Model Risk Management for Business AI After SR 26-2
US banking supervisors replaced SR 11-7 in April 2026 and left generative and agentic AI out of scope. What the guidance says, and how to apply its ideas to AI systems anyway.
Answer first
Model risk management (MRM) is the discipline of knowing which models you rely on, how wrong they could be, and what you do about it: inventory, validation by people who did not build the model, ongoing monitoring, and governance scaled to how much each model matters. On April 17, 2026 the Federal Reserve, the OCC and the FDIC issued revised model risk guidance, published by the Federal Reserve as SR 26-2, which replaces SR 11-7 (2011) and SR 21-8 (2021). The new guidance explicitly says generative and agentic AI models are not within its scope. Its principles are still a useful starting point for governing those systems.
What changed in 2026
SR 26-2 supersedes two earlier letters: SR 11-7, the 2011 model risk guidance that most bank MRM programs were built on, and SR 21-8, the 2021 interagency statement on models supporting Bank Secrecy Act and anti-money laundering compliance. The Federal Reserve's letter says it is expected to be most relevant to banking organizations with over $30 billion in total assets that it regulates.
The attachment is short and principles-based. Ideas that carry over directly to AI work:
- Materiality. Model risk reflects a model's inherent risk in the context of its exposure and purpose. Models of higher materiality warrant more comprehensive and rigorous oversight; the guidance allows that some models may be deemed immaterial and handled mainly by identifying and monitoring them.
- Effective challenge. Critical analysis by objective, qualified people throughout the model lifecycle, from development to ongoing monitoring.
- Validation. Identifying limitations and errors, and clarifying appropriate use and whether corrective action is needed.
- Ongoing monitoring. Checking whether a model still performs as expected as products, data and conditions change, with overlays, adjustment or redevelopment when it does not.
- Inventory and documentation. Keeping a comprehensive set of information about models in development or in use.
- Vendor models. The guidance notes that banks often do not receive a vendor's code, data or methodology, and says model risk management principles remain applicable anyway.
The scope line for generative and agentic AI
The attachment states that generative AI and agentic AI models are novel and rapidly evolving and are therefore not within the scope of the guidance. It says the principles apply to traditional statistical and quantitative models and to non-generative, non-agentic AI models, and that an organization's risk management and governance practices should guide controls for systems the document does not cover.
Summaries of new guidance are not always accurate, so if you are in a regulated bank, read the attachment yourself and confirm your classification with your compliance function.
Applying MRM ideas to LLM and agent systems
Outside the guidance's scope does not mean outside governance. A workable approach:
- Put every AI system in the same inventory as your other models, with a flag for generative or agentic behaviour.
- Rate materiality by exposure and purpose: who is affected, what decisions or actions depend on the output, and what a wrong output costs.
- Replace "validate the equations" with "validate the behaviour": task-specific evaluation sets, adversarial tests for prompt injection and unsafe actions, and checks of retrieval sources.
- Have someone outside the build team review the evaluation design and results, which is the effective-challenge idea applied to AI.
- Monitor in production for quality drift, refusal changes, tool-error rates and cost, and set thresholds that trigger review.
- For vendor models, test the system you deploy (your prompts, tools and data with their model), since you will not see the vendor's internals.
- Re-review on material change: a new model version, new tools, new data or a new user group.
The NIST AI RMF and its Generative AI Profile (NIST AI 600-1) give a vocabulary for generative-specific risks you can fold into this process.
What to measure
- Share of AI systems in the inventory with a current materiality rating.
- Share of high-materiality systems with an independent review in the last cycle.
- Evaluation pass rate per release and the number of releases shipped with exceptions.
- Time from a monitoring alert to a decision (accept, adjust or roll back).
Decision framing
MRM is often seen as a brake. Done in proportion to materiality it does the opposite: low-materiality assistive tools get a light path, and scarce review effort goes to the systems where a wrong output matters. That is the case to make internally, framed around faster approvals for low-risk use and better evidence for high-risk use, not around guaranteed loss avoidance.
When to go further
If you are a bank, insurer or other regulated firm, treat this as a starting point and follow your regulator's current expectations. Because the guidance leaves generative and agentic AI to each organization's own governance, expect supervisors and auditors to ask how you made that call, and keep the reasoning documented with each inventory entry.