How banks govern generative AI: key risks and controls
Banks govern generative AI by inventorying every use, rating its materiality, and applying controls proportionate to that rating: human review of outputs used in decisions, retrieval from approved sources, input and output filtering against prompt injection, contractual bans on provider training with bank data, and monthly accuracy sampling. SR 26-2 excludes generative AI, so these controls sit in an AI policy rather than model validation.
Generative AI is already inside drafting, summarisation, coding assistants and customer chat. The risks are familiar categories in a new form: wrong outputs, data leakage, misconduct and third-party dependence.
What are the main generative AI risks for financial institutions?
| Risk | What can go wrong | Best-practice controls |
|---|---|---|
| Hallucinated or inaccurate outputs | Confident but false statements or figures are relied on by staff or customers | Competent human review before use in decisions; answers drawn from approved sources with citations; accuracy tested before launch and sampled monthly |
| Prompt injection and jailbreaks | Hidden instructions in emails or documents make the model reveal data or ignore rules | Input filtering and output screening with logging; least-privilege access to data and systems; red-team testing before launch and after changes |
| Leakage of confidential or personal data | Data exposed through prompts, outputs, logs or provider retention | Contract bans provider training on your data; masking of personal data; access-controlled, monitored prompt logs |
| Misleading customer communication | A chatbot gives advice or promises that breach conduct rules | Disclose AI use and offer a human; guardrails restrict topics and products; monthly compliance sampling of conversations |
| Shadow AI | Staff use public AI tools with customer data | Approved tools list; DLP and web filtering on unapproved AI services; a safe enterprise alternative |
Does model risk guidance apply to generative AI?
Not directly in the United States: SR 26-2 places generative and agentic AI outside its scope. The FSB sound practices (October 2026) and the MAS Guidelines on AI Risk Management (October 2026) do cover all AI, including generative AI, with expectations on inventory, materiality, data governance, explainability, human oversight, testing and third-party risk. Most institutions therefore govern generative AI through an AI policy that borrows validation-style testing where materiality is high.
Which generative AI uses should be rated high risk?
- Anything that drafts or sends customer communications without human review.
- Uses that feed credit, pricing, claims or AML decisions.
- Tools connected to internal systems that can take actions.
- Uses processing personal or confidential data through an external provider.
What evidence will supervisors and auditors ask for?
Expect requests for the AI inventory entry, the materiality rating and its reasoning, test results before launch, the monitoring report, the human review records and the provider contract terms on data use. If these exist only in email threads, put them in one register: the evidence is what makes the controls real.
Primary sources (reviewed 17 Oct 2026)
- Financial Stability Board: Sound Practices for Responsible Adoption of Artificial Intelligence, final report (15 October 2026)
- Monetary Authority of Singapore: Guidelines on Artificial Intelligence Risk Management (7 October 2026)
- NIST AI 600-1: Generative Artificial Intelligence Profile (July 2024)
- Federal Reserve, OCC and FDIC: SR 26-2, Revised Guidance on Model Risk Management (17 April 2026)
This page explains supervisory guidance in plain language. It is not legal advice; confirm how each text applies to your institution and jurisdiction.
Frequently asked questions
Can banks use ChatGPT or similar tools?
Many do, through enterprise versions with contractual data protections, approved use cases and controls on what data can be entered. Public consumer tools with customer data are a common source of leakage and should be blocked or restricted.
What is the single most effective control for generative AI?
For decisions and customer communication, competent human review of outputs before use. It does not scale to every use case, which is why materiality rating decides where it is mandatory.
Put these controls to work this month
Start on the cloud risk management software within 24 hours, or ask for training or advice tailored to your institution.