources feeding GenAI target screening: Apollo, PitchBook and Crunchbase, company websites, social media networks and hiring pages](/blog/ai-driven-due-diligence-through-to-exit-how-private-equity-firms-are-using-genai-in-2025/4-4-genai-s-comprehensive-approach-to-analysing-co.webp "Sources GenAI can combine when screening target companies.") Lessons from past deals belong in the screening criteria. If deals fell through over cap table complexity or management misalignment, make those explicit checks, so the agent screens for them on every target. ### ESG review GenAI supports environmental, social and governance review by reading emissions data, sustainability disclosures, regulatory feeds and news coverage, including sources in other languages. It surfaces issues a static ESG score misses, such as recent allegations against a target with a clean compliance record, for the deal team to investigate. ### Value creation during ownership After the deal closes, agents monitor portfolio KPIs such as EBITDA against plan, customer acquisition cost and lead quality. They alert the team to deviations, explain what changed and suggest remediation options based on the data, so operating partners act earlier. ### Investment memos Deal teams can generate a first draft of an investment memo from the data room, the confidential information memorandum and meeting transcripts, using the firm's own template. The draft includes a thesis outline, supporting arguments and key risks, each linked to its source. ![Flow diagram: a data room, a confidential information memorandum and meeting transcripts go into AI diligence software, which produces a thesis outline with supporting arguments and risks](/blog/ai-driven-due-diligence-through-to-exit-how-private-equity-firms-are-using-genai-in-2025/7-7-creating-investment-memos-with-genai-1.webp "Deal inputs turned into a first-draft investment thesis for the team to review.") Consistency is a second benefit: every memo includes the baseline metrics, red flag disclosures and return calculations the investment committee expects, and version history shows what changed between drafts. ### Exit planning GenAI helps turn years of operating data into a clear equity story: the company's competitive position, the improvements made during ownership and the growth ahead. Teams can tailor materials to each buyer's priorities and keep answers to buyer questions consistent with the data room. GenAI tools don't get better on their own as they are used. They improve when reviewers correct outputs and those corrections become new test cases, instructions and data, a deliberate process the firm controls. ## Which metrics show that AI-driven diligence works? Tie metrics to returns, speed and risk, not only to usage. | Category | Metrics | | ------------------------ | ---------------------------------------------------------------------------------------- | | Investment returns | Value leakage avoided through issues found before signing; high-risk deals avoided | | Deal cycle speed | Diligence time per deal; time to surface material concerns; analyst hours redeployed | | Precision and coverage | Share of flags confirmed as material; issues missed; risk areas newly covered | | Technical risk detection | Architecture, security and third-party dependency issues found versus traditional review | | Adoption and feedback | Use across deal stages; reviewer corrections turned into tests; error rates over time | ## What roadblocks do firms hit, and how do they solve them? ### Security and confidentiality Diligence data is among the most sensitive a firm holds, and sending it to third-party services raises confidentiality and regulatory questions under laws such as the GDPR. - Run the AI platform in infrastructure the firm controls, or with providers bound by strict processing terms. - Encrypt data at rest and in transit, and apply least-privilege access by deal team. - Detect personal data before it reaches a model, and pseudonymize where the analysis allows. - Review vendors' security attestations, penetration testing and sub-processors. ### Data readiness GenAI is only as good as the documents it reads. Poor scans, inconsistent formats and missing files produce weak answers. - Parse and index data room documents properly, including tables and scanned pages. - Ground answers in retrieved documents rather than the model's general knowledge. - Keep a person in the loop to validate summaries before they are used. - Test instructions against real edge cases before a live deal. ### Explainability and trust An investment committee cannot act on an answer it cannot check. - Require every claim to cite the document and page it came from. - Keep a trace of each run: what the agent read, what it concluded and why. - Validate all critical outputs internally before they enter a memo. ### Talent and adoption Tools stall when deal teams don't know how to use them or don't trust them. - Train teams on writing requests, reviewing outputs and spotting red flags. - Appoint champions in each team who share practical know-how. - Use visual tools that let analysts adjust workflows without coding. - Run structured quality checks before outputs reach production. ## What are operating partners doing differently? Many firms now have operating partners or dedicated leaders responsible for AI across the portfolio. They treat AI as part of the value creation plan: they choose where agents should be deployed first, set shared standards for data and governance, and make sure each deployment is measured against a business outcome. Our guide to [AI agents for private equity](/post/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy) covers the portfolio playbook in detail. ## What does a roadmap to AI maturity look like? ### 1. Prototype and experiment - Start with targeted use cases such as contract extraction, red flag scanning or memo drafting. - Define success metrics before you start. - Test on real, past deal materials in a secure environment rather than demo data. - Document results, problems and lessons. ### 2. Evaluate and choose partners - Test platforms on your own use cases: full-length contracts, multilingual media checks, entity resolution. - Compare outputs with experienced analysts on several past deals. - Check security attestations such as SOC 2 and ISO 27001, and how the vendor supports your GDPR, HIPAA or CCPA obligations. - Simulate a real deal cycle and measure speed and reliability. ### 3. Embed into deal processes - Integrate with deal flow software, CRMs and virtual data rooms. - Build a shared library of instructions, output templates and red flag typologies. - Assign governance roles, and define which outputs always need human validation. - Review quality and compliance on a regular schedule. ### 4. Scale across the fund - Align AI tools with the portfolio strategy, risk appetite and compliance requirements. - Make sure models and workflows can be adapted to firm-specific data. - Plan for vendor changes: your workflows, data and outputs should be able to move if a relationship ends. An [open-source SDK](/open-source) under a permissive license is one way to keep that option. ## How does Dynamiq support AI-driven due diligence? - **Data rooms as knowledge.** Document parsing and extraction, knowledge graphs for relationships between entities, and permission-aware retrieval so each deal team sees only its deals; see [Knowledge](/product/knowledge-rag). - **Agents for each stage,** built on a visual canvas or in Python on the same engine; see [Agent Builder](/product/agents). - **Controls** that detect PII and prompt injection and pause any tool step for a person's approval; see [guardrails and approvals](/product/guardrails). - **Evidence** from traces of every run and evaluations of output quality; see [observability](/product/observability). The platform runs in Dynamiq's cloud or inside the firm's own infrastructure, and our engineers can deploy the first workflows with your team; see [how we deploy](/deployment). ## FAQ ### How is AI used in private equity due diligence? Firms use AI to read and summarize data room documents, extract key clauses and figures, flag red flags such as litigation or change-of-control terms, screen entities against sanctions and PEP lists, assess technology and cyber risk, and draft question lists and investment memos for people to review. ### Can generative AI replace due diligence advisers? No. GenAI speeds up the reading, extraction and first-pass analysis, so advisers and deal teams spend their time on judgment, negotiation and the issues that matter. Accountability for the findings stays with people. ### How do PE firms keep deal data secure when using GenAI? They run AI platforms in infrastructure they control or under strict processing agreements, restrict access by deal team, detect personal data before it reaches a model, encrypt data, and keep a trace of every run. ### What metrics should PE firms track for AI-driven diligence? Track diligence time per deal, time to surface material issues, the share of AI flags confirmed as material, issues missed, value leakage avoided and adoption across deal teams, each measured against a baseline from before the rollout. ### How do you avoid vendor lock-in with AI tools? Keep your data, instructions and evaluation sets in formats you own, prefer platforms with open-source components and self-hosting options, and confirm before signing that workflows and outputs can be exported if the relationship ends. --- # AI agents for private equity: use cases, KPIs, and deployment strategy URL: https://www.getdynamiq.ai/post/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy Published: 2025-06-12 How private equity firms use AI agents in deal sourcing, diligence and portfolio operations, which KPIs prove ROI, and how to run a first pilot. AI agents help private equity firms work through the volume of documents, data and signals that deal teams and operating partners handle every week. The most common uses are deal screening, due diligence document review, portfolio KPI reporting, FP&A support and customer service at portfolio companies. Firms that get measurable returns start with one pilot tied to a clear KPI, keep people on investment decisions and build a central playbook before they scale across holdings. ## What is an AI agent in private equity? An AI agent is a system that works toward a goal with some autonomy: it interprets a task, plans the steps, uses tools such as data rooms, CRMs and financial systems, and checks its results. Generative AI lets agents read unstructured material, such as a confidential information memorandum or an earnings call transcript, and write in plain language. ![Diagram of an agent that uses goals, abilities, prior knowledge, observations and past experiences to choose actions in its environment](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/2-2-what-is-an-ai-agent-2.webp "A general model of an agent, adapted from Poole and Mackworth: goals, abilities and knowledge shape the actions it takes.") In a PE firm, that means an agent can complete first-pass diligence checks, prepare a financial model from source files, draft investor reports or alert a deal team when a portfolio company's metrics move. A person stays accountable for every investment decision. ## How are AI agents different from traditional automation tools? | | Traditional automation and BI | AI agents | | ------------------ | ------------------------------------ | ------------------------------------------------------------------------------------------- | | Inputs | Structured data in known formats | Structured and unstructured data: contracts, decks, emails, transcripts | | Handling variation | Breaks when formats or labels change | Interprets new formats and inconsistent labels | | Output | Fixed reports and dashboards | Summaries, drafts, flags and answers to follow-up questions | | Improvement | Rebuilt by developers | Improved through evaluation: reviewer corrections become test cases and instruction changes | That last row deserves care. Agents don't learn on their own from deal outcomes. They get better when your team reviews their output, turns mistakes into test cases and updates instructions, tools and data, which is a process you control and can audit. ## Where do AI agents fit in a PE firm and its portfolio? ### Deal sourcing and screening Analysts scan startup profiles, news alerts and market signals every day, and even diligent teams miss opportunities that don't catch their eye. A screening agent reviews targets against your investment criteria, adds signals such as hiring activity, patent filings and funding rounds, explains why a target fits or not, and re-ranks the pipeline as new information arrives. The analyst decides what to pursue. ### Due diligence and document review Reviewing confidential information memorandums, financial audits, HR materials and disclosures is slow and cognitively demanding. Agents find information in embedded spreadsheets, scanned documents and unstructured PDFs, flag inconsistencies between documents and draft question lists for advisers. Our companion guide covers [AI-driven due diligence through to exit](/post/ai-driven-due-diligence-through-to-exit-how-private-equity-firms-are-using-genai-in-2025) in depth. ![Diagram of confidential information memorandums, financial audits, HR materials and disclosures flowing into a private equity AI agent that reviews and synthesizes them](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/9-9-document-processing-and-review-document-1.webp "An agent reviews and synthesizes the core documents of a deal.") ### FP&A copilots Financial planning and analysis teams spend much of their time cleaning data and reconciling numbers. An FP&A copilot connects to accounting systems, ERP (enterprise resource planning) platforms and CRMs, produces rolling forecasts, runs scenarios such as revenue underperformance or currency swings, and flags anomalies such as margin compression, with commentary a finance lead can review. ### Portfolio KPI reporting Portfolio companies report the same metrics under different names and in different formats. An agent maps each company's column names and definitions to your standard KPIs, flags gaps instead of guessing them, and prepares consistent reports for each partner. ### RFP copilots Portfolio companies that sell through tenders can use an agent to draft responses from their requirements, policies, technical specifications, pricing logic and past submissions, and to route anything unusual to the sales or deal team. ![Diagram of an RFP copilot that matches a request for proposal to company policies, specifications, pricing logic and past submissions, then drafts a response or escalates it](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/8-8-rfp-copilot-1.webp "An RFP copilot applies company requirements to draft responses and routes exceptions to people.") ### Customer support at portfolio companies Portfolio companies with large support teams can automate routine inquiries with agents grounded in product documentation and account data, handing complex cases to people with the full context. See [customer service agents](/use-cases/customer-service) for how that works. ## How are PE firms deploying AI agents across their holdings? Firms that scale agents across a portfolio tend to follow three moves: - **Portfolio-wide opportunity scans.** Assess each company for processes where agents could cut cost or improve decisions, and weigh the expected EBITDA effect against readiness and risk. - **Coverage of the deal lifecycle.** Apply agents from sourcing through diligence and ownership to exit, so insights from one phase inform the next. - **Central reuse.** Build agents once, keep them in a central registry with version control and access logs, and adapt them for each portfolio company. ![Ring diagram of five stages where AI agents help: deal sourcing, pipeline monitoring, due diligence, meeting optimization and exit strategy planning](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/10-10-streamlining-operations-across-portfolios-2.webp "Agents can support every stage from deal sourcing to exit.") Across the lifecycle, typical jobs include ranking deals and generating summary sheets, monitoring the pipeline for duplicates and inconsistencies, flagging missing clauses in diligence, preparing meeting briefs, and archiving what worked for future exits. ## What goes into a central AI playbook? A central playbook turns one successful pilot into a capability the whole portfolio can use. It covers five areas. ### Infrastructure A modular architecture keeps agents portable across companies with different systems. A common pattern has three layers: an input layer that ingests and standardizes structured and unstructured data, a reasoning layer that interprets the business context, and an action layer that writes results back to the systems people use. ![Three-tier agent architecture: input handling, cognitive and action layers connected in a cycle](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/11-11-building-a-central-ai-playbook-infrastructure.webp "A modular agent stack: input handling, reasoning and action layers.") The platform should run in environments the firm controls, so deal data stays inside the firm's perimeter; see [how we deploy](/deployment). ### Reusability Design agents as modular assets: templates for common jobs such as FP&A reporting, RFP drafting or churn monitoring, a library of tested instructions, and a defined way to validate each new adaptation. ### Governance Agent adoption needs three layers of governance: strategic (approved use cases and data boundaries), operational (a named owner for each agent and regular output reviews) and technical (traces, version control and test coverage). ### Controls - **Role-based access.** Agents see only the data their task requires. - **A complete record.** Every run is traced with its inputs, outputs and decisions, and kept for review. - **Policy checks.** Outputs that contradict policies, financial targets or controls are flagged for compliance and risk teams. Our guide to [LLM guardrails in private equity](/post/llm-guardrails-in-private-equity-how-to-scale-genai-safely-across-portfolio-companies) covers this in depth. ### Talent Relying entirely on outside vendors weakens institutional learning. Build a small center of excellence with training for deal teams and operators, a review board for new agents, and change management plans for the teams whose work changes. ## Which KPIs prove ROI from AI agents in private equity? Generic metrics such as time saved matter only when they connect to cost structures or value creation: | KPI | What it shows | Typical agent | | ------------------------ | ------------------------------------------------------------------------ | ------------------------- | | Analyst hours redeployed | Time moved from gathering data to higher-value analysis | Screening, diligence | | Adoption rate | How often teams actually use the agent in daily work | Any | | Speed to insight | Time from a new signal to a flagged risk or opportunity | Monitoring, KPI reporting | | Forecast accuracy | Gap between projected and actual financial performance | FP&A copilot | | Proposal win rate | Share of tenders won, reflected in recurring revenue | RFP copilot | | EBITDA effect | Cost savings and revenue gains attributable to agent-supported processes | Portfolio operations | Tie each pilot to one or two of these before it starts, and measure the baseline first. ## How do you start with a portfolio company pilot? ![Six numbered steps for a portfolio company pilot, from identifying a business problem to assigning ownership roles](/blog/ai-agents-for-private-equity-use-cases-kpis-and-deployment-strategy/12-12-steps-to-get-started-with-a-portfolio-company.webp "Six steps to launch a first AI agent pilot at a portfolio company.") 1. **Identify a business problem** or inefficiency that an agent can address. 2. **Focus on one clear objective** instead of spreading attention across many problems. 3. **Define success metrics** so the pilot can show a measurable return. 4. **Choose a portfolio company** with technical readiness and stakeholders willing to change how they work. 5. **Develop and test in iterations,** building a minimum viable agent with defined boundaries. 6. **Assign ownership** for data access, performance tracking and user feedback. Document each stage. During the pilot, work on accuracy, friction, false positives and false negatives; once the agent consistently hits its KPIs, add users, workflows and data sources. ## How does Dynamiq support private equity firms? Dynamiq gives firms one platform to build, govern and reuse agents across the portfolio: - **Build agents** on a visual canvas or in Python, on one engine, and reuse them as templates; see [Agent Builder](/product/agents). - **Ground them in data rooms and internal documents** with knowledge bases that respect permissions, so each person sees only what they are allowed to see; see [Knowledge](/product/knowledge-rag). - **Govern them** with PII and prompt injection detection and approvals on high-impact steps; see [guardrails and approvals](/product/guardrails). - **Measure them** with evaluations and traces of every run, including cost and latency; see [observability](/product/observability). The platform runs in Dynamiq's cloud or inside infrastructure the firm controls, with SOC 2, HIPAA and GDPR in place. ## FAQ ### How are private equity firms using AI agents? Firms use AI agents to screen and rank deals, review data room documents during diligence, standardize KPI reporting across portfolio companies, support FP&A teams with forecasts and scenarios, and automate customer support and RFP responses at portfolio companies. ### What is a good first AI agent use case for a PE firm? Pick a repetitive, document-heavy task with a clear success metric, such as first-pass review of confidential information memorandums or KPI report consolidation. It delivers value quickly, and a person already checks the output. ### How do you measure ROI from AI agents in private equity? Measure a baseline first, then track outcomes tied to cost or value creation: analyst hours redeployed, speed to insight, forecast accuracy, proposal win rates and, ultimately, the EBITDA effect of agent-supported processes. ### Can AI agents make investment decisions? They shouldn't. Agents can gather evidence, run analyses and recommend, but investment decisions carry fiduciary responsibility and belong to people. Design workflows so agents prepare and people decide. ### How do you keep deal data confidential when using AI? Run the agent platform in infrastructure the firm controls, restrict each agent's access by role, detect sensitive data before it reaches a model, keep a trace of every run, and review what model providers may do with any data sent to them. --- # Dynamiq to launch enterprise-ready AI agents in IBM watsonx Orchestrate's agent catalog URL: https://www.getdynamiq.ai/post/dynamiq-to-launch-enterprise-ready-ai-agents-within-ibm-watsonx-orchestrates-agent-catalog Published: 2025-05-22 Dynamiq announced plans to bring a Medical Research Agent, a Legal Assistant Agent and an OCR Document Agent to IBM watsonx Orchestrate's agent catalog. Dynamiq announced plans in May 2025 to make three enterprise-ready AI agents available in the [Agent Catalog within IBM watsonx Orchestrate](https://www.ibm.com/products/watsonx-orchestrate/agent-catalog): a Medical Research Agent, a Legal Assistant Agent and an OCR Document Agent. The agents connect to watsonx Orchestrate through IBM Agent Connect, so IBM clients can use them next to agents and tools from IBM and its partner ecosystem, or adapt them to their own data in Dynamiq. ## What is the watsonx Orchestrate Agent Catalog? IBM watsonx Orchestrate is IBM's platform for building, running and governing AI agents that automate business work. Its Agent Catalog is the hub where teams discover prebuilt agents and tools from IBM and its partners. Partners bring their agents to the catalog through IBM Agent Connect, IBM's partner program for agent builders. Agents built with any external framework can join, which gives IBM clients a wider choice of specialized agents inside one environment. ## Which agents did Dynamiq announce? ### Medical Research Agent The Medical Research Agent helps clinicians process a patient's history and interview notes and search current PubMed articles to identify treatment options. The final decision stays with the medical professional; the agent saves time on research and brings relevant evidence forward faster. Healthcare teams can read more about running agents under HIPAA on our [healthcare page](/industries/healthcare). ### Legal Assistant Agent The Legal Assistant Agent helps legal teams prepare and review contracts using the organization's own repository of historical contracts and documents. It identifies relationships between documents and flags risks, which helps reviewers improve contract quality and work through more agreements. ### OCR Document Agent The OCR Document Agent lets users chat with scanned documents and unstructured files, extracting information from text, charts and infographics. It turns documents that previously needed manual processing into data teams can query; see [document review agents](/use-cases/document-review) for the broader pattern. ## How can IBM clients use and customize the agents? The agents work as ready-to-use solutions and as a starting point for custom agents. With Dynamiq, IBM clients and partners can: - **Deploy the prebuilt agents** for immediate productivity gains. - **Adapt them** to specific business requirements and proprietary data. - **Test and evaluate** changes in Dynamiq before they reach users. - **Move faster to production** in the watsonx environment. - **Run on a platform with SOC 2, HIPAA and GDPR in place,** in Dynamiq's cloud or in their own infrastructure. ## What did Dynamiq and IBM say? "The ability to quickly prototype with custom agents, testing them using selected tools, proprietary data, and real-time internet data, is key to agentic AI workflow automation in the enterprise environment," said Vitalii Duk, CEO and Founder of Dynamiq. "The integration of LLM-powered non-deterministic tasks, predefined robotic process automation, and human in the loop decision-making, coordinated by IBM watsonx Orchestrate, will help IBM Clients and Partners to enhance human performance, reduce error rates, gain faster insights from proprietary data, and improve overall business process agility." "We're excited for Dynamiq to make their agent available in IBM's Agent Catalog through Agent Connect, expanding the enterprise use cases available to clients," said Nick Holda, Vice President, AI Technology Partnerships at IBM. He added that letting third-party agents work together within the watsonx Orchestrate ecosystem gives clients ready-to-use agents and integrations that help them improve productivity and streamline workflows. ## How does this fit the Dynamiq and IBM partnership? Dynamiq is an IBM AI Technology Partner, and the Dynamiq platform is listed in the [IBM Partner Plus directory](https://www.ibm.com/partnerplus/directory/solution/0484). The partnership covers more than the catalog: - **IBM models.** Dynamiq agents can use IBM watsonx models, including the Granite family, next to models from other providers. - **IBM infrastructure.** Dynamiq can be self-hosted on IBM Cloud and Red Hat OpenShift, as described in our post on [deploying Dynamiq on IBM Cloud](/post/dynamiq-now-deployable-on-ibm-cloud-simplify-scale-and-secure-your-ai-projects). - **IBM's marketplace.** Dynamiq is available through IBM's marketplace, as well as AWS Marketplace and Microsoft Azure Marketplace. The [IBM partnership page](/partners/ibm) collects the details. ## About Dynamiq Dynamiq is the enterprise AI agent platform for regulated industries: chat, voice and workflow agents on one governed stack, deployed in Dynamiq's cloud or inside your own perimeter, with engineers who help take the first agents live. It supports 29 model providers, including IBM watsonx, and 2,100+ app integrations, with SOC 2, HIPAA and GDPR in place. _Statements regarding Dynamiq and IBM's future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only._ ## FAQ ### What is IBM watsonx Orchestrate? IBM watsonx Orchestrate is IBM's platform for building, deploying and governing AI agents and assistants that automate business workflows. Its Agent Catalog offers prebuilt agents and tools from IBM and its partners. ### What is IBM Agent Connect? IBM Agent Connect is IBM's partner program for agent builders. It lets partners integrate their specialized agents into watsonx Orchestrate and offer them in the Agent Catalog. ### Which Dynamiq agents were announced for the catalog? Dynamiq announced a Medical Research Agent for clinical research support, a Legal Assistant Agent for contract preparation and review, and an OCR Document Agent for querying scanned and unstructured documents. ### Can the agents be customized with our own data? Yes. The agents are designed as a starting point: teams can adapt their instructions, tools and data sources in Dynamiq, test the changes, and deploy them in Dynamiq's cloud or in their own infrastructure. ### Is Dynamiq an IBM partner? Yes. Dynamiq is an IBM AI Technology Partner. It works with IBM watsonx models, can be self-hosted on IBM Cloud and Red Hat OpenShift, and is listed in the IBM Partner Plus directory. --- # Agentic RAG with Apache Iceberg and Dynamiq on IBM watsonx.data URL: https://www.getdynamiq.ai/post/unlocking-enterprise-intelligence-with-agentic-rag-apache-iceberg-and-dynamiq-on-watsonx-data Published: 2025-04-04 How agentic RAG combines SQL over Apache Iceberg tables with vector search in Milvus on IBM watsonx.data, built with Dynamiq agents, with an HR example. Agentic RAG is retrieval-augmented generation in which an AI agent, rather than a fixed pipeline, decides how to answer each question: which sources to query, in what order, and whether the results are good enough. On IBM watsonx.data, that means one agent can run SQL against governed Apache Iceberg tables and search unstructured documents in Milvus, then combine both into a cited answer. Dynamiq provides the agent layer, with guardrails, approvals and a trace of every step. This post explains how agentic RAG differs from classic RAG, the architecture on watsonx.data, an HR assistant example, and what to watch for in production. ## What is agentic RAG? Classic RAG follows one fixed path: embed the question, retrieve similar passages, and generate an answer from them. That works for questions whose answer sits in a document. It fails when the answer needs numbers from a table, several lookups in sequence, or a check that the retrieved text actually answers the question. Agentic RAG puts an agent in charge of retrieval: | | Classic RAG | Agentic RAG | | -------------------- | ------------------------------------- | ---------------------------------------------------------------- | | Retrieval path | Fixed: one vector search per question | Chosen per question: SQL, vector search, both, or several rounds | | Data types | Mostly unstructured text | Structured tables and unstructured documents | | Multi-step questions | Poorly supported | The agent plans and runs follow-up queries | | Self-checking | None | The agent checks whether results answer the question and retries | | Auditability | Retrieved chunks | Every query, tool call and decision in a trace | See our [glossary entry on retrieval-augmented generation](/glossary/retrieval-augmented-generation) for the basics. ## Why combine structured and unstructured data? Most real business questions need both. "How many vacation days do I have left, and can I carry them into next year?" needs the employee's balance from an HR table and the carry-over rule from a policy document. A document-only assistant cannot answer the first part; a SQL-only assistant cannot answer the second. An agent that can use both tools answers the whole question and cites each source. ## What does the architecture look like on watsonx.data? ![Layered architecture: Dynamiq as the agent layer on top of unstructured and structured data sources, which sit on a shared data platform](/blog/unlocking-enterprise-intelligence-with-agentic-rag-apache-iceberg-and-dynamiq-on-watsonx-data/1-rag-agent.webp "The agent layer sits above both structured and unstructured data on one data platform") The design has three layers. ### Structured data: Apache Iceberg tables IBM watsonx.data is a lakehouse that uses Apache Iceberg as its open table format. Iceberg gives agentic RAG several useful properties: - **Time travel.** Queries can read a table as it was at an earlier snapshot, for example to compare current and past transactions or reproduce an answer for an audit. - **Schema evolution.** Columns can be added or changed without rewriting data, so new compliance fields do not break the agent's queries. - **ACID transactions.** Concurrent updates stay consistent, so an agent never reads a half-written table. - **Open format.** The same tables are readable by several query engines and stay in your object storage. ### Unstructured data: embeddings in Milvus Policies, contracts and other documents are parsed, split into chunks and embedded, using embedding models from watsonx.ai or another provider, and stored in Milvus, the vector database available in watsonx.data. Keeping document versions in the index lets the agent cite the policy that was in force. ### The agent layer: Dynamiq A Dynamiq agent receives the question and chooses its tools: a read-only SQL tool against watsonx.data's query engine for structured data, and a retrieval tool over the Milvus index for documents. For more complex work, a manager agent can delegate to a SQL specialist and a document specialist. Guardrails detect PII before text reaches a model, and every query, retrieval and model call is recorded in a trace. ## Example: an HR policy and entitlement assistant ![Architecture of an HR assistant: Dynamiq agents answer from HR policies stored as unstructured documents and employee records stored as structured tables, both on IBM watsonx.data](/blog/unlocking-enterprise-intelligence-with-agentic-rag-apache-iceberg-and-dynamiq-on-watsonx-data/2-rag-agent-internal-hr-system-expert.webp "An HR assistant that answers from both policy documents and employee records") HR teams answer the same questions every day, and most combine a personal record with a policy: leave balances and carry-over rules, eligibility for benefits, stock plan vesting and its terms. The assistant works like this: 1. **Employee records** such as leave balances and plan enrollments live in Iceberg tables, with standardized schemas and consistent updates. 2. **HR policies** such as leave, benefits and stock plan documents are embedded into a Milvus index with their effective dates. 3. **The agent** receives a question, queries the employee's own records through the SQL tool, retrieves the relevant policy passages, and answers with citations to both. 4. **Permissions** restrict every query to the asking employee's own data, and questions outside policy are handed to the HR team. Because the agent reads the current policy at question time, a new or changed policy is used as soon as it is indexed, without retraining a model. ## What are the benefits of agentic RAG on watsonx.data? - **Complete answers.** Questions that need both numbers and rules get one answer, with a citation for each part. - **Audit-ready.** Traces show every query and retrieval, and Iceberg snapshots let you reproduce the data an answer was based on. - **One governed source.** The agent reads the same tables as your analytics and reporting, instead of a copy that drifts. - **Fast adaptation.** New policies and schema changes flow into answers without rebuilding pipelines or retraining models. - **Storage efficiency.** Iceberg tables use compressed columnar files in object storage, and Milvus supports quantized indexes for large document sets. ## What should you watch out for in production? - **Text-to-SQL errors.** Give the agent read-only access to curated views, validate generated queries, and test with real questions before launch. - **Access control.** Enforce row-level permissions in the data layer, not only in the prompt, so an agent cannot read records the user may not see. - **Personal data.** Detect PII in inputs and outputs, and keep traces in the same governed environment as the data. - **Stale indexes.** Re-index documents when policies change, and keep effective dates so old versions are not cited as current. - **Cost and latency.** Cap the number of retrieval rounds, and route simple lookups to smaller models. ## How do Dynamiq and IBM work together? Dynamiq is an IBM AI Technology Partner. Teams can build agents and RAG applications with watsonx.ai models and watsonx.data vector stores alongside other providers, and deploy Dynamiq on IBM Cloud. Read more about the [IBM partnership](/partners/ibm). For a production example on IBM technology, see how we built a [cost-aware legal research workflow with IBM watsonx](/post/how-dynamiq-built-a-cost-aware-legal-research-workflow-with-ibm-watsonx), which uses Milvus on watsonx.data for internal contracts. ## FAQ ### What is agentic RAG? Agentic RAG is retrieval-augmented generation where an AI agent decides how to retrieve information for each question: which sources to query, in what order and whether to try again. It can combine structured data, such as SQL tables, with unstructured documents in one answer. ### What is the difference between RAG and agentic RAG? Classic RAG runs one fixed retrieval step, usually a vector search, before generating an answer. Agentic RAG lets an agent plan multiple retrieval steps across different sources and check its results, which handles multi-step questions and mixed data types. ### What is Apache Iceberg used for in AI applications? Apache Iceberg is an open table format for data lakes. In AI applications it gives agents consistent, governed tables to query, with time travel for reproducing past answers and schema evolution for changing data without breaking queries. ### Can an AI agent query a data lakehouse directly? Yes, through a SQL tool. In production, give the agent read-only access to curated views, enforce row-level permissions in the data layer and validate generated queries before they run. ### Does Dynamiq work with IBM watsonx.data? Yes. Dynamiq is an IBM AI Technology Partner with access to watsonx.ai models and watsonx.data vector stores, and Dynamiq can be deployed on IBM Cloud. --- # Generative AI for enterprises: practical ways to cut costs and boost sales URL: https://www.getdynamiq.ai/post/generative-ai-practical-ways-for-enterprises-to-cut-costs-and-boost-sales Published: 2025-02-06 Where generative AI and AI agents cut costs and grow revenue in large companies: support, back office, documents, sales and marketing, and how to measure it. Enterprises cut costs with generative AI by giving agents the high-volume, text-heavy work: support requests, invoice and document processing, report drafting and internal questions, with people reviewing the exceptions. They grow sales by giving sellers and marketers researched accounts, personalized content and faster responses. The gains are real but uneven, so the practical path is one measurable process at a time. This guide lists the processes where the savings and revenue show up first, how to measure them, and what keeps the gains from eroding. ## Where does generative AI cut costs? Savings come from processes with three traits: many cases, mostly text, and clear rules for what a good outcome looks like. Today these processes are handled by agents that use tools and act in your systems, not by chatbots that only draft answers. ### Customer support Support is the most proven area. Agents resolve routine requests end to end, such as order status, account changes and policy questions, and hand complex cases to a person with the full context. For the requests that stay with people, an assistant suggests answers and pulls up the right policy. The evidence is measured, not projected. An [NBER study of 5,179 customer support agents](https://www.nber.org/papers/w31161) found that access to a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled agents. See how a full [customer service agent](/use-cases/customer-service) is put together, from knowledge base to handoff. ### Back-office processing Finance and operations teams spend their days matching, checking and routing: invoices against purchase orders, payments against ledgers, expense claims against policy. An agent does the matching, prepares the exceptions with the evidence attached, and routes them for approval. Nothing posts until a person signs off, so the saving comes from preparation time, not from removing control. Read more about [back-office agents](/use-cases/back-office). ### Document review Contracts, insurance claims, loan files and supplier onboarding packs are long and inconsistent. Agents parse them, extract the fields your policy needs, compare them with your rules and flag what is missing or out of range. Reviewers start from a prepared summary with the source page attached instead of a blank page. ### Internal questions HR, IT and policy questions take up a surprising amount of expert time. A knowledge assistant answers them from your own documents, with a citation for every claim and only the documents each employee is allowed to see. ### Software engineering Coding assistants and agents write tests, document legacy systems and handle routine migrations. The output is reviewed by an engineer before it ships, which makes this one of the lowest-risk places to start. ## Where does generative AI grow revenue? Revenue gains are harder to attribute than savings, but the mechanisms are clear. - **Sales preparation.** An agent researches accounts, summarizes past conversations and drafts meeting briefs and follow-ups, so sellers spend more time with customers and less on preparation. - **Speed to response.** Leads and customer questions get an accurate answer in minutes, at any hour and in the customer's language. - **Personalized marketing at scale.** Campaigns, product descriptions and localized versions are drafted per segment, then reviewed for brand and legal compliance before they go out. - **Retention.** Support agents that resolve issues on the first contact, and flag unhappy customers early, protect revenue you already have. - **New product features.** AI features inside your own product, such as natural-language search, can be sold or used to win deals. That last point can move fast. A B2B accounting platform [shipped document search in one week](/case-studies/smart-document-search-for-a-b2b-accounting-platform-built-in-one-week-with-dynamiq) with one of our engineers, turning a long-requested feature into part of its product. For recurring sales and research work, a ready-made agent such as [AI Coworker](/product/chat) can run a prompt on a schedule against your connected apps, with each run saved as a conversation to review. Sellers can also hand it work from Slack or Microsoft Teams, without anyone building a workflow first. ## Which processes should you start with? Start where volume is high, the rules are written down and a mistake is cheap to catch. This table shows common starting points and what to measure: | Process | What the agent does | What to measure | | ---------------------------- | ------------------------------------------------- | --------------------------------------------------- | | Support requests | Resolves routine requests, hands off the rest | Resolution rate, handle time, customer satisfaction | | Invoice and payment matching | Matches records, prepares exceptions | Cases per analyst per day, error rate | | Contract and claims review | Extracts fields, checks rules, flags gaps | Review time per file, missed issues | | Employee questions | Answers from policies with citations | Tickets avoided, answer accuracy | | Sales preparation | Researches accounts, drafts briefs and follow-ups | Selling time per rep, pipeline created | | Marketing localization | Drafts versions per market for review | Cost and time per asset | ## How do you measure the ROI of generative AI? Measure cost per case, not usage. The steps are simple: 1. **Baseline.** Record volume, handle time, error rate and cost per case for the current process. 2. **Count every cost.** Include model usage, the platform, integration work and the time people spend reviewing agent output. 3. **Run a controlled pilot.** Put real cases through the agent next to the current process and compare quality, not just speed. 4. **Track after launch.** Cost per run, latency and quality drift over time as volumes and models change. The [AI agent ROI calculator](/resources/ai-agent-roi-calculator) helps you build the first estimate for a business case. Broad estimates are useful for context. McKinsey [estimated in 2023](https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier) that generative AI and related technologies could automate work activities that take up 60% to 70% of employees' time. Your own baseline is what will convince a CFO. ## What keeps the savings from eroding? Most disappointing results trace back to the same few gaps: - **No baseline.** Without before-and-after numbers, a working agent still looks like a cost. - **Hidden review work.** If people must check every output, the saving moves rather than disappears. Evaluations and targeted approvals keep review where it matters. - **Cost creep.** Longer prompts, bigger models and retries push up the cost per run. Traces with cost and latency per step show where the money goes. - **Quality drift.** Models and data change. Scoring a sample of live traffic catches regressions early. - **Risky actions without a gate.** An agent that can send, pay or file needs an approval step before it does, or one bad run erases a quarter of savings. On Dynamiq, [evaluations](/product/evaluations) run on test sets before release and on live deployments after, and observability shows cost and latency for every step of every run. ## FAQ ### How does generative AI reduce costs for enterprises? By taking over high-volume, text-heavy work such as support requests, document processing, reconciliations and internal questions, and by preparing cases so that people only review exceptions. The saving shows up as lower cost per case and faster handling, provided review work and model costs are counted. ### How can generative AI increase sales? It gives sellers more selling time through account research and meeting preparation, responds to leads and customers faster, personalizes marketing at scale, improves retention through better support, and lets companies add AI features to their own products. ### What is a realistic ROI for generative AI? It depends on the process. High-volume support and back-office processes usually show returns first because the cost per case is easy to measure. Build the estimate from your own baseline and include model, platform and review costs. ### Which departments benefit first from generative AI? Customer support, finance and operations, legal and compliance review, IT and HR service desks, software engineering, and sales and marketing. They share high volumes of text-based work with clear rules. ### What is the difference between generative AI and AI agents? Generative AI produces content such as text, code or images. An AI agent uses a generative model to plan steps, call tools and act in other systems until a task is done. Most cost savings now come from agents, because they finish the task instead of only drafting part of it. --- # Build an intelligent agent system for market analysis with DeepSeek URL: https://www.getdynamiq.ai/post/build-an-intelligent-agent-system-for-market-analysis-with-deepseek Published: 2025-01-27 Build a market analysis agent system on DeepSeek models: a research agent with search and code, a validation agent, and a manager that coordinates them. To build a market analysis agent system with DeepSeek, pair a research agent that gathers data with a validation agent that checks it, and let a manager agent coordinate the two. With the Dynamiq SDK, each agent takes a few lines of Python, and DeepSeek is one of 29 model providers you can switch to by changing a single node. The example below produces a sourced market brief for an online seller. ## Why use DeepSeek for agent workloads? DeepSeek became a serious option for enterprise builders in January 2025, when it released DeepSeek-R1, a reasoning model with open weights under the MIT license. The [DeepSeek-R1 paper](https://arxiv.org/abs/2501.12948) describes how reinforcement learning taught the model to reason step by step, and the weights are published on Hugging Face. Three properties make DeepSeek's models worth testing for agents: - **Open weights.** You can run the models in your own environment, which matters when data may not leave your perimeter. - **Two modes of work.** Its current API models support a fast non-thinking mode and a slower thinking mode, so you can match the mode to the task. - **Price per token.** Compare current prices on [DeepSeek's pricing page](https://api-docs.deepseek.com/quick_start/pricing) against your volume before you decide. Before you send business data to any hosted API, check where the provider processes and stores it, and what its terms allow. For regulated data, the safer pattern is to run open weights on infrastructure you control, which this guide covers below. ## What does the market analysis system look like? The system has three layers: | Layer | Components | Job | | ------------- | ----------------------------------- | ------------------------------------------------------ | | Tools | Tavily web search, E2B code sandbox | Fetch current data and run calculations | | Agents | Research agent, validation agent | Gather findings; check them against sources | | Orchestration | Manager agent | Plan, delegate, and write the brief from checked facts | The manager decides which agent to call and when, so the same system can answer a narrow pricing question or produce a full market brief. That is the adaptive orchestration pattern described in our guide to [agent orchestration patterns](/post/agent-orchestration-patterns-in-multi-agent-systems-linear-and-adaptive-approaches-with-dynamiq). ## Step 1: Install and configure ```bash pip install dynamiq export DEEPSEEK_API_KEY="..." export TAVILY_API_KEY="..." export E2B_API_KEY="..." ``` ## Step 2: Set up the models Use the faster model for research and coordination and the stronger model for validation. ```python from dynamiq.connections import DeepSeek as DeepSeekConnection from dynamiq.nodes.llms import DeepSeek # Model names from DeepSeek's API documentation at the time of writing; check its current list. fast_llm = DeepSeek(connection=DeepSeekConnection(), model="deepseek-flash") strong_llm = DeepSeek(connection=DeepSeekConnection(), model="deepseek-v4-pro") ``` ## Step 3: Create the research and validation agents Each agent gets a `description`, which the manager reads to decide when to call it. ```python from dynamiq.connections import E2B as E2BConnection from dynamiq.connections import Tavily as TavilyConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.tools.e2b_sandbox import E2BInterpreterTool from dynamiq.nodes.tools.tavily import TavilyTool from dynamiq.nodes.types import Behavior, InferenceMode search = TavilyTool(connection=TavilyConnection()) sandbox = E2BInterpreterTool(connection=E2BConnection()) research_agent = Agent( name="Research agent", description="Collects current market data from the web and runs calculations.", role=( "You research markets. Find current, sourced data with web search, use the code " "interpreter for any calculation, and cite a source for every figure." ), llm=fast_llm, tools=[search, sandbox], inference_mode=InferenceMode.XML, max_loops=10, behaviour_on_max_loops=Behavior.RETURN, ) validation_agent = Agent( name="Validation agent", description="Checks research findings for unsupported claims, stale data and contradictions.", role=( "You fact-check market research. Verify each figure against its source with web search, " "flag anything unsupported or out of date, and return the corrected findings." ), llm=strong_llm, tools=[search], inference_mode=InferenceMode.XML, max_loops=8, behaviour_on_max_loops=Behavior.RETURN, ) ``` `InferenceMode.XML` makes the agents express tool calls in XML tags, which works with any chat model, including ones without native function calling. `Behavior.RETURN` makes an agent return its best answer when it reaches its loop limit, instead of raising an error. ## Step 4: Add a manager and run the analysis The manager has the two agents as its tools. It delegates research, sends the findings for validation, and writes the brief only from what survived the check. ```python from dynamiq import Workflow from dynamiq.flows import Flow manager = Agent( name="Market analysis manager", role=( "Coordinate a market analysis. Delegate research to the research agent, send its findings " "to the validation agent, and write the final brief only from validated findings. " "Call sub-agents with {'input': ''} payloads." ), llm=fast_llm, tools=[research_agent, validation_agent], inference_mode=InferenceMode.XML, max_loops=12, behaviour_on_max_loops=Behavior.RETURN, ) workflow = Workflow(flow=Flow(nodes=[manager])) result = workflow.run( input_data={ "input": ( "I sell handmade ceramics on Etsy. Analyze demand, pricing and competition for the " "next quarter and recommend three product lines, with sources." ) } ) print(result.output[manager.id]["output"]["content"]) ``` ## What does the output look like? The brief typically covers market viability, recommended product categories with the evidence behind each, pricing guidance, the main risks and a list of sources. Because the validation agent checks each figure, unsupported numbers are flagged or removed instead of passed through. Treat the brief as research, not as a verdict: the figures are only as current as the sources the agents found. Keep the source list, and spot-check the numbers that drive a decision. ## How do you run DeepSeek models on your own infrastructure? Serve the open weights with an inference server that exposes an OpenAI-compatible API, such as vLLM, then point the agents at it with the SDK's `CustomLLM` node. Nothing else in the workflow changes. ```python import os from dynamiq.connections import HttpApiKey from dynamiq.nodes.llms import CustomLLM self_hosted_llm = CustomLLM( connection=HttpApiKey(url="https://llm.internal.example.com/v1", api_key=os.environ["LLM_API_KEY"]), model="your-deepseek-deployment", # the model name your server exposes provider_prefix="openai", # route as an OpenAI-compatible endpoint ) ``` On the Dynamiq platform, the AI Gateway gives agents one endpoint for the most-used hosted models, and managed inference runs open models for you; see [models](/product/models). For teams that need everything inside their own cloud or data center, Dynamiq itself can be self-hosted; see [where Dynamiq runs](/deployment/self-hosted). ## When should you choose DeepSeek? DeepSeek is worth testing when: - **You need open weights** to keep data in your environment. - **The workload is high-volume** and token cost dominates. - **Tasks benefit from explicit reasoning,** such as validation or multi-step analysis. It is a weaker fit when your organization restricts the provider or the model's origin, or when a task needs a capability you have only validated on another model. Because the agent code doesn't depend on the provider, you can run the same workflow on several models and compare the results with [evaluations](/product/evaluations) before you commit. ## FAQ ### Is DeepSeek open source? DeepSeek publishes open weights for many of its models. DeepSeek-R1 was released in January 2025 under the MIT license, which allows commercial use and modification. Check the license of the specific model version you plan to use. ### Can I use DeepSeek models with the Dynamiq SDK? Yes. The SDK includes a `DeepSeek` LLM node for DeepSeek's API, and the `CustomLLM` node connects to any OpenAI-compatible endpoint, including a server hosting DeepSeek's open weights in your own environment. ### Why use two agents instead of one? Separating research from validation catches more mistakes. The research agent is optimized to find information quickly; the validation agent has a single job, checking each figure against its source, and can run on a stronger model. ### Is it safe to send company data to DeepSeek's API? That depends on your data and your policies. Review where the provider processes and stores data and what its terms allow. For sensitive or regulated data, run the open weights on infrastructure you control instead of calling the hosted API. ### Which search tool should the research agent use? The SDK includes several, including Tavily, Exa, Scale SERP, Firecrawl and Jina search. Tavily lets you restrict results to specific domains, which helps when only certain sources are acceptable. --- # Understanding agentic AI: definition, benefits and applications in business URL: https://www.getdynamiq.ai/post/understanding-agentic-ai Published: 2024-12-19 What agentic AI is, how it works, how it differs from generative AI and AI agents, where businesses use it, and how to implement it with the right controls. Agentic AI is artificial intelligence that pursues a goal rather than just answering a prompt: it plans the steps, uses tools and data to carry them out, checks the results and adapts when something changes. The systems that do this are called AI agents, and they are usually built on large language models. In business, agentic AI now resolves customer requests, prepares back-office cases, reviews documents and assists compliance teams, with guardrails, human approvals and full traces keeping its autonomy within set bounds. This guide explains how agentic AI works, how it differs from generative AI and traditional automation, the types of agents, where businesses use it and how to implement it safely. ## What is agentic AI? Agentic AI describes systems that act with agency: given an objective, they decide what to do next, do it, and keep going until the objective is met or they need a person. Five properties define it: - **Goal-directed.** It works toward an outcome, such as "resolve this refund request", not a single reply. - **Planning.** It breaks the goal into steps and orders them, adjusting the plan as results come in. - **Tool use.** It calls APIs, queries databases, searches documents and runs code to act in real systems. - **Memory.** It keeps context within a task and, where useful, across sessions. - **Bounded autonomy.** It acts on its own within limits, and escalates or asks for approval beyond them. The last property is what makes agentic AI usable in regulated industries. Autonomy is a setting you choose per step, not an all-or-nothing trait. See our [glossary entry on agentic workflows](/glossary/agentic-workflow) for a short definition. ## How does agentic AI work? Agentic systems run a loop: 1. **Perceive.** Gather the request and the context it needs: account data, documents, prior conversation, business rules. 2. **Reason.** A language model interprets the goal, plans the next step and chooses a tool. 3. **Act.** The system calls the tool, for example checking an order, drafting a letter or updating a record. 4. **Check.** It reads the result, validates it against rules, and decides whether to continue, retry, escalate or finish. 5. **Learn.** Outcomes, human corrections and evaluation results feed back into instructions, test sets and rules. For larger jobs, several agents coordinate: a manager agent or an explicit workflow graph assigns work to specialists and combines their results. Our guide to [multi-agent systems](/post/multi-agent-ai-systems-definition-benefits-limitations-how-to-build) covers those patterns. ## How is agentic AI different from generative AI and automation? | | Generative AI | Traditional automation (RPA, rules) | Agentic AI | | ----------------- | -------------------------------------- | ----------------------------------- | ------------------------------------------ | | What it does | Creates content from a prompt | Executes fixed steps | Plans and executes steps toward a goal | | Handles variation | Within one response | Poorly; breaks on unexpected input | Well; adapts or escalates | | Acts in systems | No, unless connected by another system | Yes, along a scripted path | Yes, choosing tools as it goes | | Typical output | Text, images, code | Completed transaction | Completed task, with a record of each step | | Main risk | Wrong or unsafe content | Brittle processes | Wrong actions, if not bounded | They work together. An agent uses generative AI to read and write, and calls existing automation as one of its tools. Agentic AI is the layer that decides what to do next. ## What is the difference between agentic AI and AI agents? The terms describe the same idea at different levels. Agentic AI is the approach: building systems that pursue goals with planning, tools and autonomy. An AI agent is a concrete system that does so, such as a support agent or a document review agent. An enterprise's agentic AI program might include one agent or dozens, plus the workflows, guardrails and monitoring that coordinate them. See our [guide to LLM agents](/post/llm-agents-explained-complete-guide-in-2025) for how a single agent works. ## What are the types of AI agents? The classic taxonomy from AI research still describes how agents decide, and each type has a place in modern systems: | Type | How it decides | Business example | | ---------------------------- | --------------------------------------------------------- | ------------------------------------------------------------- | | Simple reflex | Fixed condition-action rules on the current input | Flag every invoice above a set amount | | Model-based reflex | Rules plus a record of state | Pause a claim until a missing document arrives | | Goal-based | Plans actions to reach a goal | Find and book the best supplier slot for a delivery | | Utility-based | Weighs options by a score, such as cost against speed | Choose between a cheap model and a strong one per request | | Learning | Improves from feedback and outcomes | Adjust triage rules as reviewers correct routing mistakes | | LLM agent | A language model reasons, plans and calls tools in a loop | Resolve a customer's card issue end to end | | Multi-agent and hierarchical | A manager or graph coordinates specialized agents | Registry, screening and writing agents preparing a KYC review | Production systems mix types. Deterministic rules handle the steps that must never vary, and LLM agents handle the ones that need judgment. ## What are the benefits of agentic AI for business? - **End-to-end automation.** Agents complete tasks across systems instead of producing a draft someone else must act on. - **Exception handling.** Agents deal with the variation that breaks scripted automation, and route what they cannot resolve. - **Scale without queues.** Volume peaks do not create backlogs when agents handle routine cases. - **Consistency and records.** Every step follows the same instructions and leaves a trace that auditors can review. - **Expert time.** People spend their time on decisions and exceptions rather than gathering information. The evidence is strongest in customer service: an Asian neo-bank [automated about 85% of support inquiries](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) with an agent that uses tools, memory and a knowledge base. ## What are the risks of agentic AI, and how do you bound them? Agents that act can also act wrongly. The main risks are wrong actions from a misunderstood request, prompt injection hidden in documents or messages, agents with more permissions than their task needs, runaway cost from long loops, and weak audit trails. A practical way to manage them is to set an autonomy level for each step: | Level | The agent | Use for | | ----------------- | -------------------------------------------- | ----------------------------------------------------- | | Assist | Drafts and recommends; a person acts | Credit decisions, legal advice, anything irreversible | | Act with approval | Prepares the action; a person approves it | Payments, account changes, outbound letters | | Act within limits | Acts alone under thresholds and rules | Refunds under a set amount, routine updates | | Act and report | Acts alone; people review samples and traces | Classification, routing, internal summaries | Combine the levels with least-privilege tools, input screening for PII and prompt injection, loop and cost caps, and a trace of every run. Our [guardrails](/product/guardrails) page covers the controls. ## Where do businesses use agentic AI? - **Customer service:** resolving routine requests and handing complex ones to people with context. - **Back-office operations:** reconciliations, exception handling and report preparation in finance and operations teams. See [back-office automation](/use-cases/back-office). - **Document review and underwriting:** extracting and checking fields in loan files, claims and contracts. - **Compliance and financial crime:** preparing KYC and AML cases and monitoring regulatory change for analysts. - **IT and HR service desks:** answering policy questions and handling routine requests. - **Research and knowledge work:** answering questions across internal documents with citations. - **Software engineering:** writing and testing code for engineers to review. ## How do you implement agentic AI? 1. **Start with a high-volume, well-understood process** where exceptions are easy to spot and a mistake is cheap to catch. 2. **Measure a baseline** and tie the agent to business KPIs, such as cases handled per day or time to resolution. 3. **Plan the integrations** early; an agent is only as useful as the systems it can read and act on. 4. **Set autonomy levels and escalation paths** for every step, with named owners for exceptions. 5. **Build feedback loops** so human corrections become test cases and rule changes. 6. **Involve compliance and operations from the start**, not at the end. 7. **Expand incrementally**, reusing proven agents and controls in the next process. ## How does Dynamiq help businesses implement agentic AI? [Agent Builder](/product/agents) lets teams build agents and multi-agent workflows on a visual canvas or with the open-source Python SDK, on one engine. Choice nodes handle rule-based steps, approvals with editable fields gate any tool step, and a prompt can generate a first draft of a workflow. Guardrails detect PII and prompt injection, evaluations run on test sets and live deployments, and every run is traced with cost and latency. For work that does not need a custom workflow, AI Coworker is a ready-made agent with its own cloud sandbox, connectors, skills, scheduled tasks and subagents, in the browser, Slack, Microsoft Teams and Telegram. The platform runs in Dynamiq Cloud or self-hosted in your own cloud or data center. ## FAQ ### What is agentic AI in simple terms? Agentic AI is AI that gets things done: you give it a goal, and it plans the steps, uses tools and data to carry them out, checks the results and asks a person when it reaches the limits you set. ### What is the difference between agentic AI and generative AI? Generative AI creates content, such as text or images, in response to a prompt. Agentic AI uses generative models to plan and take actions toward a goal across multiple steps and systems. ### What is the difference between agentic AI and AI agents? Agentic AI is the approach of building goal-directed systems with planning, tools and bounded autonomy. AI agents are the individual systems that implement it, such as a support agent or a document review agent. ### Is agentic AI safe for regulated industries? It can be, when autonomy is set per step, agents have only the permissions their task needs, consequential actions require human approval, inputs are screened for PII and prompt injection, and every run is traced for audit. ### What are examples of agentic AI? A support agent that checks an order and issues a refund within limits, a back-office agent that reconciles payments and prepares exceptions for approval, a document agent that checks loan files against policy, and a compliance agent that prepares KYC reviews for analysts. --- # Build an automated social media management agent with Dynamiq URL: https://www.getdynamiq.ai/post/build-an-automated-social-media-management-agent-with-dynamiq Published: 2024-12-09 Build an agent that reads a client list, drafts a personalized email for each contact and sends it through Mailgun after a person approves the drafts. An automated social media management agent takes the repetitive part of client communication off a manager's plate: it reads a client list, writes a personalized message for each contact and sends it through your email provider. The version in this tutorial, built with the open-source Dynamiq SDK, drafts every email, asks a person to approve the batch and only then sends. That approval step is what makes the pattern safe to use with real clients. ## What does the agent do? The agent works through five steps in one run: 1. **Reads the client list** in a sandboxed code interpreter, so it can handle CSV or spreadsheet files. 2. **Drafts one email per contact** using each person's name, title, audience and interests. 3. **Shows the drafts and asks for approval** before anything leaves the building. 4. **Sends the approved emails** through Mailgun, one message per contact. 5. **Reports what it sent,** so the run leaves a clear record. Social media and marketing managers use this pattern for campaign follow-ups, event invitations and account updates, where every message should feel personal but the structure repeats. ## What do you need? - Python 3.10 or later and the SDK: `pip install dynamiq requests` - An `OPENAI_API_KEY` for the model and an `E2B_API_KEY` for the code sandbox - A Mailgun account with a verified sending domain, stored as `MAILGUN_API_KEY`, `MAILGUN_DOMAIN` and `MAILGUN_FROM` ## Step 1: Prepare the client list Any tabular file works. This example uses a CSV with fictional contacts: ```csv email,audience,first_name,last_name,title,interests jordan@example.com,Operations leaders exploring AI for support,Jordan,Lee,Head of Operations,customer experience sam@example.com,Finance teams automating month-end close,Sam,Patel,Finance Director,reporting ``` Test with addresses you own, or with your provider's sandbox domain, before you point the agent at a real list. ## Step 2: Create the send tool The `function_tool` decorator turns a Python function into a tool the agent can call. Its type hints become the input schema, and its docstring tells the agent when to use it. ```python import os import requests from dynamiq.nodes.tools.function_tool import function_tool @function_tool def send_email(to: str, subject: str, text: str, **kwargs) -> str: """Send one approved plain-text email through Mailgun. Use only after the drafts are approved.""" response = requests.post( f"https://api.mailgun.net/v3/{os.environ['MAILGUN_DOMAIN']}/messages", # EU domains use api.eu.mailgun.net auth=("api", os.environ["MAILGUN_API_KEY"]), data={"from": os.environ["MAILGUN_FROM"], "to": to, "subject": subject, "text": text}, timeout=30, ) response.raise_for_status() return f"Sent to {to}" ``` Keeping the tool this narrow is deliberate: it can send one plain-text email and nothing else. Any provider with an HTTP API works the same way. ## Step 3: Create the agent The agent gets three tools: a code interpreter to read the file, a feedback tool to ask for approval, and the send tool. ```python from dynamiq.connections import E2B as E2BConnection from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.tools.e2b_sandbox import E2BInterpreterTool from dynamiq.nodes.tools.human_feedback import HumanFeedbackTool agent = Agent( name="Client outreach agent", role=( "You write short, personal emails to a company's clients. Read the client file with the " "code interpreter. For each contact, draft one email that uses their name, title, audience " "and interests, and ends with the signature you are given. Show all drafts and ask for " "approval with the feedback tool. Send only the approved drafts, one email per contact, " "then list what was sent." ), llm=OpenAI(connection=OpenAIConnection(), model="gpt-4o", temperature=0.3), tools=[ E2BInterpreterTool(connection=E2BConnection()), # reads E2B_API_KEY HumanFeedbackTool(), # asks for approval in the console send_email(), ], max_loops=20, ) ``` ## Step 4: Run it with the client file Pass the file in the `files` input. The agent uploads it to its sandbox and reads it with code. ```python import io with open("clients.csv", "rb") as f: client_file = io.BytesIO(f.read()) client_file.name = "clients.csv" # the agent sees the file under this name result = agent.run( input_data={ "input": ( "Draft a personalized email for every contact in the attached file. " "Sign each one 'Best regards, the Dynamiq team'. Ask for approval before sending anything." ), "files": [client_file], } ) print(result.output["content"]) ``` The console shows the drafts and waits for your answer. Reply with approval, or with edits such as "shorten the second email", and the agent revises before it sends. A version of this example, with a Mailgun tool written as a full custom node, lives in the SDK repository under [examples/use_cases/smm_manager](https://github.com/dynamiq-ai/dynamiq/tree/main/examples/use_cases/smm_manager). ## How do you keep automated outreach safe? Sending email on a company's behalf is an action with consequences, especially in regulated industries. Build these controls in from the start: - **Approve before sending.** On the Dynamiq platform, an [approval](/product/guardrails) can pause any tool step for a person, who can edit fields such as the subject or body before it runs. - **Respect consent and opt-outs.** Email only contacts you have a lawful basis to contact, and include a way to unsubscribe. - **Check regulated wording.** In financial services, marketing messages can need compliance review, so route those drafts to the right reviewer. - **Keep a record.** Every run is traced, including which drafts were approved and what was sent. ## Can you do this without writing code? Yes. [AI Coworker](/product/chat) can run the same job from a chat message: connect your email and CRM, attach the client list and ask for drafts. It works in the browser and in Slack, Microsoft Teams and Telegram, and it can repeat the task on a schedule. Mailgun, Gmail and HubSpot are among the 2,100+ app integrations an agent can use, so it can send through the provider you already have. Browse them on the [integrations page](/product/integrations). ## FAQ ### Can an AI agent send emails on its own? Yes, if you give it a send tool. For anything that reaches clients, keep a person in the loop: have the agent draft first and send only what a reviewer approves, as in this tutorial. ### Which email providers can the agent use? Any provider with an API. This tutorial uses Mailgun through a small Python function; on the Dynamiq platform, the same agent can use Gmail, HubSpot, Mailgun and many other integrations without custom code. ### How does the agent read a CSV or Excel file? It uses a code interpreter running in a sandbox. The agent writes and runs Python to parse the file, which handles CSV, Excel and other formats without a special parser for each. ### How do you stop the agent from emailing the wrong people? Test against addresses you own or a sandbox domain, keep the send tool narrow, require approval before sending, and review the trace of each run. You can also check recipients against an allowlist inside the send function. ### Can you schedule the agent to run every week? Yes. AI Coworker supports scheduled tasks, and a deployed workflow can run on a cron schedule or when an event arrives from a connected app. --- # Building an intelligent Linear app assistant URL: https://www.getdynamiq.ai/post/building-an-intelligent-linear-app-assistant Published: 2024-12-05 Build a Linear assistant that answers questions about projects and teams and writes well-scoped issues from chat, using Linear's MCP server and Dynamiq. An intelligent Linear assistant turns a chat message like "create two tasks for next sprint" into well-scoped Linear issues, and answers questions about your projects and teams. The simplest way to build one is to connect an agent to Linear's hosted MCP server, which exposes Linear's own tools for finding, creating and updating issues, projects and comments. This tutorial does that with the open-source Dynamiq SDK, adds memory and streaming, and puts the agent behind a small chat UI. ## What can the assistant do? Once connected, the assistant can: - **Answer questions** about projects, teams and issues, such as "what's still open in the payments project?" - **Create issues** with a clear title, context and a checklist of deliverables, from a one-line request. - **Follow up** on earlier messages in the same conversation, such as "add a third task for QA". - **Ask before guessing** when a request could belong to more than one team or project. ## How does the agent connect to Linear? The [Model Context Protocol](/glossary/model-context-protocol) (MCP) is an open standard for giving AI agents tools from other applications. Linear runs a hosted MCP server, so the agent gets Linear's actions without custom integration code. According to [Linear's MCP documentation](https://linear.app/docs/mcp), the server lives at `https://mcp.linear.app/mcp` and accepts an API key or OAuth token in an `Authorization: Bearer` header, as well as an interactive sign-in flow. This tutorial uses an API key; create one in Linear's settings with only the permissions the assistant needs. ## Step 1: Install and configure ```bash pip install dynamiq streamlit export OPENAI_API_KEY="..." export LINEAR_API_KEY="..." ``` ## Step 2: Connect to Linear's MCP server `MCPServer` discovers the tools a server offers and hands them to the agent. Print their names once, then decide which ones the assistant should have. ```python import os from dynamiq.nodes.tools.mcp import MCPServer, MCPStreamableHTTP linear = MCPServer( name="linear", connection=MCPStreamableHTTP( url="https://mcp.linear.app/mcp", headers={"Authorization": f"Bearer {os.environ['LINEAR_API_KEY']}"}, ), ) print([tool.name for tool in linear.get_mcp_tools()]) ``` To limit what the agent can do, pass `include_tools` with the tool names it may use, or `exclude_tools` with the ones it may not, such as anything that deletes or archives. ## Step 3: Create the agent with memory and streaming Session memory lets the agent follow a conversation, and streaming sends the answer to the UI as it is written. ```python from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.flows import Flow from dynamiq.memory import Memory from dynamiq.memory.backends import InMemory from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI from dynamiq.types.streaming import StreamingConfig, StreamingMode agent = Agent( name="PM assistant", llm=OpenAI(connection=OpenAIConnection(), model="gpt-4o", temperature=0.2), role=( "You help a product team manage work in Linear. Look up teams and projects before you " "create anything. Write issues with a clear title, the context, and a checklist of " "deliverables. If the team or project is ambiguous, ask before creating the issue." ), tools=[linear], memory=Memory(backend=InMemory()), streaming=StreamingConfig(enabled=True, mode=StreamingMode.FINAL), max_loops=12, ) workflow = Workflow(flow=Flow(nodes=[agent])) ``` `StreamingMode.FINAL` streams only the final answer. Switch to `StreamingMode.ALL` if you also want to show the agent's intermediate steps, such as each Linear lookup. ## Step 4: Stream answers from the workflow Run the workflow in a background thread and read the stream as it arrives. The streaming handler stops iterating when the workflow finishes. ```python import threading from dynamiq.callbacks.streaming import StreamingIteratorCallbackHandler from dynamiq.runnables import RunnableConfig def ask(message: str, user_id: str, session_id: str): handler = StreamingIteratorCallbackHandler() runner = threading.Thread( target=workflow.run, kwargs={ "input_data": {"input": message, "user_id": user_id, "session_id": session_id}, "config": RunnableConfig(callbacks=[handler]), }, ) runner.start() for event in handler: if event.event != "streaming" or not isinstance(event.data, dict): continue delta = event.data.get("choices", [{}])[0].get("delta", {}).get("content") if isinstance(delta, str): yield delta runner.join() ``` ## Step 5: Add a chat interface Save the code above as `assistant.py`. This Streamlit app keeps one conversation per browser session and writes the streamed answer as it arrives. Start it with `streamlit run app.py`. ```python import uuid import streamlit as st from assistant import ask st.title("Linear assistant") if "session_id" not in st.session_state: st.session_state.session_id = str(uuid.uuid4()) st.session_state.messages = [] for message in st.session_state.messages: with st.chat_message(message["role"]): st.markdown(message["content"]) if prompt := st.chat_input("Ask about a project or describe an issue"): st.session_state.messages.append({"role": "user", "content": prompt}) with st.chat_message("user"): st.markdown(prompt) with st.chat_message("assistant"): reply = st.write_stream(ask(prompt, user_id="pm-team", session_id=st.session_state.session_id)) st.session_state.messages.append({"role": "assistant", "content": reply}) ``` Try "Show me our teams and active projects", then "Create an issue in the data team's project for an exploratory analysis of churn, with deliverables". Open Linear to see the issue the agent wrote. ## How do you keep the assistant from doing too much? An assistant that can write to your project tracker needs the same care as a new team member with edit access: - **Scope the credentials.** Use an API key with only the permissions the assistant needs. - **Scope the tools.** Use `include_tools` or `exclude_tools` so the agent cannot delete or archive anything. - **Require approval for writes.** On the Dynamiq platform, an [approval](/product/guardrails) can pause the step that creates an issue, so a person can edit its fields before it runs. - **Review what happened.** Every run is traced, including each Linear call the agent made. ## Can you do this without writing code? Yes. [AI Coworker](/product/chat) connects to Linear as one of its app connectors and works in the browser and in Slack, Microsoft Teams and Telegram. A team can ask it for a status summary in a Slack thread or have it draft issues from meeting notes, without building anything. Linear is one of the 2,100+ app integrations available to agents, alongside Jira, Asana and other trackers; see the [integrations page](/product/integrations). ## FAQ ### Can AI create Linear issues automatically? Yes. An agent connected to Linear's MCP server can create issues with titles, descriptions and deliverables from a plain-language request. Keep an approval step or a narrow set of tools if you want a person to confirm each change. ### What is the Linear MCP server? It is a hosted server from Linear that implements the Model Context Protocol. It gives AI agents tools for finding, creating and updating Linear objects such as issues, projects and comments, authenticated with OAuth or an API key. ### How do you stop the assistant from changing things it shouldn't? Combine a restricted API key, a limited tool list with `include_tools` or `exclude_tools`, clear instructions in the agent's role, and an approval step before any write. Traces show exactly what the agent did afterward. ### Does the assistant remember earlier messages? Within a conversation, yes: session memory is scoped by user and session, so follow-ups like "add one more task" work. For preferences that should persist across conversations, add long-term memory, as shown in our guide to [conversational agents with memory](/post/building-a-conversational-ai-agent-with-dynamiq). ### Can the same approach work with Jira or other trackers? Yes. Point the agent at another MCP server or give it tools from the platform's app integrations, which include Jira and other project management tools. The agent code stays the same. --- # Build an autonomous data analyst using Dynamiq, E2B and Together AI URL: https://www.getdynamiq.ai/post/how-to-build-an-autonomous-data-analyst-using-dynamiq-e2b-and-together-ai Published: 2024-11-06 Build a data analyst agent in Python with the open-source Dynamiq SDK: it reasons with a Together AI model and writes and runs code in an E2B sandbox. An autonomous data analyst is an AI agent that answers questions about your data by writing and running code, not by guessing from the numbers in a prompt. You give it a file and a question; it plans the analysis, runs Python in an isolated sandbox, reads the results, fixes its own errors and writes up what it found. This tutorial builds one with the open-source Dynamiq SDK, an E2B sandbox and a model served by Together AI. The code below targets the current SDK, where every agent is an `Agent`. If you followed an earlier version of this post that used `ReActAgent`, the concepts carry over unchanged. ## What can a data analyst agent do? A data analyst agent handles the parts of an analysis that follow a repeatable method: - **Load and inspect data** from CSV, Excel, JSON or other files, and check types, gaps and outliers. - **Transform data** by cleaning, joining, filtering and aggregating it with pandas. - **Run statistics and models** such as correlations, regressions, clustering and simple forecasts. - **Make charts** of distributions, trends and relationships. - **Report results** in markdown with the method, the numbers and the caveats. The important design choice is where the code runs. Executing model-written code on your laptop or server is risky, so the agent sends it to a sandbox: a short-lived, isolated machine that E2B starts in the cloud. If the code fails, the agent reads the error, fixes it and tries again. ## What do you need? - Python 3.11 to 3.13. - The Dynamiq SDK: `pip install dynamiq`. The E2B client is included. - A [Together AI](https://www.together.ai/) API key in `TOGETHER_API_KEY`. - An [E2B](https://e2b.dev/) API key in `E2B_API_KEY`. - A data file. This tutorial uses a CSV of house prices with `bedrooms`, `bathrooms` and `price` columns. ## Step 1: Set up the model and the code interpreter The `TogetherAI` node calls any chat model hosted on Together AI, and the connection reads your key from the environment. The `E2BInterpreterTool` gives the agent a sandbox where it can run Python and shell commands, install packages and read uploaded files. ```python import io import os from dynamiq import Workflow from dynamiq.callbacks import TracingCallbackHandler from dynamiq.connections import E2B as E2BConnection from dynamiq.connections import TogetherAI as TogetherAIConnection from dynamiq.flows import Flow from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import TogetherAI from dynamiq.nodes.tools.e2b_sandbox import E2BInterpreterTool from dynamiq.nodes.types import Behavior, InferenceMode from dynamiq.runnables import RunnableConfig llm = TogetherAI( connection=TogetherAIConnection(), # reads TOGETHER_API_KEY model="moonshotai/Kimi-K3", temperature=0.1, ) code_tool = E2BInterpreterTool(connection=E2BConnection()) # reads E2B_API_KEY ``` Any model on Together AI that follows instructions and tool formats reliably will work; swap the `model` string for another from Together's model list. To use a different provider, swap the LLM class, for example `OpenAI` or `Anthropic` from `dynamiq.nodes.llms`. The rest of the code stays the same. ## Step 2: Create the agent The `role` is the agent's standing instructions. Tell it where files are, how to work and what the answer should look like. Uploaded files land in `/home/user/input` inside the sandbox. ```python ANALYST_ROLE = """ You are a senior data analyst who answers questions by writing and running Python. - Uploaded files are in /home/user/input. Inspect a file before you analyze it. - Use pandas, numpy, scikit-learn and matplotlib. Install missing packages with the tool. - Combine loading, cleaning and analysis into as few tool calls as possible. - Always print results. If code fails, read the error, fix the code and run it again. - Answer in markdown: a two-sentence summary first, then method, results and caveats. """ agent = Agent( name="data-analyst", id="data-analyst", llm=llm, tools=[code_tool], role=ANALYST_ROLE, inference_mode=InferenceMode.XML, max_loops=12, behaviour_on_max_loops=Behavior.RETURN, ) ``` `InferenceMode.XML` asks the model to format its reasoning and tool calls as XML, which works reliably across providers. `max_loops` caps the number of reason, act and observe cycles, and `Behavior.RETURN` returns the best answer so far instead of raising an error if the cap is reached. ## Step 3: Pass in a file and run the analysis Files go in as `BytesIO` objects. The `name` becomes the file name in the sandbox, and the optional `description` tells the agent what the file contains. ```python def load_file(path: str, description: str) -> io.BytesIO: with open(path, "rb") as f: data = io.BytesIO(f.read()) data.name = os.path.basename(path) # saved as /home/user/input/ data.description = description return data house_prices = load_file( "data/house_prices.csv", "Comma-separated. Columns: bedrooms, bathrooms, price.", ) tracing = TracingCallbackHandler() workflow = Workflow(flow=Flow(nodes=[agent])) result = workflow.run( input_data={ "input": ( "Fit a linear regression of price on bedrooms and bathrooms. " "Use a fixed random seed and a held-out test set, report the test MSE " "and R-squared, and explain which feature matters more." ), "files": [house_prices], }, config=RunnableConfig(callbacks=[tracing]), ) print(result.output[agent.id]["output"]["content"]) ``` Wrapping the agent in a `Workflow` gives you consistent output handling and callbacks such as tracing. Inside a workflow, the agent's answer is at `result.output[agent.id]["output"]["content"]`. ## Step 4: See what the agent did Every step the agent takes, each model call and each tool call, is recorded by the tracing handler: ```python for run in tracing.runs.values(): print(run.type, run.name, run.status) ``` To see the same run tree in the Dynamiq platform, with inputs, outputs, timings and token usage for every node, replace `TracingCallbackHandler` with `DynamiqTracingCallbackHandler` and a project-scoped access key. The [SDK documentation](https://docs.getdynamiq.ai/docs/sdk/platform-integration/tracing-to-dynamiq) covers the setup. ## What happens during a run? The agent works in a loop until it can answer: 1. **Reason.** It reads the question and the file description and plans a first step, usually inspecting the file. 2. **Act.** It sends Python code to the E2B sandbox, for example `pd.read_csv("/home/user/input/house_prices.csv").describe()`. 3. **Observe.** It reads the printed output or the error message. 4. **Repeat.** It fits the model, checks the results and fixes anything that failed. 5. **Answer.** It writes the markdown report with the numbers it computed. Because the numbers come from code that ran, not from the model's memory, you can check any figure by reading the code in the trace. ## What else can you ask it? Swap the file and the question; the agent stays the same. - "Which customer attributes are most associated with churn in this file? Show a correlation table and a chart." - "Compare monthly transaction volumes by region and flag months more than two standard deviations from the mean." - "Match these two exports on invoice number and list the records that appear in only one of them, with totals." The last one is a typical back-office reconciliation, where an agent prepares the exceptions and a person reviews them. See how teams run that pattern in production in [back-office automation](/use-cases/back-office). ## How do you use it with sensitive data? Check where the sandbox runs before you upload regulated data. E2B runs sandboxes in its cloud by default, and the model provider receives whatever the agent puts in its prompts. For experiments, use synthetic or masked data. For production, run the agent where your data policies allow, screen prompts for personal data, and keep a trace of every step for review. ## How do you take it to production? The agent you built in code runs on the same engine as the visual canvas in [Agent Builder](/product/agents), so a team can build in Python, review on the canvas and deploy the same workflow as an API with versions and rollback. On the platform you can add guardrails that detect PII and prompt injection, require approval before a step runs, score answers with evaluations on test sets and live traffic, and self-host the whole platform in your own cloud or data center. ## FAQ ### What is an AI data analyst agent? It is an AI agent that answers questions about data by planning an analysis, writing code, running it in a sandbox and reporting the results. Unlike a chatbot reading numbers from a prompt, it computes every figure with code you can inspect. ### Why run the agent's code in a sandbox? Model-written code can be wrong or unsafe. A sandbox is an isolated, short-lived machine, so a failed or harmful command cannot touch your own systems, and the agent can install packages and read files freely inside it. ### Can I use a different LLM instead of Together AI? Yes. Replace the `TogetherAI` node with another LLM class, such as `OpenAI`, `Anthropic` or `WatsonX`. The Dynamiq SDK supports 29 model providers, and the agent and tool code stays the same. ### Is the Dynamiq SDK open source? Yes. The Python SDK is open source under the Apache-2.0 license and available on [GitHub](https://github.com/dynamiq-ai/dynamiq). The Dynamiq platform, with its visual builder, deployment, guardrails and observability, is a commercial product. ### Can the agent read Excel or PDF files? Yes. Pass the file the same way, and the agent can install a parsing package such as openpyxl or pdfplumber in the sandbox before reading it. Describe the file's contents in its `description` to save the agent a step. --- # Building a Search GPT with Dynamiq URL: https://www.getdynamiq.ai/post/building-a-search-gpt-with-dynamiq Published: 2024-11-03 Build a Search GPT that rephrases a question, searches the web and writes a cited answer, as a visual workflow called over HTTP or in Python with the SDK. A Search GPT is a question-answering app that searches the web for every query and writes a short answer with numbered citations. You can build one from three steps: an agent rewrites the question into a search query, a search tool fetches results, and a model writes the answer from those results only. This tutorial builds the same pipeline twice, as a visual workflow in Dynamiq that you call over HTTP and as Python code with the open-source SDK. ## How does a Search GPT work? Every question goes through the same three steps: 1. **Rephrase.** Questions such as "what's going on with Apple stock lately?" make poor search queries. An agent turns them into short, keyword-focused queries, such as "Apple stock performance this week". 2. **Search.** A search tool sends the query to a web search API and returns the top results with titles, snippets and URLs. 3. **Answer.** A model writes a concise answer using only those results, cites them as [1], [2] and so on, and lists the sources. Keeping the answer step grounded in the results is what separates a Search GPT from a chatbot answering from memory. If the results don't answer the question, the model should say so. ## Approach 1: Build it visually and call it over HTTP In [Agent Builder](/product/agents), the pipeline is five nodes on the canvas: 1. An **Input** node with an `input` field for the question. 2. An **Agent** that rephrases the question, with the rephrasing instructions below as its role. 3. A **search tool**, such as Scale SERP or Tavily, that takes the rephrased query. 4. An **LLM** node that writes the cited answer, with the answer prompt below. 5. An **Output** node that returns the answer. Test it in the builder first. Each test run is traced, so you can see every node's input and output, how long it took and what it cost. When the answers look right, deploy the workflow as an App. Every App gets its own hostname and accepts a POST request with an Access Key: ```python import os import requests ENDPOINT = "https://" # shown on the App page ACCESS_KEY = os.environ["DYNAMIQ_ACCESS_KEY"] def ask(question: str) -> dict: response = requests.post( ENDPOINT, headers={"Authorization": f"Bearer {ACCESS_KEY}"}, json={"input": {"input": question}, "stream": False}, # keys match your Input node timeout=120, ) response.raise_for_status() return response.json() # your Output node's fields ``` Send `"stream": true` to receive the answer as server-sent events while it is generated. The [HTTP reference for Apps](https://docs.getdynamiq.ai/docs/platform/deployments/call-your-app) covers streaming, async callbacks and error codes. ## Approach 2: Build it in Python with the SDK The same pipeline in code gives you full control over each step. Install the SDK with `pip install dynamiq` and set `OPENAI_API_KEY` and `SERP_API_KEY` (for Scale SERP). First, the instructions for the two model steps. The answer prompt reads the original question from the workflow input and the search results from the search step. ```python REPHRASE_ROLE = """You turn user questions into effective web search queries. Remove filler words, keep the essential keywords and names, and turn questions into short statements. Examples: - "Who was the first person to walk on the moon?" -> "First person to walk on the moon" - "How tall is the Eiffel Tower?" -> "Eiffel Tower height" Return only the rewritten query.""" ANSWER_PROMPT = """Answer the question using only the search results below. Cite sources inline as [1], [2] and so on, matching a numbered source list. If the results do not answer the question, say so and suggest a better query. Question: {{ input }} Search results: {{ search_results }} Return the answer inside tags, followed by a numbered list of sources formatted as markdown links inside tags.""" ``` Then the three nodes, wired into a workflow: ```python import re from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.connections import ScaleSerp as ScaleSerpConnection from dynamiq.flows import Flow from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.tools.scale_serp import ScaleSerpTool from dynamiq.prompts import Message, Prompt connection = OpenAIConnection() rephraser = Agent( id="rephraser", name="rephraser", role=REPHRASE_ROLE, llm=OpenAI(connection=connection, model="gpt-4o-mini", temperature=0.1), ) search = ( ScaleSerpTool( id="search", name="search", connection=ScaleSerpConnection(), # reads SERP_API_KEY limit=5, is_optimized_for_agents=True, # returns results as readable text ) .inputs(query=rephraser.outputs.content) .depends_on(rephraser) ) writer = ( OpenAI( id="writer", name="writer", connection=connection, model="gpt-4o", temperature=0.2, prompt=Prompt(messages=[Message(role="user", content=ANSWER_PROMPT)]), ) .inputs(search_results=search.outputs.content) .depends_on(search) ) workflow = Workflow(flow=Flow(nodes=[rephraser, search, writer])) def extract_tag(text: str, tag: str) -> str | None: match = re.search(rf"<{tag}>(.*?)", text, re.DOTALL) return match.group(1).strip() if match else None def search_gpt(question: str) -> tuple[str | None, str | None]: result = workflow.run(input_data={"input": question}) content = result.output[writer.id]["output"]["content"] return extract_tag(content, "answer"), extract_tag(content, "sources") ``` `.inputs()` maps one node's output into the next node's input, and `.depends_on()` sets the order. The workflow's own input, the `input` question, is available to every node, which is how the answer prompt reads the original question. A complete version of this project, with both backends and the frontend, lives in the SDK repository under [examples/use_cases/search](https://github.com/dynamiq-ai/dynamiq/tree/main/examples/use_cases/search). ## Add a Streamlit frontend A few lines of Streamlit put a search box in front of either backend. Save the code above as `search_backend.py`, then run `streamlit run app.py`: ```python import streamlit as st from search_backend import search_gpt st.title("Search GPT") question = st.text_input("Ask a question") if st.button("Search") and question.strip(): with st.spinner("Searching..."): answer, sources = search_gpt(question) st.markdown(answer or "No answer found. Try rephrasing the question.") if sources: st.markdown("**Sources**") st.markdown(sources) ``` To show the answer as it is written, enable streaming on the writer node and read the chunks with a streaming callback handler; the [streaming guide](https://docs.getdynamiq.ai/docs/sdk/concepts/streaming-and-callbacks) shows the pattern. ## How do you make a Search GPT reliable enough for work? A demo answers most questions well. A tool people rely on needs a few more controls: - **Restrict the sources.** For research in regulated fields, search only sources you trust. `TavilyTool` accepts a list of domains to include, such as regulator and central bank sites. - **Add your own documents.** Combine web search with a [knowledge base](/product/knowledge-rag) of internal documents, so answers can draw on both. - **Check for prompt injection.** Web pages can contain instructions aimed at models. A [guardrail node](/product/guardrails) can check search results for prompt injection before the writer sees them. - **Measure answer quality.** Score answers for faithfulness to the sources with [evaluations](/product/evaluations), and turn bad answers from production traces into test cases. ## FAQ ### What is a Search GPT? A Search GPT is an application that answers questions by searching the web in real time and writing a short answer with citations to the pages it used. It combines a search API with a language model, so answers reflect current information rather than only what the model learned in training. ### Which search API should I use? The Dynamiq SDK includes tools for Tavily, Exa, Scale SERP, Firecrawl and Jina search. They differ in result format, freshness and filtering options, so test two or three on your own questions before you choose. ### Why rephrase the question before searching? Search engines work best with short, keyword-focused queries. Rewriting a conversational question into a search query usually returns more relevant results, which leads to better answers. ### How do you stop the answer from making things up? Tell the model to answer only from the search results, require a citation for each claim, and have it say so when the results don't contain the answer. Then measure faithfulness with evaluations and review low-scoring answers. ### Can a Search GPT search internal documents too? Yes. Add a retrieval tool over a knowledge base of internal documents alongside web search, or instead of it, so the same pipeline answers questions about your own policies, products or research. --- # Agent orchestration patterns for multi-agent systems URL: https://www.getdynamiq.ai/post/agent-orchestration-patterns-in-multi-agent-systems-linear-and-adaptive-approaches-with-dynamiq Published: 2024-10-28 Five agent orchestration patterns for multi-agent systems, from sequential and routing to hierarchical, when to use each, and how to build them in Dynamiq. Multi-agent systems coordinate specialized agents in one of two basic ways. A linear pattern runs them in an order you define, like an assembly line; an adaptive pattern lets a manager agent decide at run time which specialist works next. In Dynamiq, the Graph Orchestrator covers linear and branching flows, and a manager agent that uses other agents as tools covers adaptive delegation. If you used an earlier version of the SDK, you may remember separate `LinearOrchestrator` and `AdaptiveOrchestrator` classes. The current SDK expresses both patterns with the building blocks in this guide, and all code below uses the current API. ## What does an orchestrator do in a multi-agent system? An orchestrator is the part of a [multi-agent system](/glossary/multi-agent-system) that decides who does what, in which order, with which context. It does five jobs: - **Breaks down the task** into steps a single agent can handle. - **Assigns each step** to the agent with the right role and tools. - **Passes context** from one step to the next, without flooding each agent with everything. - **Combines the results** into one answer. - **Enforces limits** such as maximum loops, so a confused agent cannot run forever. In an enterprise setting it has a sixth job: leaving a record. Every decision the orchestrator makes should be visible afterward, which is why every run in Dynamiq is traced. ## What are the common agent orchestration patterns? Most multi-agent systems are built from five orchestration patterns. Production systems usually combine several of them, and add a person at the steps that need approval. | Pattern | How it works | Use it when | In Dynamiq | | --- | --- | --- | --- | | Sequential (linear) | Agents run in a fixed order, each passing its output to the next | Every step must happen, such as a due diligence checklist | Graph Orchestrator states joined by edges, or a workflow | | Routing | A classifier or a rule sends each request to one specialist | Requests fall into clear types, such as billing versus fraud | The Choice node, or conditional edges in the Graph Orchestrator | | Parallel (fan-out and fan-in) | Several agents work at once and a final step merges their results | Checks or sources are independent, such as screening many parties | Independent workflow branches run at the same time; the Map node runs one step per item | | Hierarchical (adaptive) | A manager agent decides which specialist works next until the goal is met | The path cannot be planned in advance, such as open-ended research | A manager agent with the Graph Orchestrator | | Evaluator loop | One agent drafts, another checks the draft against criteria, and it loops back until it passes | Output must meet a quality bar, such as a report or a code change | A conditional edge back to an earlier state, bounded by `max_loops` | The rest of this guide goes deep on the two most common choices, linear and adaptive orchestration, and how to build each. ## What is linear orchestration? Linear orchestration runs agents in a sequence you define. Each agent receives the output of the previous step, and the last agent produces the answer. The path is known before the run starts. This suits processes that must happen the same way every time. A vendor due diligence review, for example, always gathers information first, assesses it against the firm's criteria second and writes the summary last. A bank's compliance team wants that order to be fixed and provable. ## What is adaptive orchestration? Adaptive orchestration gives a manager agent a set of specialist agents and lets it decide, step by step, which one to call next. The manager can call several specialists in parallel, check their work and loop until the task is done. This suits open-ended work where the path depends on what the agents find, such as research questions, investigations or support requests that could need any of several specialists. ## Linear or adaptive: which should you use? | | Linear (graph) | Adaptive (manager-led) | | -------------------- | --------------------------------------- | --------------------------------------------- | | Who decides the path | You, when you design the graph | The manager agent, at run time | | Predictability | High: same steps every run | Lower: the path can change between runs | | Cost and latency | Easier to predict | Varies with how many calls the manager makes | | Auditability | Straightforward to review | Needs traces to follow each decision | | Best for | Reviews, approvals, document processing | Research, investigations, open-ended requests | | Dynamiq SDK | `GraphOrchestrator` with edges | An `Agent` with other agents in `tools` | For regulated processes, start linear and add branches where you need them. Use an adaptive manager inside a well-defined step, not as the outer layer of a process an auditor needs to follow. ## How do you build a linear flow with the Graph Orchestrator? The example below runs a three-step vendor due diligence review. Install the SDK with `pip install dynamiq` and set `OPENAI_API_KEY` and `TAVILY_API_KEY` first. ```python from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.connections import Tavily as TavilyConnection from dynamiq.flows import Flow from dynamiq.nodes.agents import Agent from dynamiq.nodes.agents.orchestrators.graph import END, START, GraphOrchestrator from dynamiq.nodes.agents.orchestrators.graph_manager import GraphAgentManager from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.tools.tavily import TavilyTool llm = OpenAI(connection=OpenAIConnection(), model="gpt-4o", temperature=0.1) search = TavilyTool(connection=TavilyConnection()) # reads TAVILY_API_KEY researcher = Agent( name="Researcher", role=( "Find recent public information about the company: ownership, sanctions news, " "litigation and major events. Cite every source." ), llm=llm, tools=[search], max_loops=8, ) risk_analyst = Agent( name="Risk analyst", role=( "Assess the research against the firm's vendor risk criteria. Rate each risk as " "low, medium or high with a reason. End with one line: RISK: LOW, RISK: MEDIUM or RISK: HIGH." ), llm=llm, ) writer = Agent( name="Writer", role="Write a one-page due diligence summary for a compliance reviewer, with open questions at the end.", llm=llm, ) orchestrator = GraphOrchestrator(name="vendor-due-diligence", manager=GraphAgentManager(llm=llm)) orchestrator.add_state_by_tasks("research", [researcher]) orchestrator.add_state_by_tasks("assess", [risk_analyst]) orchestrator.add_state_by_tasks("write", [writer]) orchestrator.add_edge(START, "research") orchestrator.add_edge("research", "assess") orchestrator.add_edge("assess", "write") orchestrator.add_edge("write", END) workflow = Workflow(flow=Flow(nodes=[orchestrator])) result = workflow.run(input_data={"input": "Run vendor due diligence on Example Logistics Ltd."}) print(result.output[orchestrator.id]["output"]["content"]) ``` Each state holds one or more tasks, and edges fix the order. The `GraphAgentManager` turns the conversation so far into a clear task for the agent in each state, so the analyst sees what the researcher found. ## How do you add a branch, such as escalation to a person? Real processes are rarely a straight line. A conditional edge lets a Python function pick the next state from the shared context. Here, a high-risk rating skips the automated summary and routes the case to a compliance analyst. Add these lines before you run the workflow; the conditional edge replaces the fixed edge from `assess` to `write`. ```python def route_by_risk(context: dict, **kwargs) -> str: """Send high-risk vendors to a person instead of the writer.""" last_message = context.get("history", [{}])[-1].get("content", "") return "escalate" if "RISK: HIGH" in last_message else "write" def escalate(context: dict, **kwargs) -> dict: """Record the escalation; your case system would pick it up here.""" return {"result": "Escalated to a compliance analyst for manual review.", "escalated": True} orchestrator.add_state_by_tasks("escalate", [escalate]) orchestrator.add_conditional_edge("assess", ["write", "escalate"], route_by_risk) orchestrator.add_edge("escalate", END) ``` A plain Python function can be a state too. It receives the shared context, and the keys it returns, apart from `result`, are merged back into that context for later states. ## How do you build adaptive orchestration with a manager agent? For adaptive work, pass specialist agents in a manager agent's `tools` list. The manager reads each specialist's `description`, decides whom to call and with what task, and assembles the answer. ```python from dynamiq.nodes.types import Behavior, InferenceMode research_analyst = Agent( name="Research analyst", description="Finds recent, reputable public sources on a topic and returns cited highlights.", role="Find recent, reputable sources on the topic you are given and return cited highlights.", llm=llm, tools=[search], inference_mode=InferenceMode.XML, max_loops=6, behaviour_on_max_loops=Behavior.RETURN, ) brief_writer = Agent( name="Brief writer", description="Turns research highlights into a short brief for a named audience.", role="Turn research highlights into a concise brief. Keep every citation.", llm=llm, inference_mode=InferenceMode.XML, max_loops=4, behaviour_on_max_loops=Behavior.RETURN, ) manager = Agent( name="Manager", role=( "Plan the work, delegate research and writing to your sub-agents and check their output. " "Call sub-agents with {'input': ''} payloads and assemble the final brief." ), llm=llm, tools=[research_analyst, brief_writer], inference_mode=InferenceMode.XML, parallel_tool_calls_enabled=True, max_loops=10, behaviour_on_max_loops=Behavior.RETURN, ) workflow = Workflow(flow=Flow(nodes=[manager])) result = workflow.run( input_data={"input": "Brief the risk committee on this month's regulatory news about AI in EU banking."} ) print(result.output[manager.id]["output"]["content"]) ``` `parallel_tool_calls_enabled=True` lets the manager fan out independent subtasks, such as researching two regulators at once. `behaviour_on_max_loops=Behavior.RETURN` makes an agent return its best answer instead of raising an error when it hits its loop limit. The [orchestrators documentation](https://docs.getdynamiq.ai/docs/sdk/agents/orchestrators) covers the full API, and the SDK repository includes complete graph examples, from a code assistant to a customer service concierge. ## How do you see what the orchestrator did? Multi-agent runs are only as trustworthy as your ability to replay them. Every run on the Dynamiq platform is traced: each manager decision, agent step, tool call and model call, with inputs, outputs, cost and latency. SDK workflows can send the same traces to the platform with a callback handler. On top of the traces, an AI agent reviews production runs and surfaces failures, regressions and other issues for your team to check. See [observability](/product/observability) for how that works. ## Can you build these patterns without code? Yes. [Agent Builder](/product/agents) offers the same building blocks on a visual canvas, including a Graph Orchestrator and agents that call other agents. The canvas and the SDK run on the same engine, so a team can start in one and continue in the other. In regulated workflows, pair either pattern with approvals: an approval pauses any tool step for a person, who can edit its fields before it runs. Our guide to [multi-agent systems](/post/multi-agent-ai-systems-definition-benefits-limitations-how-to-build) covers the design trade-offs in more depth. ## FAQ ### What is agent orchestration? Agent orchestration is the coordination layer of a multi-agent system. It decides which agent handles each step, passes context between them, combines their results and enforces limits, such as how many steps a run may take. ### What is the difference between linear and adaptive orchestration? In linear orchestration, you define the order of the steps before the run, and every run follows it. In adaptive orchestration, a manager agent decides at run time which specialist to call next, so the path can change from run to run. ### What happened to LinearOrchestrator and AdaptiveOrchestrator in Dynamiq? The current SDK replaces them with two more general building blocks. Linear and branching flows use `GraphOrchestrator`, and adaptive delegation uses an `Agent` that lists other agents in its `tools`. The concepts in this guide carry over; only the class names changed. ### Can you combine linear and adaptive orchestration? Yes, and it is often the best design. Use a graph for the outer process so the order of steps is fixed and auditable, and put an adaptive manager agent inside a single state where the work is open-ended, such as research. ### How do you keep a multi-agent system auditable? Trace every run, keep the outer process as a defined graph, route high-risk cases to a person with a conditional edge or an approval, and turn failed runs into evaluation cases so the same mistake is caught next time. --- # Automating customer support in banking with agentic RAG URL: https://www.getdynamiq.ai/post/automating-customer-support-in-banking-using-agentic-rag Published: 2024-10-21 How agentic RAG automates bank customer support: one agent retrieves your procedures, another acts through your APIs, with approvals and traces. Agentic RAG automates banking support by giving an AI agent a search tool over your own procedures instead of a fixed retrieval step. The agent decides what to look up, searches again when the results don't fit, and hands a cited procedure to a second agent that completes the request through your banking APIs. Approvals, PII detection and a trace of every run keep the whole thing within what a bank can explain to an auditor. ## What is agentic RAG? [Retrieval-augmented generation](/glossary/retrieval-augmented-generation) (RAG) grounds a model's answer in documents you control. The system embeds the question, retrieves the closest passages from a vector store and asks the model to answer from them. Classic RAG runs that pipeline once, in a fixed order. Agentic RAG turns retrieval into a tool that an agent calls when it needs to, as often as it needs to. That matters in banking, where a customer's words ("I was charged twice at a hotel abroad") rarely match the title of the procedure that answers them. | | Classic RAG | Agentic RAG | | --------------------- | ------------------------------- | -------------------------------------------------- | | Who decides to search | The pipeline, once per question | The agent, as often as the task needs | | Weak results | Answers from whatever came back | Rephrases and searches again, or asks the customer | | Output | An answer | An answer, or an action through a tool | | Best for | FAQ-style questions | Requests that need lookup, reasoning and an action | ## Why does banking support need more than a chatbot? Three things make bank support different from a generic help desk: - **Procedures change often.** Fees, dispute windows and card controls get updated, so answers must come from the current document, not from what a model remembers. - **Many requests end in an action.** Blocking a card, opening a dispute or reissuing a statement needs tools, and the bank needs control over which actions run without a person. - **Everything is reviewable.** Supervisors and auditors expect to see what the agent read, what it did and why. Support automation of this kind is already in production. For one example, read how [an Asian neo-bank automated about 85% of support inquiries and went live in 30 days](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai). ## How does the two-agent design work? The workflow in this tutorial splits the job between two agents: 1. **The procedures agent** receives the customer's request, searches your internal procedures and returns the steps that apply, plus the details the bank needs from the customer. 2. **The action agent** receives the request and that procedure, asks the customer for anything missing, and calls the banking API for the steps it is allowed to take. 3. **The workflow** returns a summary of what was done, and the full run is traced. Splitting the roles keeps each agent's instructions short and its permissions narrow. The procedures agent can read but not act; the action agent can act only through the tools you give it. ## How do you build it with the Dynamiq SDK? The code below uses the [open-source Dynamiq SDK](/open-source) (Apache-2.0), the same engine that runs the visual Agent Builder. It assumes your procedures are already indexed in a Pinecone index called `bank-procedures`. The [RAG pipeline guide](https://docs.getdynamiq.ai/docs/sdk/rag/rag-pipeline) shows the indexing flow. The SDK supports 8 vector stores, and swapping Pinecone for another one changes a single node. ### Install and configure ```bash pip install dynamiq export OPENAI_API_KEY="..." export PINECONE_API_KEY="..." ``` ### Step 1: Give the agent a search tool over your procedures `VectorStoreRetriever` bundles a query embedder and a retriever into one tool the agent can call. Use the same embedding model you indexed with. ```python from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.connections import Pinecone as PineconeConnection from dynamiq.nodes.embedders import OpenAITextEmbedder from dynamiq.nodes.llms import OpenAI from dynamiq.nodes.retrievers import PineconeDocumentRetriever from dynamiq.nodes.retrievers.retriever import VectorStoreRetriever llm = OpenAI( connection=OpenAIConnection(), # reads OPENAI_API_KEY model="gpt-4o-mini", temperature=0.1, ) procedures_search = VectorStoreRetriever( name="procedures-search", description="Searches the bank's internal support procedures. Input: a short search query.", text_embedder=OpenAITextEmbedder( connection=OpenAIConnection(), model="text-embedding-3-small", # the model used at indexing time ), document_retriever=PineconeDocumentRetriever( connection=PineconeConnection(), # reads PINECONE_API_KEY index_name="bank-procedures", top_k=5, ), ) ``` ### Step 2: Create the procedures agent The agent's role tells it to search, retry with a better query when results miss, and say so when no procedure applies. ```python from dynamiq.nodes.agents import Agent procedures_agent = Agent( name="Procedures agent", role=( "You look up the bank's internal procedure for a customer request. " "Search the procedures, rephrase the query if the results do not fit, " "and return the steps to follow and the details needed from the customer. " "Name the procedure you used. If nothing applies, say so." ), llm=llm, tools=[procedures_search], max_loops=6, ) ``` ### Step 3: Give the action agent its tools The action agent gets two tools: an HTTP tool for your internal API and a feedback tool for asking the customer a question. ```python from dynamiq.connections import Http as HttpConnection from dynamiq.nodes.tools.http_api_call import HttpApiCall from dynamiq.nodes.tools.human_feedback import HumanFeedbackTool BANK_API = "http://127.0.0.1:8000" # your internal API gateway bank_api = HttpApiCall( name="bank-api", connection=HttpConnection(method="POST", url=BANK_API), description=( "Internal bank API for the signed-in customer. Call an endpoint by its full URL " "and send parameters in `data`:\n" f"- {BANK_API}/cards/block (card_last4, reason)\n" f"- {BANK_API}/statements/request (account_last4, month)" ), ) ask_customer = HumanFeedbackTool() # asks in the console; connect your chat channel in production ``` The tool description is the agent's only documentation for your API, so list every endpoint and its parameters. Notice what is not on the list: nothing moves money. Keep those endpoints out of the agent's reach, or put an approval in front of them. ### Step 4: Chain the agents in a workflow The action agent runs after the procedures agent and receives both the original request and the procedure it returned. ```python from dynamiq import Workflow from dynamiq.flows import Flow def with_procedure(inputs: dict, outputs: dict) -> str: return ( f"Customer request: {inputs['input']}\n\n" f"Procedure to follow:\n{outputs[procedures_agent.id]['content']}" ) action_agent = ( Agent( name="Action agent", role=( "You resolve the customer's request by following the procedure you are given. " "Ask the customer for a missing detail with the feedback tool. " "Never ask for a PIN, password or one-time code. " "Finish with a short summary of what you did and what happens next." ), llm=llm, tools=[bank_api, ask_customer], max_loops=10, ) .inputs(input=with_procedure) .depends_on(procedures_agent) ) workflow = Workflow(flow=Flow(nodes=[procedures_agent, action_agent])) session_token = "..." # from the customer's authenticated app session result = workflow.run( input_data={ "input": "I lost my debit card this morning. Please block it.", "tool_params": { "by_name": {"bank-api": {"headers": {"Authorization": f"Bearer {session_token}"}}}, }, } ) print(result.output[action_agent.id]["output"]["content"]) ``` `tool_params` passes the customer's session token to the bank API at run time. The model never sees it, so it cannot repeat it in a reply, and the agent never has to ask the customer for credentials. A runnable version with a mock bank API lives in the SDK repository under [examples/use_cases/customer_support](https://github.com/dynamiq-ai/dynamiq/tree/main/examples/use_cases/customer_support). Its mock checks a PIN to stay self-contained. In production, authenticate the customer in your app and never let a model collect a PIN or password. ## What does it take to run this in a bank? The two agents are the easy part. These controls are what make the workflow acceptable to risk and compliance teams: - **Keep data in your perimeter.** Run Dynamiq in our cloud or self-hosted in your own cloud account or data center; see [how we deploy](/deployment). - **Detect PII and prompt injection** before a request reaches a model, with [guardrail nodes](/product/guardrails) in the workflow. - **Put approvals on actions that matter.** An approval pauses a step for a person, who can edit its fields before it runs. - **Retrieve with permissions.** [Knowledge bases](/product/knowledge-rag) apply permission-aware retrieval, so an agent sees only what the asking user is allowed to see. - **Trace and evaluate.** Every run is traced with its inputs, outputs, cost and latency. Turn bad runs into test cases and attach [real-time evals](/product/evaluations) to the live deployment. For the full picture of support automation across chat and inbound calls, see [customer service agents](/use-cases/customer-service). Banks, lenders and insurers can find the rest of the platform, from document review to compliance, on the [financial services page](/industries/financial-services). ## FAQ ### What is the difference between RAG and agentic RAG? Classic RAG retrieves once and answers from whatever came back. Agentic RAG gives retrieval to an agent as a tool, so the agent decides when to search, rephrases weak queries and combines retrieval with other tools, such as an API call or a question to the customer. ### Can an AI agent block a card or move money on its own? Technically yes, if you give it the tool. In practice, banks let agents take low-risk, reversible actions such as blocking a card, and route anything that moves money or changes an account through an approval, where a person reviews and can edit the exact request before it runs. ### How do you keep customer data safe when an LLM handles bank support? Keep processing inside infrastructure you control, detect PII and prompt injection before text reaches a model, inject credentials at run time instead of asking customers for them, and scope retrieval to what each user may see. A trace of every run gives your auditors the evidence. ### Do you need to retrain the model when procedures change? No. The agent answers from the retrieved procedure, not from training, so you update the document and re-index it. That also gives compliance teams one place to review what the agent is told to do. ### Which LLM and vector database should I use? Use the model and store your bank has already approved. The SDK supports 29 model providers and 8 vector stores behind the same interfaces, so changing either one means swapping a single node, not rewriting the workflow. --- # Building a conversational AI agent with memory in Dynamiq URL: https://www.getdynamiq.ai/post/building-a-conversational-ai-agent-with-dynamiq Published: 2024-10-10 Build a conversational AI agent that remembers: session memory scoped by user and session, plus long-term facts the agent saves and recalls itself. To build a conversational AI agent that remembers, give it two kinds of memory: session memory for the current conversation and long-term memory for durable facts about the user. In the Dynamiq SDK you attach a `Memory` object for the first and a `LongTermMemoryConfig` for the second, then pass a `user_id` and `session_id` on every run. The agent loads the relevant history before it answers and decides on its own which facts are worth keeping. ## Why does a conversational agent need memory? Language models are stateless: each call sees only what is in its prompt. Without memory, an agent forgets the customer's name between two messages and asks for the same details twice. Memory fixes three things: - **Continuity.** The agent follows a conversation across turns, so "and what about the second one?" makes sense. - **Personalization.** The agent adapts to a user's stated preferences, such as language, format or channel. - **Efficiency.** Users don't repeat themselves, and conversations get shorter. In regulated settings, memory also needs limits. A bank's support assistant should remember that a customer prefers Spanish and reported a lost card this morning. It has no business remembering the card number. ## What kinds of memory can an agent have? The SDK separates memory into two stores that you can use together or on their own. | | Session memory | Long-term memory | | ------------------------ | --------------------------------------- | --------------------------------------------------------- | | What it stores | The messages of a conversation | Durable facts about a user | | Scope | `user_id` and `session_id` | `user_id`, across every session | | Who decides what to keep | Saved automatically on every turn | The agent, through remember and recall tools | | Retrieval | Recent messages, relevant ones, or both | Semantic search over facts, with duplicate results merged | | Configured with | `memory=Memory(...)` | `long_term_memory=LongTermMemoryConfig(...)` | Session memory works with 8 memory backends, from in-process storage for tests to PostgreSQL, DynamoDB and vector stores such as Qdrant, Pinecone and Weaviate. The [agent memory glossary entry](/glossary/agent-memory) covers the concepts in more depth. ## Step 1: Install the SDK The SDK is [open source](/open-source) under the Apache-2.0 license and runs on Python 3.10 or later. ```bash pip install dynamiq export OPENAI_API_KEY="..." ``` ## Step 2: Add session memory Attach a `Memory` instance to the agent and pass the same `user_id` and `session_id` on every turn of a conversation. The agent loads prior messages for that pair before it reasons, and saves the new turn afterward. ```python from dynamiq import Workflow from dynamiq.connections import OpenAI as OpenAIConnection from dynamiq.flows import Flow from dynamiq.memory import Memory, MemorySaveMode from dynamiq.memory.backends import InMemory from dynamiq.nodes.agents import Agent from dynamiq.nodes.llms import OpenAI llm = OpenAI(connection=OpenAIConnection(), model="gpt-4o-mini", temperature=0.2) agent = Agent( name="assistant", id="assistant", llm=llm, role="Friendly, concise assistant. Use earlier turns of the conversation when they help.", memory=Memory(backend=InMemory(), save_mode=MemorySaveMode.INPUT_OUTPUT), ) workflow = Workflow(flow=Flow(nodes=[agent])) conversation = {"user_id": "user-42", "session_id": "session-1"} workflow.run(input_data={"input": "Hi, I'm Dana. I work on the payments team.", **conversation}) result = workflow.run(input_data={"input": "Which team am I on?", **conversation}) print(result.output[agent.id]["output"]["content"]) # mentions the payments team ``` `MemorySaveMode.INPUT_OUTPUT` stores only the user's message and the final answer, which keeps the next prompt small. The default, `FULL`, also stores tool calls and observations, which is useful when you need to replay exactly what happened. ## Step 3: Use a persistent backend and pick a retrieval strategy `InMemory` disappears when the process stops. For a real assistant, switch the backend; nothing else in the agent changes. Vector-store backends such as Qdrant embed each message, which lets the agent retrieve by relevance, not only by recency. ```python from dynamiq.connections import Qdrant as QdrantConnection from dynamiq.memory import MemoryRetrievalStrategy from dynamiq.memory.backends import Qdrant from dynamiq.nodes.embedders import OpenAIDocumentEmbedder memory = Memory( backend=Qdrant( connection=QdrantConnection(), # reads QDRANT_URL and QDRANT_API_KEY embedder=OpenAIDocumentEmbedder(connection=OpenAIConnection(), model="text-embedding-3-small"), index_name="conversations", dimension=1536, # must match the embedding model ), save_mode=MemorySaveMode.INPUT_OUTPUT, ) agent = Agent( name="assistant", id="assistant", llm=llm, role="Friendly, concise assistant. Use earlier turns of the conversation when they help.", memory=memory, memory_limit=50, memory_retrieval_strategy=MemoryRetrievalStrategy.BOTH, ) ``` The retrieval strategy decides which messages reach the prompt: | Strategy | Loads | Good for | | --------------- | ----------------------------------------------- | ------------------------------------ | | `ALL` (default) | The most recent messages, up to `memory_limit` | Short conversations | | `RELEVANT` | Messages that match the current question | Long threads with many topics | | `BOTH` | Recent messages merged with relevant older ones | Support conversations that span days | ## Step 4: Add long-term memory Session memory is a transcript. Long-term memory is a small set of facts that should outlive any single conversation, such as a preferred language or a role. When you enable it, the agent gets two tools, `remember_fact` and `recall_facts`, and decides during its reasoning loop when to use them. When a new fact says nearly the same thing as a stored one, the agent updates the stored fact instead of saving a near duplicate, and `recall_facts` merges the results of several phrasings without duplicates. ```python from dynamiq.memory.long_term import LongTermMemoryConfig from dynamiq.memory.long_term.backends import QdrantLongTermMemoryBackend from dynamiq.nodes.embedders import OpenAITextEmbedder facts = QdrantLongTermMemoryBackend( connection=QdrantConnection(), embedder=OpenAITextEmbedder(connection=OpenAIConnection(), model="text-embedding-3-small"), collection_name="user_facts", dimension=1536, ) agent = Agent( name="assistant", id="assistant", llm=llm, role=( "Friendly, concise assistant. Save durable facts the user shares about their " "preferences and role, and recall them when they help. Never save account numbers, " "passwords or other secrets." ), memory=memory, # the conversation so far long_term_memory=LongTermMemoryConfig(backend=facts), # facts across conversations ) ``` Long-term memory is scoped to the user, so every run must include a `user_id`; the agent raises an error if it is missing. Facts saved in one session are available in the next one, while the session transcript stays separate. ## Step 5: Put it in a chat loop Each new conversation gets a fresh `session_id`. Long-term facts carry over because the `user_id` stays the same. ```python import uuid workflow = Workflow(flow=Flow(nodes=[agent])) def chat(user_id: str) -> None: session_id = str(uuid.uuid4()) # a new conversation print("Type 'exit' to quit.") while (text := input("You: ").strip()).lower() != "exit": result = workflow.run(input_data={"input": text, "user_id": user_id, "session_id": session_id}) print("Agent:", result.output[agent.id]["output"]["content"]) chat("user-42") ``` Tell the agent "I prefer answers in Spanish" in one session, then start a new one and ask a question: the agent can recall the preference without being told again. ## How do you inspect and delete what an agent remembers? Both stores are also usable directly, which is how you build an admin view or honor a deletion request. ```python from dynamiq.prompts import MessageRole # Session memory memory.add( role=MessageRole.USER, content="Please contact me by email, not phone.", metadata={"user_id": "user-42", "session_id": "session-1"}, ) matches = memory.search(query="contact preference", filters={"user_id": "user-42"}, limit=5) recent = memory.get_all(limit=20) memory.delete(user_id="user-42") # every session for this user # Long-term facts print(facts.list_all(user_id="user-42")) facts.clear_user(user_id="user-42") # returns the number of facts deleted ``` Scoped deletion matters under data protection law. If a customer asks you to erase their data, you need to find it and remove it everywhere the agent stored it; our guide to [LLMs under GDPR](/post/balancing-innovation-and-privacy-llms-under-gdpr) covers the wider picture. ## What should a production agent remember? Memory is where an assistant becomes useful, and also where it can quietly accumulate data you never meant to keep. A few rules keep it safe: - **Remember preferences, not secrets.** Put the boundary in the agent's role and check it in evaluations. - **Flag personal data before it reaches the model.** [Guardrail nodes](/product/guardrails) detect PII and prompt injection at the start of a workflow, so you decide what happens to flagged input. - **Scope everything by user.** Never share a memory store across users without the `user_id` filter. - **Keep a record.** Every run is traced, so you can see which facts an agent saved and when. To put the agent in front of users, wrap the workflow in a small HTTP server and deploy it to Dynamiq as a service with the `dynamiq` CLI, or build the same agent on the [Agent Builder](/product/agents) canvas, which runs on the same engine. Either way, the memory stores can live in your own infrastructure. ## FAQ ### What is the difference between short-term and long-term memory in AI agents? Short-term (session) memory is the transcript of the current conversation, scoped to a user and a session. Long-term memory is a separate store of durable facts about a user that persists across sessions. In the Dynamiq SDK, the agent decides what goes into long-term memory through its remember and recall tools. ### Where does a Dynamiq agent store its memory? Wherever you choose. The SDK supports 8 memory backends for session memory, including in-process storage, SQLite, PostgreSQL, DynamoDB and vector stores. Long-term facts can live in PostgreSQL with pgvector, Pinecone, Qdrant, Weaviate or an in-memory store for tests. ### Does long-term memory require a user ID? Yes. Every fact is scoped to a user, so runs with long-term memory enabled must pass a `user_id`. The agent raises an error without one, which prevents facts from being saved without an owner. ### How do I delete a user's memory for a GDPR request? Call `memory.delete(user_id=...)` to remove the user's session history and `clear_user(user_id=...)` on the long-term backend to remove their facts. Also check traces and any other system where you copied conversation data. ### Can the agent remember too much? It can, if you let it. Tell the agent in its role which kinds of facts to save and which to avoid, keep the save mode lean with `INPUT_OUTPUT`, and review saved facts during evaluations. --- # Dynamiq now deployable on IBM Cloud: simplify, scale and secure your AI projects URL: https://www.getdynamiq.ai/post/dynamiq-now-deployable-on-ibm-cloud-simplify-scale-and-secure-your-ai-projects Published: 2024-10-07 Dynamiq runs self-hosted on IBM Cloud Kubernetes Service or Red Hat OpenShift, so agents, knowledge bases and traces stay in an IBM Cloud account you control. Yes, Dynamiq can be self-hosted on IBM Cloud. You install it with a Helm chart on IBM Cloud Kubernetes Service, or on Red Hat OpenShift on IBM Cloud, backed by IBM Cloud Databases for PostgreSQL and IBM Cloud Object Storage. Agents, workflows, knowledge bases and traces then stay in an IBM Cloud account and region your organization controls, which suits banks, insurers, healthcare providers and public agencies with data residency rules. We first announced IBM Cloud support in October 2024. This post covers what the deployment looks like today. ## Why run Dynamiq on IBM Cloud? Three reasons come up most often: - **Data residency and control.** You choose the IBM Cloud region and own the account, the network and the keys. Customer data, documents and run history stay there. - **Proximity to IBM watsonx.** Agents can use IBM watsonx models, including the Granite family, and watsonx embedding models, alongside models from other providers. - **Existing IBM relationships.** Organizations that already run on IBM Cloud or OpenShift can add an agent platform without a new cloud vendor, and Dynamiq is available through IBM's marketplace. Dynamiq is an IBM AI Technology Partner; the [IBM partnership page](/partners/ibm) covers the models, data services and agent catalog work in more detail. ## What runs in your IBM Cloud account? The Helm chart deploys the Dynamiq services into your cluster: the web application and its API, deployed Apps, model inference endpoints, knowledge bases, managed databases and service deployments. You provide the supporting services, which on IBM Cloud map to managed offerings: | Dynamiq needs | On IBM Cloud | | ---------------------------------------- | ---------------------------------------------------------------------- | | Kubernetes 1.32+ | IBM Cloud Kubernetes Service on VPC, or Red Hat OpenShift on IBM Cloud | | PostgreSQL 16+ (three logical databases) | IBM Cloud Databases for PostgreSQL | | S3-compatible object storage | IBM Cloud Object Storage, through its S3-compatible endpoint | | NATS with JetStream | Installed in the cluster from the upstream NATS chart | | DNS and wildcard TLS | IBM Cloud Internet Services or your DNS provider, with cert-manager | Model calls go to the providers you configure, such as IBM watsonx or models you serve yourself. If your environment must make no outbound calls at all, fully isolated and air-gapped setups are delivered with our engineers; see [where Dynamiq runs](/deployment/self-hosted). ## How does the installation work? The installation follows the same Helm flow as any Kubernetes cluster, with IBM Cloud specifics for storage and ingress: 1. **Create the cluster:** an IBM Cloud Kubernetes Service cluster on VPC running Kubernetes 1.32 or newer, or a managed OpenShift cluster. 2. **Provision PostgreSQL:** an IBM Cloud Databases for PostgreSQL deployment at version 16 with three logical databases, using TLS for every connection. 3. **Create object storage:** a Cloud Object Storage bucket with HMAC credentials, reached through its regional S3-compatible endpoint. 4. **Run NATS with JetStream** in the cluster. 5. **Set up DNS and certificates** for the platform's hostnames, with wildcard TLS issued through a DNS-01 challenge. 6. **Install the Helm chart** and run the database migrations. The [IBM Cloud installation guide](https://docs.getdynamiq.ai/docs/platform/self-hosted/install-ibm-cloud) walks through every step. ## What can you build once it is running? Once the platform is running, teams can: - **Build agents** on a visual canvas or in Python with the open-source SDK, on one engine; see [Agent Builder](/product/agents). - **Ground them in your documents** with knowledge bases that apply permission-aware retrieval; see [Knowledge](/product/knowledge-rag). - **Govern what they do** with guardrails that detect PII and prompt injection and approvals that pause any tool step for a person; see [guardrails and approvals](/product/guardrails). - **Measure quality** with evaluations, including real-time evals on live deployments; see [Evals](/product/evaluations). - **See every run** as a trace with cost and latency; see [observability](/product/observability). - **Run and fine-tune open models** with managed inference and LoRA adapters; see [models](/product/models). Common first projects include [customer service agents](/use-cases/customer-service), research and knowledge assistants, and back-office automation. ## Who is this deployment for? Self-hosting on IBM Cloud fits organizations that must show a regulator or an auditor where their data lives and who can reach it. That includes [financial services](/industries/financial-services), healthcare and the public sector. Dynamiq has SOC 2, HIPAA and GDPR in place, and SSO is available on the Enterprise plan. For teams that want help, our engineers deploy the platform and the first agents with you and then hand over; see [how we deploy](/deployment). ## FAQ ### Can Dynamiq run on IBM Cloud? Yes. Dynamiq can be self-hosted on IBM Cloud Kubernetes Service or on Red Hat OpenShift on IBM Cloud, using IBM Cloud Databases for PostgreSQL and IBM Cloud Object Storage for its data. ### Does Dynamiq support Red Hat OpenShift? Yes. Dynamiq installs on Red Hat OpenShift, including managed OpenShift on IBM Cloud, with OpenShift-specific settings for routes and security contexts. ### What infrastructure does a self-hosted install need? Kubernetes 1.32+, a PostgreSQL 16+ server with three logical databases, S3-compatible object storage, NATS with JetStream, and DNS with wildcard TLS certificates for the platform's hostnames. ### Can we use IBM watsonx models with Dynamiq? Yes. Dynamiq agents can call IBM watsonx models, including the Granite family, and use watsonx embedding models, next to models from other providers in the same workflow. ### Can Dynamiq run fully isolated or air-gapped? Fully isolated and air-gapped environments are delivered with our engineers, who plan the setup with your infrastructure and security teams. --- # A comprehensive guide to transforming your business with multimodal AI URL: https://www.getdynamiq.ai/post/a-comprehensive-guide-to-transforming-your-business-with-multimodal-ai Published: 2024-09-11 What multimodal AI is, how models combine text, images, audio and video, where businesses use it, and how to adopt it safely in regulated industries. Multimodal AI is AI that understands, and often produces, more than one type of data, such as text, images, audio and video, within a single model. Instead of chaining a speech recognizer, an OCR engine and a text model, a multimodal model can read a scanned contract, listen to a customer's voice message and answer questions about both. For businesses, that makes it practical to automate work that involves documents, screens and conversations, not just typed text. ## What is multimodal AI? A modality is a type of data: text, images, audio, video or structured numbers. Traditional AI models handle one modality each. A multimodal model accepts several at once and reasons across them. | Aspect | Single-modality AI | Multimodal AI | | ------------ | ---------------------------------------------------- | --------------------------------------------------------------- | | Data handled | One type at a time, such as only text or only images | Several types together, such as text, images and audio | | Context | Understands a situation from one kind of data | Builds a fuller picture by combining data types | | Results | Based on one source | Cross-checked across sources | | Best for | Simple tasks with one input type | Tasks that mix documents, images, speech or video | | Example | Classifying the sentiment of survey responses | Checking an insurance claim's photos against the policy wording | Take an insurance claim. The file holds a claim form, photos of the damage, a PDF of the policy and a recorded call with the customer. A multimodal model can read all four and flag where the photos don't match the description, work that used to need a person to switch between tools. ## How does multimodal AI work? Most multimodal models follow the same pattern, even when the details differ: 1. **Encode.** A dedicated encoder turns each input into a numerical representation: a vision encoder for images and document pages, an audio encoder for speech. 2. **Align.** Training on paired data, such as images with captions or audio with transcripts, teaches the model to place related inputs close together, so a photo of a dented bumper and the words "dented bumper" mean the same thing to it. 3. **Fuse.** The model combines the representations and uses attention to weigh the parts that matter for the question, such as the damaged area in a photo rather than the parking lot around it. 4. **Generate.** A language model produces the output: an answer, extracted fields, a summary or, in models that support it, speech or images. Some models combine separate encoders with a language model; others are trained on mixed data from the start. For a business user, the difference shows up in quality and cost, not in how you use the model. Very few enterprises train multimodal models from scratch. The usual path is to start from a foundation model, ground it in your own documents with retrieval, and fine-tune only when evaluation shows a gap the other steps can't close. ## What types of multimodal models are there? | Type | Inputs and outputs | Typical business uses | | ----------------------- | ----------------------------------------------- | ------------------------------------------------------ | | Vision-language models | Text with images, document pages or screenshots | Document extraction, visual inspection, screen reading | | Speech and audio models | Speech in, text out, or text in, speech out | Call transcription, voice agents, meeting summaries | | Video understanding | Video frames and audio, with text | Training content, quality control, compliance review | | Omni or any-to-any | Text, images and audio in and out | Assistants that see, listen and talk in one exchange | The right type depends on the data your process already produces. A bank's back office mostly needs vision-language models for documents; a contact center needs speech. ## Where do businesses use multimodal AI? The strongest use cases are the ones where people currently read, look at or listen to something before they type into a system: - **Document processing.** Extract fields from scanned forms, bank statements, invoices and contracts, including tables and charts, and check them against your rules. See [document review agents](/use-cases/document-review). - **Customer support.** Understand a customer's screenshot, photo or voice message along with their account history, and resolve more requests without a handoff. See [customer service agents](/use-cases/customer-service). - **Insurance and claims.** Compare damage photos with the claim description and the policy, and route mismatches to an adjuster. - **Healthcare administration.** Process referral documents and intake forms that mix typed text, handwriting and images, with clinicians reviewing anything flagged. - **Inspection and quality control.** Combine camera images with sensor readings to spot defects on a production line. - **Computer use.** Agents that read screens can operate web applications without an API, for example to collect data from a supplier portal. ## What are the benefits of multimodal AI? - **Fewer handoffs.** One model replaces a chain of OCR, speech-to-text and text models, with fewer points of failure. - **Richer context.** Each data type can confirm or contradict the others, which catches errors a single source would miss. - **More natural interaction.** Customers and employees can show and say what they mean instead of describing it in text. - **New automation.** Work that depended on reading documents or looking at images becomes automatable for the first time. ## What are the challenges, and how do you manage them? | Challenge | Why it matters | How to manage it | | --------------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------------ | | Accuracy on your data | Models can misread handwriting, stamps and low-quality scans | Test on a representative sample of your own files before launch | | Personal and biometric data | Images and audio carry faces, ID numbers and voices | Keep processing in your perimeter and detect PII in extracted text | | Cost and latency | Images, audio and video use far more tokens than text | Route simple tasks to smaller models and compress inputs | | Explainability | Reviewers need to see why a field was extracted or a claim flagged | Keep traces and link each answer to its source page or frame | Regulation adds weight to the second row. Under the GDPR, biometric data used to identify a person is a special category of data, and sector rules such as HIPAA apply to images and recordings as much as to text. ## How do you get started with multimodal AI? 1. **Pick one workflow** where people spend time reading documents or listening to calls, and the outcome is easy to check. 2. **Collect a representative sample** of real inputs, including the messy ones: poor scans, handwriting, background noise. 3. **Compare two or three models** on that sample, scoring accuracy, latency and cost. 4. **Build the workflow with a review step,** so a person approves low-confidence results while you learn where the model fails. 5. **Measure in production** and turn failures into test cases before you expand to the next workflow. ## How does Dynamiq support multimodal AI? Dynamiq lets teams build multimodal agents on the models they choose and run them inside their own infrastructure: - **Agents that see documents and images.** Agents accept images and files alongside text, and document parsing and extraction feed knowledge bases for retrieval; see [Knowledge](/product/knowledge-rag). - **Voice agents for inbound calls,** with 15 voice model vendors to choose from; see [Voice Agents](/product/voice-agents). - **A choice of models.** Use any of 29 model providers from the SDK, call the most-used models through one AI Gateway endpoint, or run open models yourself; see [models](/product/models). - **Controls for sensitive data.** Detect PII and prompt injection, pause tool steps for human approval and trace every run; see [guardrails and approvals](/product/guardrails). The whole platform can run in Dynamiq's cloud or self-hosted in your own cloud account or data center; see [how we deploy](/deployment). ## FAQ ### What is multimodal AI in simple terms? Multimodal AI is AI that can work with several kinds of information at once, such as text, pictures, sound and video. It can look at a photo, read a document and listen to a recording, then answer a question that needs all three. ### What is the difference between multimodal AI and generative AI? Generative AI describes what a model does: it creates new content. Multimodal describes what a model can take in or produce: more than one type of data. Many current generative models are multimodal, such as a model that reads an image and writes a description of it. ### What are examples of multimodal AI in business? Common examples include extracting data from scanned documents, customer support agents that understand screenshots and voice messages, insurance claims review that compares photos with policy text, and quality inspection that combines camera images with sensor data. ### Is multimodal AI safe for sensitive documents? It can be, with the right controls: process documents in infrastructure you control, detect personal data in extracted text, limit who can access the results, keep a person in the loop for decisions, and keep a trace of every run for audit. ### Do I need to train my own multimodal model? Usually not. Most businesses start from an existing foundation model, add their own documents through retrieval and carefully written instructions, and fine-tune only when evaluation on their own data shows a gap the other steps can't close. --- # How the EU AI Act impacts your business: deadlines, obligations and fines URL: https://www.getdynamiq.ai/post/how-the-eu-ai-act-will-impact-your-business Published: 2024-08-23 What the EU AI Act requires of businesses in 2026: risk categories, what applies now, the new high-risk deadlines after the AI Omnibus, fines and next steps. The EU AI Act affects any business that develops, sells or uses AI systems in the European Union, including companies based outside it. What you must do depends on your system's risk category and your role: most companies are deployers of AI and face lighter duties than the providers who build it. As of September 2026, bans and transparency rules already apply, and obligations for high-risk AI start on 2 December 2027. This guide covers who the Act applies to, the risk categories, the current timeline including the 2026 AI Omnibus changes, the fines, and practical steps to prepare. It is general information, not legal advice. ## What is the EU AI Act? The [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) (Regulation (EU) 2024/1689) is the European Union's law on artificial intelligence. It was adopted by the European Parliament and the Council, entered into force on 1 August 2024, and applies in stages. It regulates AI by risk: the more an AI system can harm health, safety or fundamental rights, the stricter the rules. It sits alongside existing law rather than replacing it. GDPR still governs personal data, and sector rules in finance, healthcare and product safety still apply. See our [glossary entry on the EU AI Act](/glossary/eu-ai-act) for a short definition. ## Does the EU AI Act apply to my business? It applies if you are one of these, regardless of where your company is based: - **Provider:** you develop an AI system or model, or have one developed, and place it on the EU market or put it into service under your name. - **Deployer:** you use an AI system in a professional capacity in the EU, for example a bank using a chatbot or an employer using a screening tool. - **Importer or distributor:** you bring AI systems from outside the EU to the EU market or resell them. - **Non-EU provider or deployer** whose AI system's output is used in the EU. Most enterprises are deployers. You can become a provider without meaning to, for example by putting your own name on a high-risk system or substantially modifying one. ## What are the risk categories under the EU AI Act? | Category | Examples | What it means | | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------- | | Unacceptable risk | Social scoring, harmful manipulation, emotion recognition at work or school, untargeted scraping of facial images | Banned | | High risk | Credit scoring of individuals, recruitment and promotion decisions, access to essential services, AI in medical devices and machinery | Strict requirements before and after launch | | Transparency risk | Chatbots, AI-generated or altered images, audio and video | People must be told, and content must be marked | | Minimal or no risk | Spam filters, recommendation of internal documents, most productivity tools | No new obligations | General-purpose AI models, such as the large language models behind most assistants, have their own set of rules for the companies that train and release them. ## What applies now, and what comes next? | Date | What applies | | --------------- | ------------------------------------------------------------------------------------------------------------------- | | 1 August 2024 | The Act enters into force | | 2 February 2025 | Bans on unacceptable-risk practices and the AI literacy duty | | 2 August 2025 | Rules for general-purpose AI models, governance and penalties framework | | 27 July 2026 | The AI Omnibus amendment enters into force | | 2 August 2026 | Transparency rules for chatbots and generated content; enforcement by the AI Office and national authorities begins | | 2 December 2026 | Ban on AI that generates non-consensual intimate images or child sexual abuse material | | 2 December 2027 | Obligations for high-risk AI systems listed in Annex III, such as credit scoring and recruitment | | 2 August 2028 | Obligations for high-risk AI built into products covered by EU product safety law, listed in Annex I | The dates after July 2026 reflect the AI Omnibus, as summarized in the European Commission's [announcement that it entered into force](https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force). ## What changed with the 2026 AI Omnibus? The European Commission proposed the AI Omnibus in November 2025 to simplify the Act. After a political agreement in May 2026, it entered into force on 27 July 2026. The main changes for businesses: - **Later high-risk deadlines.** Annex III systems now have until 2 December 2027, and AI in regulated products until 2 August 2028. - **Simpler AI literacy duty.** The requirement on companies was simplified, with the Commission and member states taking a larger role. - **Lighter registration.** Registering AI systems that are exempt from high-risk status in the EU database was simplified. - **More relief for growing companies.** Some measures previously reserved for small and medium-sized enterprises now also cover small mid-caps. - **A new ban.** AI systems that generate non-consensual sexually explicit images or child sexual abuse material are prohibited. ## What must you do if your AI system is high-risk? Providers of high-risk systems must have: - A risk management system that runs for the whole life of the system. - Data governance for training, validation and test data. - Technical documentation and automatic logging of events. - Clear instructions for deployers, including limitations. - Human oversight built into the design. - Appropriate accuracy, robustness and cybersecurity. - A quality management system, a conformity assessment, registration in the EU database and post-market monitoring, including reporting serious incidents. Deployers of high-risk systems must use them as instructed, assign trained people to oversee them, monitor them, keep the logs they generate and inform affected workers. Public bodies, private entities providing public services, and any deployer using AI for credit scoring or for life and health insurance pricing must also carry out a fundamental rights impact assessment. For banks, the most common high-risk case is creditworthiness assessment. Our post on [generative AI in banking](/post/generative-ai-and-llms-in-banking-examples-use-cases-limitations-and-solutions) covers where that line falls. ## What are the transparency obligations? Since 2 August 2026, the transparency rules require that: - Chatbots and other AI systems that interact with people tell them they are dealing with an AI, unless it is obvious. - Deepfakes, meaning AI-generated or altered images, audio or video of real-looking people, places or events, are labeled. - AI-generated or altered content carries machine-readable marks so it can be detected. - AI-generated text published to inform the public on matters of public interest is disclosed as such, unless a person reviewed it and takes editorial responsibility. For most enterprises, the chatbot rule is the one that applies today. A customer service agent is usually not high-risk, but it must say that it is an AI. ## What are the penalties for non-compliance? The Act sets maximum fines as a fixed amount or a share of worldwide annual turnover, whichever is higher: | Violation | Maximum fine | | ----------------------------------------------------------------------------- | ---------------------------------------------- | | Using a prohibited AI practice | €35 million or 7% of worldwide annual turnover | | Breaching most other obligations, including high-risk and transparency duties | €15 million or 3% | | Supplying incorrect or misleading information to authorities | €7.5 million or 1% | For small and medium-sized enterprises, the lower of the two amounts applies. Providers of general-purpose AI models face separate fines of up to €15 million or 3%. ## How should a business prepare? 1. **Inventory your AI.** List every AI system you build, buy or embed, including features inside vendor software. 2. **Classify each system.** Decide whether it is prohibited, high-risk, subject to transparency rules or minimal risk, and record why. 3. **Confirm your role.** Provider or deployer duties differ, and customizing a system can change your role. 4. **Close the transparency gaps now.** Make sure chatbots disclose that they are AI and generated media is labeled. 5. **Plan for high-risk obligations.** Use the time until December 2027 to build logging, human oversight, testing and documentation into those systems. 6. **Question your vendors.** Ask AI providers for documentation, logging capabilities and the data they keep. 7. **Train your people.** Staff who operate or oversee AI need to understand what it can and cannot do. ## How does Dynamiq support EU AI Act compliance? No platform makes you compliant on its own; compliance depends on your use case, your documentation and your processes. A platform can supply the controls and evidence the Act asks for. This is how Dynamiq's capabilities map to common obligations: | Obligation | How Dynamiq helps | | -------------------------------- | ----------------------------------------------------------------------------------------- | | Record-keeping and logging | Every run is traced step by step with inputs, outputs, cost and latency | | Human oversight | Approvals pause any tool step for a reviewer, who can edit fields before approving | | Accuracy and robustness | Evaluations on test datasets before release and real-time evaluations on live deployments | | Risk controls | Guardrails detect PII and prompt injection before a model sees the text | | Change control and documentation | Versioned workflows with rollback, so you can show what changed and when | | Data governance | Self-hosting in your own cloud account or data center | These features are part of Dynamiq's [guardrails and approvals](/product/guardrails), evaluations and observability. The platform runs in Dynamiq Cloud or self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes cluster or on-prem, and Dynamiq has SOC 2, HIPAA and GDPR in place. ## FAQ ### When does the EU AI Act take effect? It entered into force on 1 August 2024 and applies in stages. Bans apply since 2 February 2025, general-purpose AI model rules since 2 August 2025, and transparency rules since 2 August 2026. High-risk obligations apply from 2 December 2027, or 2 August 2028 for AI in regulated products. ### Does the EU AI Act apply to companies outside the EU? Yes. It applies to providers that place AI systems on the EU market and to providers and deployers outside the EU whose AI output is used in the EU, in the same way GDPR reaches non-EU companies that process EU residents' data. ### Is a customer service chatbot high-risk under the EU AI Act? Usually not. A customer service chatbot falls under the transparency rules, so it must tell people they are talking to an AI. It becomes high-risk only if it is used for a listed purpose, such as deciding on credit or on access to essential services. ### What are the fines under the EU AI Act? Up to €35 million or 7% of worldwide annual turnover for prohibited practices, up to €15 million or 3% for most other breaches, and up to €7.5 million or 1% for misleading information to authorities, whichever is higher. For SMEs, the lower amount applies. ### What did the AI Omnibus change? It delayed high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for AI in regulated products, simplified the AI literacy and registration duties, extended some SME relief to small mid-caps, and banned AI that generates non-consensual intimate imagery or child sexual abuse material. --- # Mastering LLM security: an air-gapped approach for high-security deployments URL: https://www.getdynamiq.ai/post/mastering-llm-security-an-air-gapped-solution-for-high-security-deployments Published: 2024-08-20 How to secure LLMs and AI agents in high-security environments: the OWASP 2025 risks, when an air-gapped deployment pays off, and the controls it still needs. LLM security means protecting the models, the data they read, the prompts and outputs they exchange, and the actions agents take. An air-gapped deployment is the strictest option: models, data and the agent platform run on infrastructure with no connection to outside networks, so nothing can leak to a third party. It removes the network as an attack path, but not the risks inside the application, such as prompt injection, over-privileged agents or poisoned models, so those still need their own controls. This guide covers the main LLM security risks, when an air gap is worth its cost, what it does not protect against, and what running LLMs in an isolated environment takes. ## What are the main security risks for LLM applications? The [OWASP Top 10 for LLM applications](https://genai.owasp.org/llm-top-10/) is the most widely used reference. Its 2025 edition lists these risks, with the controls that address each: | Risk | What it means | Main controls | | -------------------------------- | ---------------------------------------------------------------------------------------- | ---------------------------------------------------------- | | Prompt injection | Input, including text inside documents or web pages, overrides the system's instructions | Input detectors, least-privilege tools, approvals | | Sensitive information disclosure | The system reveals personal, financial or confidential data | PII detection, permission-aware retrieval, output checks | | Supply chain | Compromised models, datasets, packages or plugins | Verified sources, checksums, scanning, internal registries | | Data and model poisoning | Tampered training, fine-tuning or retrieval data changes behavior | Data provenance, evaluations before release | | Improper output handling | Model output is passed to other systems without validation | Output validation, sandboxed execution | | Excessive agency | An agent has more permissions or autonomy than its task needs | Narrow tools, approvals on consequential actions | | System prompt leakage | Instructions or secrets in prompts are exposed | Keep secrets out of prompts, test for leakage | | Vector and embedding weaknesses | Retrieval stores leak data across users or accept poisoned content | Access control on retrieval, ingestion checks | | Misinformation | Confident but false outputs are trusted | Grounding with citations, human review | | Unbounded consumption | Runaway usage drives cost or denies service | Rate limits, loop caps, cost monitoring | ## What is an air-gapped LLM deployment? An air-gapped deployment runs every component inside a network with no path to the internet or other outside networks: the models and inference servers, vector stores and databases, the agent platform, identity, logging and monitoring. Models, software updates and data enter only through controlled, audited transfers. See our [glossary entry on air-gapped deployment](/glossary/air-gapped-deployment) for a short definition. It differs from the more common self-hosted setup, where the platform runs in your own cloud account or data center but can still call outside model providers through egress you control. The table below compares the options: | | Managed cloud | Self-hosted with controlled egress | Fully isolated (air-gapped) | | ------------------ | --------------------------------- | ---------------------------------- | ------------------------------------- | | Where data lives | The vendor's cloud | Your cloud account or data center | Your isolated network | | Outside calls | To the vendor and model providers | Only to providers you allow | None | | Models | Hosted and open models | Hosted and self-hosted models | Self-hosted open-weight models only | | Updates | Continuous, by the vendor | Your schedule | Controlled transfers on your schedule | | Operational effort | Lowest | Moderate | Highest | ## When does an air-gapped deployment make sense? Choose an air gap when the data or the mission cannot tolerate any outside connection: - **Government and defense** work with classified or mission-critical information. - **Critical infrastructure** operators whose control networks are isolated by design. - **Financial institutions** processing core banking or trading data under strict internal rules, where supervisors or policy rule out external processing. - **Healthcare and life sciences** organizations handling patient or genomic data in isolated research environments. For many regulated organizations, a self-hosted deployment with egress limited to approved providers, or to no providers at all for sensitive workflows, meets the requirement at lower cost. Decide per workload, not per organization: the same bank may run a public-data research assistant in a managed cloud and a customer-data agent fully inside its own network. ## What does an air gap not protect against? An air gap stops data from leaving over the network. It does not make the application itself safe: - **Prompt injection still works.** A malicious instruction inside an internal document or an email the agent reads needs no internet connection. - **Insiders keep their access.** Users with legitimate access can still misuse an assistant or extract data through it. - **Over-privileged agents can still act.** An agent that can modify records will do so if it is manipulated or simply wrong. - **Imported models can be compromised.** Weights and packages brought across the gap must be verified, or the air gap protects a poisoned model. - **Wrong answers are still wrong.** Isolation does nothing for hallucinations. So an isolated deployment needs the same application controls as any other: input detectors, least-privilege tools, approvals on consequential actions, output validation and full tracing. ## What does it take to run LLMs in an isolated environment? - **Open-weight models** whose licenses allow your use, sized for the hardware you have. Our [overview of open-source LLMs](/post/less-risk-and-no-fee-overview-of-open-source-llms-for-enterprises) covers the main families. - **GPU capacity** for inference, and for fine-tuning if you customize models. - **A controlled transfer process** for models, container images and packages: download outside, scan, verify checksums, move through an approved channel. - **Internal mirrors and registries** for containers and packages, so nothing reaches outward during installs or updates. - **Identity and logging inside the boundary**, integrated with your internal identity provider and monitoring. - **Offline evaluation**, so every new model or prompt version is tested on your cases before promotion. - **People** who can operate GPUs, Kubernetes and the platform without vendor access to the environment. ## Which security controls does every LLM deployment need? Whether in a managed cloud or an isolated network: 1. **Least privilege** for users, agents and tools, with read-only access by default. 2. **Input screening** for PII and prompt injection, including retrieved documents. 3. **Approvals** before payments, record changes, outbound messages and other hard-to-undo actions. 4. **Output validation** against schemas and policies before results reach other systems. 5. **Sandboxed code execution**, separated from the platform and from other workloads. 6. **Secrets management**, with credentials encrypted and never placed in prompts. 7. **Tracing and retention** of every run, for investigation and audit. 8. **Testing** on adversarial cases before each release, and incident response when something slips through. ## How does Dynamiq support high-security deployments? Dynamiq can be [self-hosted](/deployment/self-hosted) on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes cluster or on-prem, with outbound calls going only to the model and tool providers you configure. Fully isolated and air-gapped environments are delivered with our engineers, who set up the platform and the models it uses to run entirely inside your boundary. Inside any deployment, guardrails detect PII and prompt injection before a model reads the text, approvals with editable fields gate any tool step, agent code runs in isolated sandboxes, credentials are encrypted in a Vault-based secrets manager, and every run is traced with inputs, outputs, cost and latency. Dynamiq has SOC 2, HIPAA and GDPR in place, with SSO on the Enterprise plan. The [Security and Trust Center](/security) lists the controls in detail. ## FAQ ### What is an air-gapped LLM? It is a large language model deployed on infrastructure with no connection to outside networks, together with the data and applications that use it. Prompts, documents and outputs never leave the isolated environment. ### Is an air-gapped deployment more secure than a private cloud? It removes network paths for data to leave, which a private cloud with outbound access does not. It does not protect against application-level risks such as prompt injection or misuse by insiders, and it costs more to operate, so it is best reserved for workloads that truly require it. ### Can you use GPT or Claude in an air-gapped environment? Not through their public APIs, which need a network connection to the provider. Air-gapped deployments run open-weight models, such as Llama, Mistral, Qwen or Gemma, on local GPUs. ### How do you update models in an air-gapped environment? Through a controlled transfer: download and scan the model outside, verify its checksums, move it through an approved channel, then run your evaluation set inside the environment before promoting it. ### Does Dynamiq offer air-gapped deployment? Yes. Fully isolated and air-gapped environments are delivered with our engineers, who set up the platform and its models to run entirely inside your boundary. Standard self-hosted installs run in your cloud account or data center with outbound access limited to providers you configure. --- # AMD Instinct MI250 vs NVIDIA A100: the LLM inference showdown URL: https://www.getdynamiq.ai/post/amd-instinct-mi250-vs-nvidia-a100-the-llm-inference-showdown Published: 2024-07-16 Our vLLM benchmark of the AMD Instinct MI250 against NVIDIA A100 40GB and 80GB on 7B to 14B models: throughput, latency, setup and what it means today. In our mid-2024 benchmark with vLLM 0.5.0, the NVIDIA A100 was faster than the AMD Instinct MI250 on every model we tested. The MI250 reached about 84% of the A100 40GB's throughput on 7B to 8B models and about 76% on 11B to 14B models, while the A100 80GB led both. The MI250's strength is memory: 128 GB of HBM2e per card, which matters more as models and context windows grow. ## Which GPUs did we compare? We compared AMD's Instinct MI250 with two versions of NVIDIA's A100, all data center GPUs aimed at AI and high-performance computing: | Specification | NVIDIA A100 40GB SXM | NVIDIA A100 80GB SXM | AMD Instinct MI250 | | ------------------ | -------------------- | -------------------- | ------------------ | | Architecture | NVIDIA Ampere | NVIDIA Ampere | AMD CDNA 2 | | Power | 400 W | 400 W | 500 W | | GPU memory | 40 GB HBM2 | 80 GB HBM2e | 128 GB HBM2e | | FP32 | 19.5 TFLOP/s | 19.5 TFLOP/s | 45.3 TFLOP/s | | FP16/BF16 (matrix) | 312 TFLOP/s | 312 TFLOP/s | 362 TFLOP/s | | Memory bandwidth | 1,555 GB/s | 2,039 GB/s | 3,277 GB/s | | L2 cache | 40 MB | 40 MB | 16 MB | The MI250 figures are for the whole card. An MI250 package holds two graphics compute dies (GCDs), each with 64 GB of memory, and the ROCm runtime treats each die as a separate GPU. That detail matters for reading the results below. ## How did we run the benchmark? We used vLLM, a widely used open-source LLM serving engine, with the same settings on both platforms: - **Batch size:** 8, to measure how each GPU handles several prompts at once - **GPU memory utilization:** 90% - **Tensor parallel size:** 1, to measure a single GPU | Stack | AMD MI250 machine | NVIDIA A100 machines | | ------------------ | ----------------- | -------------------- | | Linux distribution | Ubuntu 20.04 | Ubuntu 20.04 | | GPU software | ROCm 6.1.2 | CUDA 12.1 | | GPU communication | RCCL 2.18.6 | NCCL 2.18.3 | | Inference engine | vLLM 0.5.0 | vLLM 0.5.0 | We ran the `benchmark_throughput.py` and `benchmark_latency.py` scripts from the vLLM repository. Throughput runs used prompts from the [ShareGPT dataset](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered) in batches of 8. Latency runs used randomly generated prompts of a fixed length, so every request in a batch had the same input and output size. With a tensor parallel size of 1, each vLLM process runs on a single ROCm device. On the MI250 that is one of its two dies, so the MI250 results compare one die with one full A100. ## How fast was the MI250 compared with the A100? ### 7B to 8B models These models, such as Llama 3 8B, Mistral 7B and Gemma 7B, are popular starting points for fine-tuning. | Model | Metric | MI250 | A100 40GB | A100 80GB | | ----------------- | ------------------- | ----- | --------- | --------- | | Meta-Llama-3-8B | Latency (s/request) | 2.64 | 1.93 | 1.70 | | Meta-Llama-3-8B | Requests per second | 10.90 | 13.24 | 14.22 | | Meta-Llama-3-8B | Tokens per second | 4,517 | 5,487 | 5,879 | | Mistral-7B-v0.3 | Latency (s/request) | 2.42 | 1.83 | 1.62 | | Mistral-7B-v0.3 | Requests per second | 9.38 | 11.09 | 12.78 | | Mistral-7B-v0.3 | Tokens per second | 4,739 | 5,626 | 6,032 | | gemma-7b-it | Latency (s/request) | 3.32 | 2.23 | 1.96 | | gemma-7b-it | Requests per second | 8.40 | 8.71 | 10.34 | | gemma-7b-it | Tokens per second | 3,703 | 3,837 | 4,556 | | Qwen2-7B-Instruct | Latency (s/request) | 2.73 | 1.84 | 1.63 | | Qwen2-7B-Instruct | Requests per second | 10.37 | 14.65 | 15.67 | | Qwen2-7B-Instruct | Tokens per second | 4,377 | 6,164 | 6,559 | On these models, the MI250 delivered 71% to 96% of the A100 40GB's throughput (about 84% on average) and 66% to 81% of the A100 80GB's. Its latency per request was 1.3 to 1.5 times the A100 40GB's. ![Bar chart of latency per request for four 7B to 8B models, with the MI250 slowest and the A100 80GB fastest on each](/blog/amd-instinct-mi250-vs-nvidia-a100-the-llm-inference-showdown/2-latency-by-model-seconds-per-request.webp "Latency for 7B to 8B models, in seconds per request (lower is better).") ![Bar chart of requests per second for four 7B to 8B models, with the A100 80GB highest and the MI250 lowest on each](/blog/amd-instinct-mi250-vs-nvidia-a100-the-llm-inference-showdown/3-throughput-by-model-request-per-second-1.webp "Throughput for 7B to 8B models, in requests per second (higher is better).") ![Bar chart of tokens per second for four 7B to 8B models, with the A100 80GB highest and the MI250 lowest on each](/blog/amd-instinct-mi250-vs-nvidia-a100-the-llm-inference-showdown/4-throughput-by-model-tokens-per-second-1.webp "Throughput for 7B to 8B models, in tokens per second (higher is better).") ### 11B to 14B models | Model | Metric | MI250 | A100 40GB | A100 80GB | | -------------------------- | ------------------- | ----- | --------- | --------- | | Llama-2-13b-chat-hf | Latency (s/request) | 4.16 | 3.78 | 2.67 | | Llama-2-13b-chat-hf | Requests per second | 2.60 | 3.61 | 6.36 | | Llama-2-13b-chat-hf | Tokens per second | 1,271 | 1,737 | 3,062 | | Phi-3-Medium-4k-instruct | Latency (s/request) | 3.45 | 2.88 | 2.11 | | Phi-3-Medium-4k-instruct | Requests per second | 4.86 | 6.19 | 7.59 | | Phi-3-Medium-4k-instruct | Tokens per second | 2,289 | 2,976 | 3,650 | | falcon-11B | Latency (s/request) | 3.89 | 3.21 | 2.49 | | falcon-11B | Requests per second | 4.64 | 6.17 | 9.19 | | falcon-11B | Tokens per second | 2,444 | 3,303 | 4,205 | | Phi-3-Medium-128k-instruct | Latency (s/request) | 3.78 | 2.99 | 2.20 | | Phi-3-Medium-128k-instruct | Requests per second | 4.58 | 5.81 | 7.11 | | Phi-3-Medium-128k-instruct | Tokens per second | 2,202 | 2,793 | 3,420 | On the larger models, the MI250 delivered 72% to 79% of the A100 40GB's throughput (about 76% on average) but only 41% to 64% of the A100 80GB's. The A100 80GB's extra memory and bandwidth pay off here: more room for the KV cache means more requests in flight. ![Bar chart of tokens per second for four 11B to 14B models, with the A100 80GB highest and the MI250 lowest on each](/blog/amd-instinct-mi250-vs-nvidia-a100-the-llm-inference-showdown/7-graph-6-throughput-by-model-tokens-per-second.webp "Throughput for 11B to 14B models, in tokens per second (higher is better).") ## What do the results mean? Three things stood out: - **The comparison was one die against one GPU.** Each MI250 result used one of the card's two dies, which has about half of the card's compute and memory bandwidth. Seen that way, one die reaching 76% to 84% of an A100 40GB's throughput is a solid result, and a full card can serve a second model replica on its other die. We did not measure that configuration, but it belongs in any price-performance comparison. - **Software maturity mattered.** CUDA's mature kernels and vLLM's longer history on NVIDIA hardware were an advantage in 2024. vLLM's support for ROCm has improved since. - **Memory decides what fits.** The MI250's 64 GB per die holds larger models or longer contexts than an A100 40GB can, which is worth more than raw speed for some workloads. ## Do these results still hold in 2026? Treat them as a snapshot. Both GPUs are now a generation or more behind AMD's Instinct MI300 series and NVIDIA's H100 and later GPUs, and vLLM and ROCm have both changed a great deal since version 0.5.0 and 6.1.2. The method still applies: fix the serving engine and settings, test the models you actually run, and measure throughput and latency at your real batch sizes. ## How should you choose inference hardware for enterprise agents? Agent workloads make many model calls per task, so the cost and speed of each call add up. When you size hardware for self-hosted models, check: 1. **Memory per GPU** for your largest model plus the KV cache your context lengths and concurrency need. 2. **Throughput at your real batch sizes,** measured with your prompts, not a vendor's. 3. **Latency per request,** which users feel directly in chat and voice agents. 4. **Software support** for your serving engine and model architectures on that hardware. 5. **Availability** in your cloud region or data center, which often decides more than benchmarks do. On Dynamiq, you can run open models on managed inference and fine-tune them with LoRA adapters, or call the most-used hosted models through one AI Gateway endpoint; see [models](/product/models). For hardware inside your own perimeter, Dynamiq runs self-hosted on any Kubernetes cluster; see [where Dynamiq runs](/deployment/self-hosted). ## FAQ ### Is the AMD MI250 faster than the NVIDIA A100 for LLM inference? Not in our 2024 tests. With vLLM 0.5.0, one MI250 die reached about 84% of an A100 40GB's throughput on 7B to 8B models and about 76% on 11B to 14B models, and the A100 80GB was faster than both on every model. ### Why does the A100 80GB beat the A100 40GB on the same chip? It has twice the memory and more bandwidth. For inference, extra memory holds a larger KV cache, so the server can keep more requests in flight, which raises throughput, especially on 11B to 14B models. ### Does vLLM support AMD GPUs? Yes. vLLM supports AMD Instinct GPUs through ROCm, and we ran this benchmark on vLLM 0.5.0 with ROCm 6.1.2. Support has expanded considerably since then, so test current versions before you decide. ### Which GPU should I choose for self-hosted LLMs today? Start from your workload: the models you will serve, their context lengths, your concurrency and your latency target. Then benchmark the GPUs you can actually get in your cloud region or data center, with the serving engine you plan to use. ### What is LLM inference? LLM inference is running a trained model to generate output from a prompt. Its cost and speed depend on the model's size, the hardware and the serving engine; our [glossary entry on LLM inference](/glossary/llm-inference) covers the basics. --- # GenAIOps for enterprises: the build vs. buy decision URL: https://www.getdynamiq.ai/post/genaiops-for-enterprises-the-build-vs-buy-dilemma Published: 2024-07-03 GenAIOps is how enterprises run generative AI and agents in production. Compare building a platform with buying one: costs, risks and a checklist. Most enterprises should buy a GenAIOps platform and build their use cases on top of it, rather than build the platform itself. Building makes sense when AI infrastructure is your core product and you can fund a dedicated platform team for years. Everyone else reaches production sooner, with fewer security gaps, on a platform that runs in their own environment and can be extended in code. This guide explains what GenAIOps covers today, where the costs hide on each side, and how to make the call. ## What is GenAIOps? GenAIOps (generative AI operations) is the set of practices and tools for taking generative AI applications from prototype to production and keeping them reliable, safe and affordable there. It grew out of MLOps, but the unit of work changed. Instead of training and serving one predictive model, teams now combine hosted and open models, prompts, retrieval, tools and agents into workflows that call other systems. Today most GenAIOps work is about agents. An agent plans steps, calls tools and reads documents, and it can act on its own. So the operational questions go well beyond model accuracy: what can the agent touch, who approved an action, what did the run cost, and can you replay it for an auditor? ## What does a GenAIOps platform include? A complete platform covers the whole path from a first prototype to an audited production system: | Capability | What it covers | | ------------------------ | ------------------------------------------------------------------------------------ | | Model access | One gateway to hosted and self-hosted models, with shared credentials and fallbacks | | Agents and workflows | Visual or code-first design of agents, tools and multi-step flows | | Knowledge and retrieval | Document parsing, chunking, embeddings, vector stores and permission-aware retrieval | | Integrations | Connectors to business apps, databases and MCP servers | | Guardrails and approvals | PII and prompt injection detection, output validation, human sign-off on risky steps | | Evaluation | Test datasets, LLM-as-a-judge and code metrics, scoring of live traffic | | Observability | Traces of every step with inputs, outputs, cost and latency | | Deployment | APIs, versions and rollback, schedules, roles and SSO | If one of these is missing, a team usually ends up building it by hand for each use case, which is how "buy" projects quietly turn into "build" projects. ## How is GenAIOps different from MLOps? MLOps manages models you train. GenAIOps manages systems you compose, mostly from models someone else trained. | | MLOps | GenAIOps | | ----------------------- | ------------------------------ | ------------------------------------------------------------------------- | | Main asset | A model you train | Prompts, retrieval indexes, tools and agents around models you consume | | Most frequent change | Retrained weights | Prompt, workflow and model version changes | | How quality is measured | Accuracy on a labeled test set | Task-specific evals, often with an LLM as judge, plus review of live runs | | Main risks | Drift and bias | Hallucination, prompt injection, data leakage, unsafe agent actions | | Main cost driver | Training compute | Inference tokens and tool calls at run time | ## What does it take to build a GenAIOps platform in-house? Building means assembling and running a dozen components: a model gateway, an orchestration runtime, vector storage and ingestion pipelines, an evaluation harness, tracing, guardrails, secrets management, access control and deployment tooling. Open-source frameworks cover parts of this. Integrating them, securing them and keeping them current is a platform team's full-time job. Most of the cost arrives after launch: - **Model churn.** Providers release and retire models every few months. Each change touches prompts, evals and cost baselines. - **New standards.** Protocols such as the [Model Context Protocol](/glossary/model-context-protocol) changed how agents connect to tools in a short time, and a homegrown platform has to keep up. - **Security review.** Every new capability that lets an agent act on a system needs threat modeling and controls. - **Compliance evidence.** Audit trails, access reviews and retention policies have to be produced for auditors and regulators, not just built. - **Key-person risk.** A small platform team is hard to hire and harder to replace when someone leaves. ## When does building in-house make sense? Build when most of these are true: - AI infrastructure is your product, or a core differentiator you sell. - You have, or can hire and keep, a team with ML, backend, security and SRE skills for several years. - After an honest evaluation, no platform meets a hard requirement you have. - You accept that your first production use case waits for the platform. ## When does buying make sense? Buy when most of these are true: - You want use cases in production this quarter, not after a platform project. - Your engineers are worth more on products and integrations than on infrastructure. - You work under regulation and need audit trails, access control and certifications without assembling them yourself. - Several teams, technical and not, need to build on one governed platform. ## Build vs. buy at a glance | Factor | Build in-house | Buy a platform | | ----------------------- | ------------------------------------------ | --------------------------------------------------------------- | | Fit | Tailored to your workflows | Configurable, and extensible in code if the platform has an SDK | | Data control | Full, by design | Full only if the platform runs in your environment | | Time to first use case | Long: platform first, use cases after | Short: use cases from the start | | Upfront cost | High: team, infrastructure, tooling | Lower: license and integration work | | Ongoing cost | Platform team, infrastructure, upgrades | Renewals and usage, plus infrastructure if self-hosted | | Security and compliance | You build and evidence every control | Inherited controls plus your own review | | Main risk | Delays, technical debt, key people leaving | Vendor dependency and integration gaps | ## What does each option really cost? Model inference costs the same whichever way you go: you pay per token or per GPU hour for the models your use cases call. The difference is people and time. Building puts the largest line item on salaries for a multi-disciplinary platform team, followed by infrastructure, tooling, and the cost of use cases that wait while the platform is built. Technical debt grows as components age and the team shrinks. Buying moves most of that into a license, plus integration work and change management. If you self-host, you also run the infrastructure, but you do not maintain the software. To size the benefit side of the business case, the [AI agent ROI calculator](/resources/ai-agent-roi-calculator) estimates what an agent could save a team. ## How do you buy without getting locked in? Vendor lock-in is the most common objection to buying, and procurement questions can reduce most of it: - Can the whole platform run in our cloud account or data center, not only the model? - Can we bring our own models, vector stores and identity provider? - Is there a code path, such as an SDK, so engineers are not limited to a visual builder? - Where are prompts, traces and evaluation data stored, and can we export them? - Which certifications does the vendor hold, and which controls stay with us? Our vendor-neutral [platform evaluation kit](/resources/agent-platform-evaluation-kit) turns these into a scored checklist you can use with any vendor, including us. ## Where does Dynamiq fit? Dynamiq is a platform you buy and still control. [Agent Builder](/product/agents) gives business teams a visual canvas and engineers the open-source Python SDK (Apache-2.0), both running on one engine, so a workflow designed on the canvas and one written in code ship the same way. The platform runs in Dynamiq Cloud or [self-hosted](/deployment/self-hosted) on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes cluster or on-prem. Fully isolated setups are delivered with our engineers. It includes guardrails that detect PII and prompt injection, approvals on tool steps, evaluations on test sets and live traffic, and traces with cost and latency for every run. Dynamiq has SOC 2, HIPAA and GDPR in place, with SSO on the Enterprise plan. For teams without spare engineering capacity, our [forward-deployed engineers](/deployment) build the first use cases with you, inside your environment. That is how an Asian neo-bank [automated about 85% of support inquiries and went live in 30 days](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai), with one of our engineers working alongside its team. ## A five-step checklist for the decision 1. List your first three use cases, their owners and their deadlines. 2. Write down the hard requirements: data residency, identity provider, audit and retention rules. 3. Estimate the team you would need to build the platform and keep it current for three years. 4. Run the same use case on two or three platforms, in your environment, with your data. 5. Decide on time to production and three-year total cost, not on license price alone. ## FAQ ### What is GenAIOps? GenAIOps is the set of practices and tools for running generative AI applications and agents in production. It covers model access, retrieval, guardrails, evaluation, tracing and deployment, so teams can ship changes safely and explain what the system did. ### Is GenAIOps the same as LLMOps? The terms overlap. LLMOps usually refers to operating large language models specifically, while GenAIOps also covers agents, retrieval pipelines, tools and other generative models. In practice, vendors and teams use them almost interchangeably. ### How long does it take to build a GenAIOps platform in-house? Longer than most plans assume, because the platform has to work before the first use case can ship, and it needs a permanent team afterward. Buying lets the first use case start right away, with the effort going into integrations and testing. ### Can an off-the-shelf GenAIOps platform run in our own cloud? Some can. Ask whether the whole platform, including traces, knowledge bases and credentials, runs in your account or data center, or only the model. Dynamiq can be self-hosted on the major clouds, OpenShift, any Kubernetes cluster or on-prem. ### How do we avoid vendor lock-in with a GenAIOps platform? Choose a platform you can self-host, that works with your own models and identity provider, and that has an SDK so workflows live in code you control. Confirm that traces and evaluation data are stored in your database and can be exported. --- # Think global, act local: how data residency regulations influence banks' LLM use URL: https://www.getdynamiq.ai/post/think-global-act-local-how-data-residency-regulations-influence-banks-llm-use Published: 2024-06-26 How data residency and transfer rules shape banks' use of LLMs, where AI systems touch regulated data, and architectures that keep that data in-country. Data residency regulations influence banks' LLM use by limiting where customer data may be stored and processed, and those limits apply to everything an AI system touches: prompts, retrieved documents, embeddings, logs and fine-tuning data. Few laws ban cloud AI outright. Instead, cross-border transfer rules, sector-specific localization requirements and supervisors' outsourcing expectations combine, so banks increasingly run models and agent platforms inside the jurisdiction, often in their own infrastructure. This guide explains the terms, where LLMs touch regulated data, the types of rules involved, and architectures that let banks use LLMs while keeping data where it must stay. It is general information, not legal advice. ## What is data residency? Three related terms are often confused: | Term | Meaning | | ----------------- | --------------------------------------------------------------------------------------------------------------------------------- | | Data residency | Where data is physically stored and processed, often chosen by the organization or promised by a vendor | | Data localization | A legal requirement that certain data stay in a country, or that a copy be kept there | | Data sovereignty | The principle that data is subject to the laws of the country where it is located, including access by that country's authorities | For banks, the practical question combines all three: which data may leave which jurisdiction, under what conditions, and who could access it once it does. See our [glossary entry on data residency](/glossary/data-residency) for a short definition. ## Where do LLMs touch regulated data? An LLM application creates more copies of data than most teams expect. Each of these needs a location, an access policy and a retention period: - **Prompts and responses** sent to a model, including the customer details an agent adds from other systems. - **Retrieved documents** pulled into the model's context from knowledge bases. - **Embeddings** in vector stores, which are derived from the source text and can reveal it. - **Traces and logs** that record every step of an agent run for debugging and audit. - **Memory** that stores conversation history or long-term facts about a customer. - **Fine-tuning data** and the adapted model weights trained on it. - **Backups and support access**, including where a vendor's staff can see data from. A model API in another country can receive customer data even when the bank's databases never move, simply because an agent put that data into a prompt. ## Which rules affect where banks can process data with LLMs? Four kinds of rules apply, and a bank usually faces several at once: | Type of rule | Examples | What it means for LLMs | | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------ | | Cross-border transfer rules in privacy law | GDPR in the EU, UK GDPR, Switzerland's FADP, Brazil's LGPD, Japan's APPI, South Korea's PIPA | Personal data can leave only with a legal transfer mechanism, such as an adequacy decision or approved contract clauses | | Sector and national localization rules | India's requirement that payment system data be stored only in India; Russia's rule that citizens' personal data be stored in Russian databases; China's security assessment and other conditions for exporting personal information | Some data must stay in the country, whatever the transfer mechanism | | Supervisory outsourcing and resilience rules | The EU's DORA and the EBA outsourcing guidelines, MAS guidelines in Singapore, OSFI's third-party risk guideline in Canada, APRA's CPS 230 in Australia, US interagency guidance on third-party relationships | AI vendors and model APIs are third-party services: they need risk assessment, contracts with audit and exit rights, and ongoing oversight | | Banking secrecy | Swiss, Singaporean and Luxembourg banking secrecy provisions | Client-identifying data may not be disclosed to service providers without safeguards or consent, wherever they are | Two points are often misunderstood. GDPR does not require data to be stored in the EU; it restricts transfers outside the European Economic Area unless a transfer mechanism applies. And enforcement of localization rules is real: in 2021, India's central bank stopped several international card networks from onboarding new customers until they met its payment data storage rules. In the EU, the [Digital Operational Resilience Act](https://eur-lex.europa.eu/eli/reg/2022/2554/oj), which applies from 17 January 2025, adds a register of ICT third-party providers and stricter contract, testing and exit requirements. An AI platform or model provider used in a critical function falls within that scope. ## How can banks use LLMs and keep data where it must stay? Most banks combine several of these patterns: - **Deploy in-region.** Run the agent platform, vector stores and logs in the jurisdiction whose rules apply, using a local cloud region or the bank's own data center. - **Self-host models for regulated data.** Run open-weight models inside your environment for steps that handle customer data, and use hosted models only for data that may leave. - **Route by data classification.** Detect personal data in a request, then send it to an in-region model or a person, and send everything else wherever it is cheapest to process. - **Minimize what reaches the model.** Retrieve only the fields a task needs and keep identifiers out of prompts where possible. - **Keep traces with the data.** Store logs in the same region, with retention periods aligned to your record-keeping obligations. - **Deploy per jurisdiction.** International banks often run one deployment per region with shared workflow definitions, so the same agent logic runs everywhere while the data stays local. Our [overview of open-source LLMs](/post/less-risk-and-no-fee-overview-of-open-source-llms-for-enterprises) covers the model options for self-hosting. ## What should banks ask an AI vendor about data residency? - Where are prompts, responses, embeddings, traces, memory and backups stored, and can we choose the region? - Which sub-processors and model providers receive our data, and from which countries? - From where can the vendor's staff access our environment for support? - Can the whole platform, not only the model, run in our cloud account or data center? - How is data deleted when we delete a project or end the contract? - What audit rights, exit support and incident notification terms does the contract include? - Which certifications does the vendor hold, and is a data processing agreement available? ## How does Dynamiq help banks meet data residency requirements? Dynamiq can be [self-hosted](/deployment/self-hosted) in a bank's own AWS, Azure, GCP or IBM Cloud account in the region it chooses, on Red Hat OpenShift, on any Kubernetes cluster or on-prem. In a self-hosted deployment, documents, knowledge bases, traces and credentials live in the bank's own database and storage, and outbound calls go only to the model and tool providers the bank configures. Fully isolated setups are delivered with our engineers, who set up the platform and the models it uses to run entirely inside the bank's boundary. Guardrails detect PII before text reaches a model, so workflows can route customer data to an in-region model or a person. Dynamiq has SOC 2, HIPAA and GDPR in place, and a data processing agreement is available. Read more about Dynamiq for [financial services](/industries/financial-services). ## FAQ ### What is data residency in banking? It is the requirement or practice of storing and processing customer and transaction data in specific countries or regions. For banks it covers core systems and every copy an application creates, including LLM prompts, embeddings and logs. ### Does GDPR require data to be stored in the EU? No. GDPR allows personal data to be stored anywhere, but transfers outside the European Economic Area need a legal mechanism, such as an adequacy decision or standard contractual clauses, plus an assessment of the risks in the destination country. ### Can banks use cloud-hosted LLMs under data residency rules? Often, for data that may leave the jurisdiction or through in-region services with the right contracts. For regulated customer data, many banks run models and the agent platform in-region or in their own infrastructure, and route sensitive steps there. ### Are embeddings personal data? They can be. Embeddings derived from personal data can reveal information about the person, so treat vector stores built from customer documents with the same residency, access and retention rules as the source data. ### How does DORA affect banks' use of AI vendors? Under DORA, AI platforms and model APIs are ICT third-party services. Banks must record them in their register of information, assess their risks, secure audit and exit rights in contracts, and monitor them, with stricter requirements when they support critical or important functions. --- # Less risk and no fee: an overview of open-source LLMs for enterprises URL: https://www.getdynamiq.ai/post/less-risk-and-no-fee-overview-of-open-source-llms-for-enterprises Published: 2024-06-19 Open-source and open-weight LLMs for enterprises: model families to consider, how licenses differ, what they really cost to run, and when to self-host. Open-source LLMs, more precisely open-weight LLMs, are language models whose trained weights you can download, run on your own infrastructure and fine-tune. For enterprises, they lower data risk because prompts and documents never leave your environment, and they carry no per-token license fee. They are not free to run, though: you pay for GPUs, operations and evaluation, and each model's license sets its own conditions. This overview explains what "open" means for LLMs, the model families worth evaluating, the real costs, and how to decide between open-weight and proprietary models. ## What is an open-source LLM? Most models called open source are open-weight: the developer publishes the trained weights and a license to use them, but not necessarily the training data or the full training code. The Open Source Initiative's [Open Source AI Definition](https://opensource.org/ai/open-source-ai-definition) sets a higher bar, requiring enough information about the training data, plus the code and weights, for others to study and rebuild the system. For enterprise use, the practical questions are simpler. Can you run the model where you want? Can you use it commercially? Can you modify and fine-tune it? The license answers all three, and licenses differ more than the "open" label suggests. See our [glossary entry on open-weight models](/glossary/open-weight-model) for a short definition. ## Why do enterprises use open-weight LLMs? - **Data control.** Prompts, documents and outputs stay inside your cloud account or data center, which simplifies data residency, privacy and supervisory questions. - **Cost at scale.** With steady, high volumes, owned inference can cost less than paying per token, because you pay for capacity rather than usage. - **Customization.** You can fine-tune a model on your own data, for example with LoRA adapters, to match your formats, terminology and tasks. - **Stability.** A model version you host does not change or disappear unless you decide it should, which helps with validation and model risk management. - **No single-vendor dependency.** You can switch between model families without renegotiating a contract. - **Transparency.** Published weights and model cards let you inspect, test and document the model more deeply than a closed API allows. ## What are the risks and hidden costs? Open weights move risk rather than remove it: - **Operations.** You run GPU capacity, serving, scaling, monitoring and upgrades, or pay someone to. - **Security.** Model files, serving endpoints and fine-tuning data become part of your attack surface, and model provenance matters: download weights only from the publisher's official channels. - **License terms.** Some licenses restrict use above a user threshold, in certain fields or regions, or require attribution. Legal review belongs in model selection. - **Quality on the hardest tasks.** Leading proprietary models are usually ahead on the most demanding reasoning and coding work. For extraction, classification, summarization and grounded question answering, open-weight models often perform well. - **Safety tuning.** Guardrails and evaluations are your job, whichever model you use, but a self-hosted model gets no provider-side filtering. ## Which open-weight model families should enterprises consider? These families are widely used in enterprise deployments. Check each model card and license for the exact version you plan to run, since terms can differ between releases: | Family | Developer | License | Notes | | -------- | ---------- | -------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- | | Llama | Meta | Llama Community License | Commercial use allowed with conditions, including a separate license for very large consumer services | | Mistral | Mistral AI | Apache 2.0 for many open models; others under research or commercial terms | Strong European option with small and mid-size models | | Qwen | Alibaba | Apache 2.0 for most recent models | Broad range of sizes, strong multilingual and coding performance | | Gemma | Google | Gemma Terms of Use | Small, efficient models that suit a single GPU | | DeepSeek | DeepSeek | MIT for recent releases | Large mixture-of-experts models with strong reasoning | | gpt-oss | OpenAI | Apache 2.0 | Open-weight reasoning models released in 2025 | | Granite | IBM | Apache 2.0 | Built for enterprise use, with documented training data sources | | Phi | Microsoft | MIT | Small models for constrained hardware | Smaller models, from a few billion to a few tens of billions of parameters, run on one or a few GPUs and handle most classification, extraction and routing work. The largest mixture-of-experts models need multi-GPU servers but come closest to proprietary quality. ## How do open-weight models compare with proprietary models? | | Open-weight, self-hosted | Proprietary API | | ---------------------------- | ---------------------------------------------- | ------------------------------------------ | | Where data goes | Stays in your environment | Sent to the provider, under its data terms | | Cost structure | GPU capacity and operations, no per-token fee | Pay per token, no infrastructure | | Quality on the hardest tasks | Good and improving, usually behind the leaders | Usually highest | | Customization | Full fine-tuning and adapters | Limited to what the provider offers | | Version control | You decide when to upgrade | The provider updates and retires models | | Effort to start | Higher | Lowest | Many enterprises end up using both: proprietary APIs for low-sensitivity, high-difficulty tasks, and open-weight models for sensitive data, high volumes and narrow tasks, behind one gateway so applications do not care which is which. ## Are open-source LLMs really free? The license fee is zero; the system is not. The main costs are: - GPU capacity, reserved or on demand, sized for peak load and context length. - Serving and scaling software, such as vLLM, plus monitoring and on-call support. - Evaluation: test sets and people to check quality before each upgrade. - Fine-tuning compute and data preparation, if you customize. - Security and compliance work for a new component in your stack. Self-hosting usually pays off when data rules require it, when volumes are high and steady enough to keep GPUs busy, or when a small, tuned model can replace a large general one. For spiky or low volumes, per-token APIs are often cheaper. ## How do you choose and deploy an open-weight model? 1. **Define the task and build a test set** from real examples, with the answers you expect. 2. **Shortlist two or three families** whose licenses fit your use, and pick sizes your hardware can serve. 3. **Test on your data**, comparing quality, latency and cost per task with a proprietary baseline. 4. **Decide on customization.** Try better prompts and retrieval first; fine-tune with LoRA adapters only if the gap remains. 5. **Serve it behind a standard API**, ideally OpenAI-compatible, so applications can switch models without code changes. 6. **Add guardrails and monitoring**, including PII and prompt injection detection, output validation and traces. 7. **Plan upgrades.** Re-run your test set on each new release before switching. If data residency drives the decision, our post on [how data residency rules shape banks' LLM use](/post/think-global-act-local-how-data-residency-regulations-influence-banks-llm-use) covers the regulatory side. ## How does Dynamiq support open-weight models? Dynamiq lets you deploy open models such as Llama, Gemma, Mistral and Qwen on managed inference with a vLLM runtime, each behind its own OpenAI-compatible endpoint with replica autoscaling. You can fine-tune LoRA adapters on a base model and serve them dynamically over that shared model instead of standing up a deployment per adapter. See [models and fine-tuning](/product/models) for details. Each step in a workflow can use a different model: an open-weight deployment for sensitive steps and a hosted provider elsewhere. The SDK supports 29 model providers, and the AI Gateway puts the most-used models behind one endpoint and one credential. The whole platform can be self-hosted on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes cluster or on-prem, and fully isolated setups are delivered with our engineers. ## FAQ ### What is the difference between open-source and open-weight LLMs? An open-weight model publishes its trained weights under a license that lets you run and usually modify it. An open-source model, in the strict sense, also makes available the code and enough information about the training data for others to study and rebuild it. Most popular "open" LLMs are open-weight. ### Are open-source LLMs free for commercial use? Often, but not always. Apache 2.0 and MIT licensed models allow commercial use with few conditions. Custom licenses, such as those for Llama and Gemma, allow commercial use but add conditions you must follow. Read the license for the exact model version. ### Are open-weight models as good as GPT or Claude? For many enterprise tasks, such as extraction, classification, summarization and answering questions from retrieved documents, current open-weight models perform well. On the hardest reasoning and coding tasks, the leading proprietary models are usually ahead. Test on your own cases before deciding. ### Is it safer to self-host an LLM? It keeps data inside your environment, which removes the risk of a third party processing it. It also makes you responsible for securing the model, the serving endpoints and the fine-tuning data, and for the guardrails a provider would otherwise apply. ### Can open-weight models be fine-tuned? Yes. Full fine-tuning changes all the weights; parameter-efficient methods such as LoRA train small adapters on top of a frozen base model, which is cheaper and lets you serve many adapters over one base model. --- # Balancing innovation and privacy: LLMs under GDPR URL: https://www.getdynamiq.ai/post/balancing-innovation-and-privacy-llms-under-gdpr Published: 2024-06-04 How to use LLMs under GDPR: lawful basis, data minimization, DPIAs, erasure and data residency, and where self-hosting and PII detection help. Yes, you can use large language models under the GDPR, as long as you can show a lawful basis for the personal data they process, keep that data to a minimum and honor people's rights over it. In practice, that means mapping every place personal data enters an LLM system, controlling where it is processed, and keeping records that prove what happened. None of this blocks innovation, but it has to be designed in from the first pilot. ## Does GDPR apply to LLMs? The GDPR applies whenever you process personal data of people in the EU, whether your company is established in the EU or offers services to people there. An LLM system processes personal data as soon as a prompt, a retrieved document or an output contains information about an identifiable person. It makes no difference that the model "only predicts text". The stakes are real. Under [Article 83 of the GDPR](https://eur-lex.europa.eu/eli/reg/2016/679/oj), the most serious infringements can be fined up to €20 million or 4% of worldwide annual turnover, whichever is higher. ## Where does personal data enter an LLM system? Most teams think about training data first. In an enterprise deployment, far more personal data flows through the system at run time. | Where it enters | Example | The GDPR questions it raises | | ----------------------------- | --------------------------------------------- | -------------------------------------------------------- | | Training and fine-tuning data | Support transcripts used to fine-tune a model | Lawful basis, compatible purpose, what the model retains | | Prompts | A customer types account details into a chat | Minimization, retention, transfer to a model provider | | Retrieved documents (RAG) | An HR assistant retrieves an employee's file | Access control, purpose limitation | | Uploaded files | A loan applicant uploads bank statements | Special categories of data, retention | | Agent memory | An assistant remembers a user's preferences | Transparency, erasure | | Logs and traces | Full prompts and outputs kept for debugging | Retention, access, security | | Outputs | A generated summary about a named person | Accuracy, automated decision-making | Map each of these for your use case before you pick a model. The map becomes the backbone of your records of processing and your impact assessment. ## Which GDPR principles matter most for LLMs? Six parts of the regulation come up in almost every LLM project: - **Lawful basis (Article 6).** Consent is only one of six bases. Many enterprise uses rely on contract or legitimate interests, which require a documented balancing test. Health, biometric and other special categories of data (Article 9) need an additional condition. - **Purpose limitation and data minimization (Article 5).** Use personal data only for the purpose you stated, and send the model only what it needs for the task. - **Storage limitation (Article 5).** Set retention periods for prompts, agent memory and traces, not only for source systems. - **Security (Article 32).** Encrypt data, restrict access by role and pseudonymize where you can. - **Automated decisions (Article 22).** Decisions with legal or similarly significant effects, such as a credit refusal, need meaningful human involvement and safeguards. - **Data protection impact assessments (Article 35).** A DPIA is required when processing is likely to result in a high risk to people. Large-scale processing of customer or employee data with a new technology often qualifies. Transparency runs through all of them. People should know when an AI system processes their data and why, which usually means updating privacy notices before launch. ## What have EU regulators said about AI models? In December 2024, the European Data Protection Board published [Opinion 28/2024](https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en) on personal data in AI models. Three points from it matter for anyone building on LLMs: 1. **Anonymity is assessed case by case.** For a model to count as anonymous, it should be very unlikely both to identify the people whose data trained it and to let someone extract that data through queries. 2. **Legitimate interest can work, with a three-step test.** You identify the interest, show the processing is necessary for it, and balance it against people's rights, including what they would reasonably expect. 3. **Unlawful training can follow the model.** A model developed with unlawfully processed personal data can affect the lawfulness of its deployment, unless the model has been duly anonymized. For deployers, the practical takeaway is due diligence: ask your model provider how it trained the model and what it does with your data. ## How does the EU AI Act fit in? The EU AI Act entered into force on August 1, 2024 and applies alongside the GDPR, not instead of it. The AI Act sets obligations by an AI system's risk level, while the GDPR keeps governing the personal data those systems process. Our [guide to the EU AI Act](/post/how-the-eu-ai-act-will-impact-your-business) covers its timeline and requirements. ## Should you use private LLMs or hosted APIs? Where the model runs decides who else processes your data and where. None of the options makes you compliant by itself; you still need a lawful basis, minimization and a way to handle rights requests. | Option | Where data is processed | What to put in place | | ---------------------------------------- | --------------------------------------------------- | ----------------------------------------------------------------- | | Hosted model API | The provider's infrastructure and regions | A data processing agreement, transfer safeguards, retention terms | | Hosted API with regional data residency | The provider's infrastructure in your chosen region | The same agreement, plus a check of which services stay in region | | Open-weight model on your infrastructure | Your cloud account or data center | Your own security, patching and capacity planning | Many enterprises combine them: an [open-weight model](/glossary/open-weight-model) in their own environment for the most sensitive data, and an approved hosted model for the rest. ## A GDPR checklist for LLM projects 1. **Map the data flows** for training, prompts, retrieval, files, memory, logs and outputs. 2. **Choose and document a lawful basis** for each purpose, with a balancing test where you rely on legitimate interests. 3. **Run a DPIA** before launch when the processing is likely to be high risk, and revisit it when the use case changes. 4. **Minimize what reaches the model.** Remove or mask personal data the task does not need. 5. **Set retention periods** for prompts, agent memory and traces, and enforce them. 6. **Control where processing happens,** with processor agreements and transfer safeguards for any provider outside your region. 7. **Build rights handling** that can find and erase a person's data everywhere it was copied, including memory and logs. 8. **Keep a person in the loop** for decisions with legal or similarly significant effects. 9. **Train the people** who build and operate the system on these rules. 10. **Keep records** of processing activities and of what each AI system did. ## How does Dynamiq support GDPR programs? Dynamiq is built for teams that must answer these questions for an auditor, and the platform covers SOC 2, HIPAA and GDPR, with a DPA available. - **Processing where you choose.** Run Dynamiq in our cloud or self-hosted in your own AWS, Azure, GCP, IBM Cloud or OpenShift environment, on any Kubernetes cluster or on-premises; see [where Dynamiq runs](/deployment/self-hosted). - **Your choice of model.** Use the model providers your data protection officer approved from the SDK, call the most-used models through one AI Gateway endpoint, or run open models on [managed inference](/product/models). - **PII detection before the model.** [Guardrail nodes](/product/guardrails) detect PII and prompt injection at the start of a workflow, so you decide what happens to flagged input. - **Memory you can erase.** Agent memory is scoped by user and session and can be deleted per user when a request comes in. - **A record of every run.** Each run is traced with its inputs, outputs and steps, which gives your DPIA and your auditors real evidence; see [observability](/product/observability). The [security and trust center](/security) covers our compliance, data protection and sub-processors in more detail. ## FAQ ### Can you put personal data in LLM prompts under GDPR? Yes, if you have a lawful basis for that processing, the data is necessary for the task, and the model provider is bound by a processing agreement with suitable safeguards. Minimize first: mask or remove personal data the model does not need, and set a retention period for stored prompts. ### Do you need a DPIA for an LLM project? You need one when the processing is likely to result in a high risk to people, which is common for large-scale processing of customer, patient or employee data with a new technology. Even when it is not strictly required, a DPIA is the clearest way to document your decisions. ### Does the right to erasure apply to LLMs? Yes. People can ask you to erase their personal data, and that includes copies in prompts, retrieved documents, agent memory and logs. Removing data from a trained model's weights is much harder, which is one reason to keep personal data out of training sets and rely on retrieval instead. ### Is a self-hosted LLM automatically GDPR compliant? No. Self-hosting controls where data is processed and removes a third party from the chain, but you still need a lawful basis, minimization, retention rules, security and a way to handle rights requests. ### Does the EU AI Act replace GDPR for AI systems? No. The AI Act regulates AI systems by risk level, and the GDPR continues to govern any personal data those systems process. Most enterprise AI projects in the EU have to satisfy both. --- # Guardrails for LLMs in banking: essential measures for secure AI use URL: https://www.getdynamiq.ai/post/guardrails-for-llms-in-banking-essential-measures-for-secure-ai-use Published: 2024-06-04 The guardrails banks need around LLMs and AI agents: input screening, output checks, limits on actions, human approvals and audit trails, with examples. Guardrails for LLMs in banking are the controls around a model or agent that decide what it may read, say and do. They screen input for personal data and prompt injection, restrict retrieval to what each user may see, check outputs against policy, require a person's approval before risky actions, and record every step for audit. Banks need all of these layers, because a filter on the model's answer does nothing to stop an agent from taking a wrong action. This guide explains the risks guardrails address, the types of guardrails and where each one sits, and how to put them in place in a regulated bank. ## Why do banks need guardrails for LLMs? A bank's LLM application touches customer data, makes statements customers may rely on, and increasingly acts in core systems. Each of those creates a specific risk: - **Data exposure.** Account numbers, card data and personal details can leak into prompts sent to external models, into logs, or into answers shown to the wrong person. - **Wrong answers that bind you.** A confident but incorrect statement about a fee, rate or policy is a conduct risk. - **Prompt injection.** Attackers hide instructions in messages, emails or documents that an agent reads, trying to override its rules. - **Excessive agency.** An agent with broad tool access can move money, change records or send messages nobody approved. - **Unfair outcomes.** Credit and pricing decisions influenced by a model must not discriminate, and adverse decisions need specific reasons. - **Regulatory exposure.** Supervisors expect traceability, human oversight and documented controls for any model that affects customers. These risks map closely to the industry's reference list, the [OWASP Top 10 for LLM applications](https://genai.owasp.org/llm-top-10/), which ranks prompt injection, sensitive information disclosure and excessive agency among the most serious. ## What happens without guardrails? Courts and tribunals already treat a chatbot's words as the company's words. In [Moffatt v. Air Canada](https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html), a Canadian tribunal ruled in 2024 that the airline was responsible for its website chatbot's incorrect explanation of its bereavement fare policy, and ordered it to compensate the customer. The airline argued the chatbot was responsible for its own statements; the tribunal rejected that. For a bank, the equivalent is an assistant that misstates an overdraft fee, promises a rate or gives what sounds like investment advice. Other public incidents, such as support bots coaxed into swearing at customers or into "agreeing" to absurd deals, show how quickly a missing guardrail becomes a reputational problem. ## What are the types of LLM guardrails? Guardrails sit at five points in the path from a user's request to an agent's action: | Layer | What it checks | Banking example | | ---------- | ------------------------------------------------------ | ------------------------------------------------------------------------------------------ | | Input | PII, prompt injection, off-topic or abusive requests | A customer pastes a full card number into chat, and the message is routed to a secure flow | | Retrieval | Which documents and records the agent may read | A branch employee's assistant cannot retrieve private banking files | | Output | Grounding, format, policy and tone | An answer about fees must cite the current fee schedule | | Action | Which tools the agent may call, and with what approval | A payment reversal waits for an operations manager to approve the exact amount | | Monitoring | Every step, verdict and cost, after the fact | Compliance reviews a flagged conversation step by step | Input and output checks are what most people mean by guardrails. For agents, the action layer matters most, because it is where a mistake stops being words and becomes a transaction. ## Which guardrails matter most for banks? - **Sensitive data detection.** Detect account numbers, card numbers, national IDs and other personal data before text reaches an external model or a log, and route flagged messages deliberately. - **Prompt injection detection.** Screen user messages and, just as important, the documents and emails an agent reads during KYC, onboarding or complaints handling. - **Grounded answers.** Answer only from approved sources, such as current product terms and policies, and cite them. - **No unsolicited advice or commitments.** Check outputs for investment advice, rate promises or statements outside the approved script, and hand those conversations to a person. - **Least-privilege tools.** Give each agent only the tools and data its task needs, with read-only access by default. - **Human approval on consequential steps.** Payments, account changes, filings and customer letters wait for sign-off, and the reviewer can edit the exact fields before approving. - **AI disclosure.** Under the EU AI Act, since 2 August 2026 chatbots must tell people they are talking to an AI system, not a human. - **Audit trail.** Keep a replayable record of what the agent read, which checks it passed, what it did and who approved it. ## How do you implement guardrails in a bank? 1. **Classify each use case by risk.** Separate internal assistants from customer-facing agents, and assistive tasks from ones that influence decisions such as credit. 2. **Write policies as testable rules.** "Never quote a rate that is not in the rate sheet" can be tested; "be careful with rates" cannot. 3. **Place detectors before the model.** Screen for PII and prompt injection on every input, including retrieved documents where the risk is high. 4. **Constrain tools and add approvals.** Limit tool permissions, set thresholds, and require sign-off where an action is hard to reverse. 5. **Validate outputs.** Check structure, allowed values and required citations before anything reaches a customer or another system. 6. **Test before release.** Build evaluation sets from real cases and red-team prompts, and run them on every change to prompts, models or tools. 7. **Monitor live traffic.** Trace every run, score a sample continuously and alert on failures and unusual cost. 8. **Review and improve.** Treat guardrail misses as incidents, add the failing case to the test set and fix the rule. ## How do guardrails fit bank regulation? Guardrails are how a bank shows regulators that its AI controls exist and work: - **Model risk management.** Validation, monitoring and documentation apply to LLM-based systems. In the US, the interagency guidance was revised in April 2026 ([SR 26-2](https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm)), replacing SR 11-7. - **EU AI Act.** Transparency duties, including chatbot disclosure, apply from 2 August 2026. Systems that assess the creditworthiness of individuals are high-risk, and after the 2026 AI Omnibus amendment their obligations, including human oversight and logging, apply from 2 December 2027. - **Data protection.** GDPR and similar laws require data minimization and a lawful basis for processing, which input screening and self-hosting help demonstrate. See our post on [LLMs under GDPR](/post/balancing-innovation-and-privacy-llms-under-gdpr). - **Operational resilience.** In the EU, DORA requires banks to manage risk from ICT providers, which includes AI platforms and model APIs. ## How does Dynamiq implement guardrails? In Dynamiq, [guardrails](/product/guardrails) are workflow nodes. A PII detector, a prompt injection detector and a policy detector screen text before an agent sees it, and each returns a verdict you branch on, sending clean messages to the agent and flagged ones to a refusal, a secure flow or a person. Detectors flag what they find; they do not rewrite the message, so the decision stays with you. Validators check outputs, including valid JSON, allowed choices and regex patterns. Approvals pause a tool call, or an agent's plan, until a reviewer signs off, and the reviewer can edit fields such as the amount or the body of a letter before approving. A rejection flows back to the agent as feedback it can act on. Every verdict and approval is recorded in the run's trace. The platform can be self-hosted in the bank's own cloud account or data center, with fully isolated setups delivered with our engineers, and Dynamiq has SOC 2, HIPAA and GDPR in place. ## FAQ ### What are LLM guardrails? LLM guardrails are controls that limit what a language model or agent can read, say and do. They include input screening, retrieval permissions, output validation, limits and approvals on actions, and monitoring of every run. ### What is the difference between input and output guardrails? Input guardrails check what goes into the model, such as personal data, prompt injection or off-topic requests. Output guardrails check what comes out, such as unsupported claims, wrong formats or statements that break policy. Agents also need action guardrails, which control the tools a model can call. ### Can guardrails stop prompt injection completely? No single check can. Combine an injection detector with least-privilege tools, approvals on consequential actions and monitoring, so that a successful injection still cannot do real damage. ### What should happen when a guardrail detects personal data? Route the message deliberately: to a secure flow, to a model hosted inside your environment, or to a person. Record the verdict in the trace so the decision can be reviewed later. ### Are banks required to use AI guardrails? Regulations rarely use the word, but they require what guardrails provide: data protection, human oversight, logging and documented controls. The EU AI Act, model risk management guidance and data protection law all expect banks to show these controls in practice. --- # Should enterprises consider implementing large language models? URL: https://www.getdynamiq.ai/post/should-enterprises-consider-implementing-large-language-models Published: 2024-05-23 Should your enterprise implement LLMs? Where they pay off, the risks and costs, when not to use them, and a step-by-step plan from first use case to scale. Yes. For most enterprises the question is no longer whether to use large language models (LLMs), but where they pay off and how to run them safely. LLMs work best today inside agents that use your tools and data, in processes with high volumes of text: customer service, document review, compliance preparation, internal knowledge and software engineering. The practical path is one measurable process at a time, with data controls, evaluations and human approvals in place before you scale. This guide covers where LLMs deliver value, when not to use them, the costs and risks, and a step-by-step implementation plan. ## Should your enterprise implement LLMs? Implement LLMs where three things are true: the work is mostly language (reading, writing, searching, classifying), it happens at volume, and you can measure the outcome. That describes a large share of work in banking, insurance, healthcare, telecom and the public sector. Do not use an LLM where it adds risk without value: - **Deterministic, stable processes.** If fixed rules already handle a task reliably, keep the rules; they are cheaper and fully predictable. - **Decisions you cannot explain or review.** Keep a person on decisions such as credit approvals, where the outcome must be justified. - **No access to the data.** An LLM that cannot reach your systems or documents will guess, and guesses are expensive. - **No way to measure success.** Without a baseline, you cannot tell whether the system helps. The [AI agent readiness assessment](/resources/ai-agent-readiness-assessment) helps you check whether a process and your team are ready. ## Where do LLMs deliver value in enterprises? | Function | What LLM agents do | Typical measure | | -------------------- | ------------------------------------------------------------------ | ---------------------------- | | Customer service | Resolve routine requests, hand complex ones to people with context | Resolution rate, handle time | | Back office | Match records, prepare exceptions and reports for approval | Cases per person, error rate | | Document review | Extract and check fields in contracts, claims and loan files | Review time per file | | Compliance | Prepare KYC and AML cases, track regulatory change | Case preparation time | | Knowledge work | Answer questions from internal documents with citations | Time to answer, accuracy | | Software engineering | Write code, tests and documentation for review | Cycle time | | Sales and marketing | Research accounts, draft content for review | Selling time, content cost | Customer service is often the first production use case because volumes are high and outcomes are easy to measure. An Asian neo-bank [automated about 85% of support inquiries](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) and went live in 30 days. ## What are the benefits and risks of LLMs for enterprises? | Benefit | Matching risk | How to manage it | | ---------------------------------------- | ---------------------------------- | ------------------------------------------------------------------------ | | Handles unstructured text at scale | Wrong answers stated confidently | Ground answers in your documents with citations; evaluate before release | | Automates multi-step work through agents | Agents acting beyond their remit | Least-privilege tools and approvals on consequential actions | | Works with your internal data | Data leaving your control | Self-host where needed; detect PII before text reaches a model | | Quick to prototype | Pilots that never reach production | Pick measurable use cases and plan for integration early | | Many models to choose from | Lock-in or sudden model changes | Keep models swappable behind one gateway | | Natural-language interface for everyone | Prompt injection and misuse | Input detectors, monitoring and clear usage policies | Regulation adds structure rather than a ban. In the EU, the [AI Act](/post/how-the-eu-ai-act-will-impact-your-business) requires chatbots to disclose that they are AI and sets stricter obligations for high-risk uses such as credit scoring, while GDPR continues to govern personal data. ## What does it cost to implement LLMs? - **Model usage:** per-token fees for hosted models, or GPU capacity for self-hosted ones. - **Platform:** orchestration, retrieval, guardrails, evaluation and tracing, built or bought. - **Integration:** connecting agents to your systems of record, usually the largest effort. - **People:** a product owner, engineers, domain experts to review outputs, and change management. - **Operations:** monitoring, evaluation runs and updates as models change. Measure cost per completed case, not cost per token. Our guide to [GenAIOps build vs. buy](/post/genaiops-for-enterprises-the-build-vs-buy-dilemma) covers the platform decision in detail. ## What are the risks of waiting? Waiting has a cost: competitors that automate support and operations serve customers faster at lower cost, and teams without hands-on experience take longer to judge vendors and risks. Rushing has a cost too: an unsupervised chatbot that misstates a policy can create legal and reputational damage. The answer is not speed or caution alone but a controlled first use case that proves value and builds the capability. ## How should an enterprise implement LLMs? 1. **Pick one process with a baseline.** Choose high volume, clear rules and a measurable outcome, and record current cost and quality. 2. **Check data readiness.** Confirm the agent can reach the documents and systems it needs, with permissions that match your roles. 3. **Choose the deployment model by data sensitivity.** Managed cloud, self-hosted in your cloud or data center, or fully isolated for the most sensitive work. 4. **Select models by testing.** Compare two or three models on your own cases for quality, latency and cost. 5. **Build the agent.** Give it instructions, tools, knowledge retrieval and memory, and keep the model swappable. 6. **Add guardrails and approvals.** Screen inputs for PII and prompt injection, validate outputs and require approval before consequential actions. 7. **Pilot against your KPIs.** Run real cases alongside the current process and compare. 8. **Scale with monitoring.** Trace every run, score live traffic, train the people who work with the agent, and reuse what works in the next process. The [AI agent ROI calculator](/resources/ai-agent-roi-calculator) helps you estimate the value of a first use case before you start. ## How does Dynamiq help enterprises implement LLMs? Dynamiq is one platform to build, run and govern LLM agents. Teams build on a visual canvas or with the open-source Python SDK on the same engine, connect to a large catalog of app integrations and your own databases, and ground agents in knowledge bases with permission-aware retrieval. Guardrails detect PII and prompt injection, approvals gate any tool step, evaluations run on test sets and live traffic, and every run is traced with cost and latency. The platform runs in Dynamiq Cloud or [self-hosted](/deployment/self-hosted) on AWS, Azure, GCP, IBM Cloud, Red Hat OpenShift, any Kubernetes cluster or on-prem, and our forward-deployed engineers can build the first use cases with your team. ## FAQ ### What is a large language model? A large language model is an AI model trained on large amounts of text to understand and generate language. In enterprises, LLMs usually run inside agents that combine the model with instructions, tools and company data to complete tasks. ### Should we build our own LLM? Rarely. Training a model from scratch is expensive and seldom necessary. Most enterprises use pre-trained hosted or open-weight models, ground them in company data with retrieval, and fine-tune only for narrow needs such as a specific output format. ### Is it safe to use LLMs with company data? It can be, with the right controls: a deployment model that matches the data's sensitivity, permission-aware retrieval, PII and prompt injection detection, approvals on consequential actions and full tracing. Check each vendor's data terms and certifications. ### How long does it take to implement an LLM use case? A prototype can be built quickly; production takes longer because of integrations, testing and security review. With a focused use case and experienced engineers, production in weeks is realistic: the neo-bank above went live in 30 days. ### Which departments should adopt LLMs first? Start where volume is high, rules are clear and outcomes are measurable: customer service, finance and operations, document review, compliance preparation, IT and HR service desks, and software engineering. --- # Generative AI and LLMs in banking: examples, use cases, limitations and solutions URL: https://www.getdynamiq.ai/post/generative-ai-and-llms-in-banking-examples-use-cases-limitations-and-solutions Published: 2024-05-10 How banks use generative AI and LLMs today, from customer service to KYC triage, what still limits adoption, and the controls that get agents into production. Banks use generative AI and large language models (LLMs) to answer customer questions, read and extract data from documents, triage compliance alerts, draft reports and write code. The main barriers are not model quality but data security, explainability and control. That is why production deployments pair agents with guardrails, human approvals and full audit trails, often running inside the bank's own infrastructure. This guide covers where generative AI works in banking today, with examples, the limitations that still hold banks back, and the controls that address them. ## How are banks using generative AI today? Banks have moved from chat pilots to agents that do work inside their systems. The first wave, in 2023 and 2024, was mostly internal assistants that summarized documents and drafted text. The current wave is agents that look up accounts, call internal APIs, fill forms and prepare decisions, with a person approving anything that commits money or reaches a customer file. The economic case is large. McKinsey estimated that generative AI could add [$200 billion to $340 billion in value a year](https://www.mckinsey.com/industries/financial-services/our-insights/capturing-the-full-value-of-generative-ai-in-banking) to the banking industry, mostly through productivity. Capturing that value depends less on the model than on how safely an agent can be connected to a bank's systems. ## What are the main use cases for LLMs in banking? ### Customer service Agents answer routine requests on chat and phone, such as card status, balances, fees and password resets, and hand anything that needs judgment to a person with the full conversation attached. Grounding answers in the bank's own policies keeps them accurate, and calling backend APIs lets the agent resolve a request instead of only describing how. For example, an Asian neo-bank [automated about 85% of support inquiries and went live in 30 days](/case-studies/automating-customer-support-at-scale-how-a-neo-bank-saves-1-5m-a-year-with-ai) with a single agent that uses tools, memory and a knowledge base, and escalates disputes and edge cases to its support team. ### Back-office operations Payments, lending and finance operations run on reconciliations, exception queues and reports. An agent can match failed transfers against ledger entries, draft the exception report and route it for approval, so the team reviews a prepared case instead of assembling one. Our guide to [back-office agents](/use-cases/back-office) covers the pattern in more detail. ### Document review, lending and underwriting Loan files, account-opening packs and trade documents are long and inconsistent. LLMs parse them, extract the fields a policy needs and check those fields against the bank's rules. Clean files move forward and exceptions go to an underwriter with the source page attached. Keep a person in the decision for credit. The EU AI Act classifies AI systems that evaluate the creditworthiness of individuals or set their credit score as high-risk, and those obligations apply from 2 December 2027. ### Compliance and financial crime Transaction monitoring and fraud scoring still run mostly on traditional machine learning models. Generative AI helps the analysts who work the alerts. An agent gathers customer data, screens adverse media, summarizes the case against policy and drafts the narrative, and an analyst decides. The same approach works for KYC refreshes and for tracking regulatory change across the sources a compliance team follows. See how this works in [compliance and financial crime](/use-cases/compliance-financial-crime). ### Knowledge assistants for employees Relationship managers, contact center staff and operations teams spend a large share of their day looking for answers in policies and product documentation. A knowledge assistant answers from internal documents with a citation for every claim, and permission-aware retrieval makes sure each employee only sees what their role allows. ### Software engineering Banks run large, old codebases. Coding assistants help engineers document legacy systems, write tests and migrate code, which is often the fastest internal win because the output is reviewed by an engineer before it ships. ## What does a banking agent look like in practice? Take a periodic KYC refresh for a small-business customer. Each step pairs work the agent does with a control the bank keeps: | Step | What the agent does | Control | | ------- | ---------------------------------------------------------------------- | ------------------------------------------------------------------------- | | Collect | Pulls the customer file, registry data and recent transactions | Read-only access, scoped to the case | | Screen | Checks names against adverse media and sanctions sources | Every source and result recorded in the trace | | Analyze | Compares ownership and activity with the last review and flags changes | PII detection before text reaches an external model | | Draft | Writes the review summary and a recommended risk rating | Output validated against a required template | | Decide | Sends the case to an analyst | The analyst approves, edits or rejects; nothing is filed without sign-off | The agent does the collection and drafting that used to take most of the analyst's time. The analyst keeps the decision, and the trace shows an auditor exactly what the agent read and wrote. ## What limits generative AI adoption in banking? - **Data security and residency.** Customer data sent to a third-party model API leaves the bank's control, and residency rules may forbid it outright for some data and jurisdictions. - **Hallucination.** An LLM can answer fluently and wrongly. In banking, a wrong fee, rate or policy statement is a conduct risk, not a typo. - **Explainability and model risk management.** Model risk teams expect LLM-based systems to be inventoried, validated and monitored like any other model. In the US, that framework was updated in April 2026 by revised interagency guidance, [SR 26-2](https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm), which replaced SR 11-7. - **Fairness.** Credit and pricing decisions must not discriminate, and adverse decisions need specific reasons a customer can understand. - **New attack surfaces.** Prompt injection, data leakage through outputs and agents with too much authority appear in the [OWASP Top 10 for LLM applications](https://genai.owasp.org/llm-top-10/). - **Third-party risk.** In the EU, the [Digital Operational Resilience Act](https://eur-lex.europa.eu/eli/reg/2022/2554/oj), which applies from 17 January 2025, requires banks to manage and document the risk from ICT providers, including AI vendors. - **Cost at scale.** Per-token pricing adds up quickly across millions of customer interactions. ## How do banks overcome these limitations? Each limitation has a practical control. None of them requires waiting for better models. | Limitation | Solution | | -------------------------------- | -------------------------------------------------------------------------------------------------- | | Data leaving the bank | Run the platform, and if needed open-weight models, inside the bank's cloud account or data center | | Hallucination | Retrieval over approved sources with citations, and evaluations before every release | | Explainability | A trace of every step: what the agent read, which tools it called, what it wrote, what it cost | | Prompt injection and PII | Detectors that screen input before it reaches a model | | Agents acting beyond their remit | Least-privilege tools and human approval on any tool step that commits money or data | | Model risk management | Versioned workflows, test datasets and real-time evaluations on live traffic | | Cost | Routing simple tasks to smaller models and measuring cost per run | The pattern is consistent: keep data in your environment, make every run reviewable, and put a person at the points where a mistake would be expensive. Our guide to [guardrails for LLMs in banking](/post/guardrails-for-llms-in-banking-essential-measures-for-secure-ai-use) goes deeper on each control. ## How does Dynamiq help banks deploy generative AI? Dynamiq gives banks one platform to build, run and govern agents for customer service, document review and compliance. It runs in Dynamiq Cloud or self-hosted in the bank's own AWS, Azure, GCP or IBM Cloud account, on Red Hat OpenShift, on any Kubernetes cluster or on-prem, and fully isolated setups are delivered with our engineers. Guardrails detect PII and prompt injection before a model sees the text, approvals with editable fields gate any tool step, and every run is traced with cost and latency. Knowledge bases use permission-aware retrieval, so an agent only sees what the asking employee may see. Dynamiq has SOC 2, HIPAA and GDPR in place. Read more about Dynamiq for [financial services](/industries/financial-services). ## FAQ ### How are banks using generative AI? Mostly for customer service, back-office operations, document review in lending, compliance and KYC case preparation, employee knowledge assistants and software engineering. Most production deployments keep a person approving any decision that affects a customer or moves money. ### What are the risks of using LLMs in banking? The main risks are customer data leaving the bank, hallucinated answers, prompt injection, agents acting beyond their remit, unfair outcomes in credit and pricing, and weak auditability. Each has a control: self-hosting, grounded retrieval, input detectors, approvals, fairness testing and full traces. ### Can banks use public LLM services such as ChatGPT? Most banks do not let staff paste customer data into consumer chatbots. They use enterprise model agreements with data-use restrictions, or open-weight models hosted in their own environment, behind a platform that logs and controls every call. ### Is AI for credit decisions high-risk under the EU AI Act? Yes. AI systems that evaluate the creditworthiness of individuals or establish their credit score are listed as high-risk, with fraud detection excluded. After the 2026 AI Omnibus amendment, those high-risk obligations apply from 2 December 2027. ### Should banks host LLMs on-premises? Host in your own environment when data residency rules or internal policy require it, or when volumes make owned inference cheaper. Many banks mix both: hosted frontier models for low-sensitivity tasks and self-hosted open-weight models for customer data, behind one platform with the same controls.