SR 26-2 and AI Supervision: What State-Chartered Banks Must Evidence

What SR 26-2 is, and what it is not

SR 26-2 and its attachment, Supervisory Guidance on Model Risk Management, are dated April 17, 2026. The OCC, the Federal Reserve Board and the FDIC issued the revised guidance. It updates the model risk management approach in place since SR 11-7, reflecting supervisory experience and industry feedback, and it emphasizes a risk-based approach tailored to a bank's model risk profile, size and complexity.

Two features shape how a risk leader should read it. First, the guidance states that it does not set enforceable standards or prescriptive requirements, and that non-compliance with it will not result in supervisory criticism. Second, in a May 1, 2026 speech, Governor Bowman said the agencies amended the guidance to clarify that it does not apply to generative or agentic AI.

The practical consequence is that a comparison framed as one AI supervisory document against another starts from a mistaken premise. SR 26-2 governs model risk management for traditional models. Whatever applies to generative and agentic AI, at the federal level or elsewhere, has to be assessed separately and recorded as a deliberate decision. Applicability to a particular charter or Federal Reserve membership status should be confirmed with the bank's primary regulator.

Evidence crosswalk by artifact type

Organizing by evidence type is more useful than organizing by regulatory text. The table shows what SR 26-2 describes for traditional models and what an examiner could reasonably ask to see. For generative and agentic AI, the right-hand reading is that the artifact is the bank's own commitment, not an SR 26-2 requirement.

• Governance and policy: SR 26-2 says effective policies define risk management expectations and set a framework for assessing the magnitude of model risk. Artifact: approved policy tying risk tiering to the intensity of practices, plus a documented statement of which framework governs generative and agentic AI.
• Model information and inventory: The guidance notes that banks commonly maintain a comprehensive set of information for models under development or in use. Artifact: a per-model information set, with generative and agentic systems listed separately so scope is not ambiguous.
• Risk-based tailoring: Practices vary among banks, and each is responsible for practices appropriate to its risks. Artifact: written rationale showing how risk profile, size and complexity set the depth of controls.
• Ongoing monitoring: An effective plan may include regularly assessing model limitations at development and over time, with procedures for responding to issues before and after approval. Artifact: monitoring plans and issue-response records showing the procedures were used.
• Internal audit: Internal audit generally evaluates whether practices are rigorous and effective and whether policies are implemented accordingly. Artifact: evidence that lets audit test implementation, not only the existence of documents.
• Vendor-supplied AI: Treatment depends on whether the system is a traditional model or generative or agentic AI. Artifact: a documented classification of each vendor-supplied system and the framework applied to it.

Documentation versus runtime evidence

The SR 26-2 evidence types are mostly documentary: policies, information sets, monitoring plans and audit evaluations. Nothing in that description calls for runtime policy-enforcement records or records of AI interactions and tool calls. That is an observation about the guidance, not a gap in it, because generative and agentic AI sit outside its scope.

The gap appears one step later. If a bank applies its own risk framework to generative or agentic AI, the question an examiner or internal auditor is likely to ask is whether the controls in that framework actually operated in production. A policy that prohibits sending certain data to a model, or requires human approval before an agent takes an action, is only documentation until there is a record that the rule was applied to a specific interaction.

This is the distinction worth building the checklist around. Artifacts such as policy definitions, tiering rationale and monitoring plans can be satisfied through documentation. Evidence that a control operated, what an AI system was allowed to do, and what it did, can only be produced by records generated at the point of execution. Reviewing behavior after the fact cannot reconstruct a decision that was never recorded when it was made.

Governance tradeoffs and open decisions

Because SR 26-2 is non-enforceable, the compliance question is less about meeting a checklist and more about whether the bank can show that its chosen practices are appropriate and effective for its risks. The guidance itself places that responsibility on each bank. Examiners may still evaluate that judgment, so the rationale needs to be written down and consistent with what runs in production.

Federal Reserve leadership has also said that banks currently rely on existing risk-management frameworks to guide AI use, and that the Fed should assess whether its supervisory guidance is fit for the future. Expectations for generative and agentic AI may therefore change. Evidence designed around a specific document's wording is more fragile than evidence designed around what the bank can demonstrate about system behavior.

The main tradeoff is effort. Maintaining separate inventories and evidence sets for traditional models and for generative and agentic AI avoids scope confusion, but it adds ownership overhead. A shared evidence structure with a clear scope flag per system usually limits duplicated work while keeping the SR 26-2 boundary visible.

Where runtime enforcement fits

Trussed is the control, provenance and evidence infrastructure for consequential AI. It enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. Enforcement outcomes include allow, redact, block, rate limit and escalate for human approval. Those decisions are enforced at runtime, not reviewed after the fact, and the enforcement itself generates audit-ready evidence, mapped to frameworks including NAIC, NYDFS, HIPAA, the EU AI Act and NIST.

Insurance is the regulated financial sector where Trussed is deepest, and the platform has run continuously in a regulated enterprise production environment for more than 18 months. It does not cover model lineage, dataset provenance or transaction-level governance. For a bank risk leader, the relevance is narrow and specific: it addresses the last checklist row, durable records of what AI systems were allowed to do and did, and leaves the documentary rows to the bank's own model risk program.