Most GenAI pilots in regulated institutions never reach production. The business team believes the assistant works. The risk team has no evidence it can sign.
The gap is real. Traditional model validation asks whether a model's score is accurate and stable. A generative AI system gives a different answer to the same question, has no single accuracy number, and can fail with a confident, fluent, wrong sentence.
This guide is for model risk, compliance and business leaders at banks and credit unions who have been asked to approve a GenAI use case. It covers what changes, what evidence to assemble, and how to keep it current after launch.
- Traditional validation does not fit GenAI: there is no single accuracy number, the output varies, the system is larger than the model, and it changes without anyone touching it.
- SR 26-2 puts generative and agentic AI out of scope, so the bank sets its own standard. The most defensible one is the standard it already runs for traditional models, adapted.
- Validate the whole system: model, prompts, retrieval sources, guardrails and escalation paths.
- Build the approval pack around six items: inventory entry and owner, materiality tier, test results against agreed thresholds, independent challenge, known limits and controls, and a monitoring plan.
- Approval is the start of the evidence. Monitor live traffic against the approved version and turn confirmed failures into new test cases.
Why generative AI breaks traditional model validation
Four properties of a GenAI system change the validator's job.
- There is no single accuracy number. A credit score can be backtested against defaults. A summary of a loan file has no ground truth to score against. Testing has to cover groundedness, policy adherence, fair lending, personal data leakage and safe escalation, each against agreed thresholds.
- The output varies. The same prompt can return different text on different runs. A test that passes once proves little. Tests need enough repeated runs to show how often a failure occurs.
- The system is larger than the model. Prompts, retrieval sources, guardrails and the vendor's foundation model all shape the answer. Validating the model alone leaves most of the system untested.
- It changes without anyone touching it. Vendors update models. Source documents change. A system that passed review in March can fail quietly by June.
None of this makes GenAI impossible to approve. It makes the approval evidence different.
Where the rules stand
In April 2026, the Federal Reserve, the OCC and the FDIC issued SR 26-2, which replaced SR 11-7. Footnote 3 of the attachment says generative AI and agentic AI models "are not within the scope of this guidance." It adds that a bank's "risk management and governance practices should guide the determination of appropriate governance and controls" for systems the guidance does not cover.
So the bank sets its own standard. The OCC bulletin says the agencies plan a request for information on model risk management that considers banks' use of generative and agentic AI. Our SR 26-2 guide covers what changed and what it means for GenAI in detail.
The most defensible standard is the one your team already runs for traditional models, adapted to the four properties above.
What changes between a traditional model and a GenAI system
| Question | Traditional model | GenAI system |
|---|---|---|
| What is validated | The model and its inputs. | The whole system: model, prompts, retrieval sources, guardrails and escalation paths. |
| How it is tested | Backtesting against known outcomes. | Curated scenario sets scored for groundedness, policy adherence, fair lending and data leakage, with expert review. |
| What a result looks like | One metric, such as accuracy or stability. | Several scores with thresholds, plus the rate of critical failures. |
| What triggers a re-review | A model change or a scheduled review. | A change to the model version, prompt, retrieval source or guardrail, and drift seen in production. |
| Vendor dependence | Sometimes a vendor model. | Almost always a third-party foundation model, with no access to code, data or method. |
The approval evidence pack
A committee that approves a GenAI launch will want the same answers it asks of any model. Build the pack around these six items.
- Inventory entry and owner. What the system is, which prompts, retrieval sources and vendor model it uses, and who is accountable. SR 26-2 describes an inventory with "sufficient information to understand model risks" (Section VI).
- Materiality tier. A customer-facing collections agent warrants more rigor than an internal search assistant. SR 26-2 ties rigor to purpose and exposure (Section III). Apply the same logic here.
- Test results against agreed thresholds. Scores for each risk, the datasets used, and the critical-case failure rate beside the overall score. Keep the tests so they can be rerun.
- Independent challenge. A record that someone independent of the builders reviewed the results, and that business and compliance experts rated real scenarios.
- Known limits and controls. What the system must not do, which guardrails enforce it, and when it hands off to a person.
- A monitoring plan. What is tracked after launch, who reviews it, and what triggers a re-review.
For a deeper look at the testing step, read our LLM evaluation guide. For how a model can score outputs at scale, see LLM-as-a-judge.
Monitoring after launch
Approval is the start of the evidence, not the end. SR 26-2 asks whether a model is "performing as expected given potential changes in products, exposures, activities, clients, data relevance, or market conditions" (Section V). For GenAI, that means watching live traffic against the approved version.
- Track the same scores used in approval on a sample of production conversations.
- Alert when a score moves from the approved baseline, for example after a vendor model update.
- Turn confirmed failures into new test cases, so the next review starts with more evidence than the last.
- Log every change to the model version, prompt, retrieval source and guardrail, with who approved it.
A checklist for your next model risk committee
- The system is in the inventory, with a named owner and its vendor model listed.
- It has a materiality tier and a reason for that tier.
- Tests cover groundedness, policy adherence, fair lending, data leakage and escalation, with thresholds set before testing.
- Results show critical-case failures, not only an overall score.
- An independent reviewer challenged the results, and the record shows who and when.
- Guardrails and handoff rules are documented.
- A monitoring plan names the scores, the reviewer and the re-review triggers.
- Prompts, retrieval sources and model versions are versioned.
How GenGuardX supports the evidence
GenGuardX gives business and risk teams one shared environment for this work. Inventory management registers datasets, models, prompts, RAGs and agents, and lineage tracking shows how they connect. Standardized evaluations are versioned and re-runnable, and Human Integrated Testing lets subject matter experts rate real scenarios.
The Governed Workbench adds role-based access and approval workflows, so the record shows who challenged what and who signed off. Monitoring for drift compares live traffic with the approved version.
What are the trade-offs of GenAI model risk management?
Governing a GenAI system like a model gives the committee evidence it can defend, and it costs setup time, reviewer hours and slower changes.
| What you gain | What it costs |
|---|---|
| An approval tied to one version: the committee knows which model, prompts and retrieval sources it signed off. | Every change to the model version, prompt, retrieval source or guardrail goes back to review, so updates ship more slowly. |
| Test results against thresholds set before testing, with critical-case failures shown. | Curated scenario sets and agreed thresholds take setup time before the first approval. |
| A record of independent challenge and expert ratings. | Independent reviewers and business and compliance experts spend hours rating scenarios, time taken from other work. |
| Monitoring that catches drift, for example after a vendor model update. | Someone has to review alerts and sampled conversations on a schedule, for as long as the system runs. |
| Confirmed failures become new test cases. | The suite grows, and each rerun takes longer and adds model costs. |
| One shared record for business and risk teams, if you use a platform such as GenGuardX. | A platform adds license cost, integration and setup work, and training for reviewers. |
To keep the costs down, scale rigor to the materiality tier, rerun one saved test suite at every re-review, and reserve full independent review for changes that affect customer outcomes.
FAQ
What is model risk management for generative AI?
It is the work of identifying, testing, approving and monitoring the risk that a generative AI system gives wrong, unsafe or unfair outputs. For a bank, it means applying the discipline of model risk management (inventory, independent challenge, testing, monitoring and documentation) to systems that produce language instead of a score.
Does SR 26-2 cover generative AI model risk?
No. Footnote 3 of the SR 26-2 attachment states that generative AI and agentic AI models "are not within the scope of this guidance." It says the bank's own risk management and governance practices should guide controls for systems the guidance does not cover.
Who should validate a generative AI system?
Someone independent of the team that built it. SR 26-2 describes effective challenge as "critical analysis conducted by objective experts" with "sufficient independence to maintain objectivity." For generative AI, add business and compliance experts who can judge whether a fluent answer is correct.
How often should a generative AI system be revalidated?
SR 26-2 sets no annual floor for in-scope models. Timing and frequency "vary based on model purpose, model methodology, frequency and scope of model changes." For generative AI, a practical trigger is any change to the model version, prompt, retrieval source or guardrail, plus a scheduled review for drift.
Last reviewed: November 15, 2026. This article summarizes public guidance and is not legal advice.
See how GenGuardX builds GenAI approval evidence for model risk teams.