Finance
Generative AI and LLMs in banking: examples, use cases, limitations and solutions
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.
Vitalii DukFounder and CEO7 min readUpdated 
In short
Banks use generative AI and LLMs most for customer service, back-office work, compliance triage, employee knowledge assistants and software engineering. What slows adoption is rarely model quality; it is data security, explainability and control. Banks that run agents in their own infrastructure, with guardrails, approvals and traced runs, are moving from pilots to 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 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 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 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.
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, 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.
- Third-party risk. In the EU, the Digital Operational Resilience Act, 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 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.
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.
