SR 11-7 and Generative AI: What the 2026 Supersession Actually Changed
SR 11-7 is no longer in effect: Federal Reserve SR letter 26-2 (April 17, 2026) supersedes and replaces it, and OCC Bulletin 2026-13 rescinded the OCC's parallel 2011 guidance the same day. The twist that matters for this page's question: the revised joint guidance expressly excludes generative and agentic AI from its scope and tells organizations their own risk management and governance practices should govern such tools, with an interagency request for information announced. As of 2026-08-28, there is no dedicated model-risk framework for generative AI — this page maps the guidance's enduring disciplines onto one you can adopt as your own.
What changed, from the primary sources
01SR 11-7 was the reference for fifteen years
- What happened:
- The Federal Reserve's Guidance on Model Risk Management (SR letter 11-7, April 4, 2011) and its OCC twin (Bulletin 2011-12) defined how supervised banking organizations governed models: development, validation, effective challenge, inventory, governance. Nearly every model-risk program in the industry was built on its vocabulary.
- Why it matters:
- Which is why pages still describing SR 11-7 as the operative guidance are now wrong — and why what replaced it, and what the replacement deliberately left out, both matter.
02April 17, 2026: superseded and replaced
- What happened:
- Federal Reserve SR letter 26-2, Revised Guidance on Model Risk Management, states that it supersedes and replaces SR letter 11-7 and SR letter 21-8, the 2021 Interagency Statement on Model Risk Management for Bank Systems Supporting Bank Secrecy Act/Anti-Money Laundering Compliance. The same day, OCC Bulletin 2026-13 rescinded OCC Bulletin 2011-12, the Model Risk Management booklet of the Comptroller's Handbook, Bulletin 1997-24 on credit scoring models, and Bulletin 2021-19 — SR 21-8's OCC twin. The attached guidance is joint: Board of Governors, Federal Deposit Insurance Corporation, and Office of the Comptroller of the Currency.
- Why it matters:
- The supersession is total, not partial. Citing SR 11-7 as current guidance in a model-risk document is now a defect an examiner can spot on page one.
03The revised guidance is risk-based and narrower in reach
- What happened:
- The 2026 guidance sets forth a risk-based approach tailored to an organization's model risk profile and the size and complexity of its operations, with materiality — model exposure together with model purpose — driving the rigor of oversight. It says it is expected to be most relevant to banking organizations with over $30 billion in total assets, with smaller organizations generally excluded consistent with a tailored supervisory approach, and it states that it does not set forth enforceable standards or prescriptive requirements — while noting, in the same footnote, that supervisory action may still result for violations of law or unsafe or unsound practices stemming from insufficient management of model risk.
- Why it matters:
- Model risk management stopped being one-size-fits-all in the regulators' own framing. The disciplines survived; the presumption that every model everywhere gets full-dress treatment did not.
04Generative and agentic AI are expressly out of scope
- What happened:
- Footnote 3 of the joint guidance: generative AI and agentic AI models are novel and rapidly evolving, and as such are not within the scope of the guidance. The same footnote adds that a banking organization's risk management and governance practices should guide the determination of appropriate governance and controls for tools not covered, and that the principles described apply to traditional statistical and quantitative models and non-generative, non-agentic AI models.
- Why it matters:
- This is the sentence the target query turns on. The guidance that replaced SR 11-7 deliberately does not govern generative AI — it hands the control decision back to each organization's own governance, by name.
05A request for information is the announced next step
- What happened:
- OCC Bulletin 2026-13 states the agencies plan to issue a request for information addressing model risk management generally and considering, in particular, banks' use of AI — including generative AI, agentic AI, and AI-based models.
- Why it matters:
- As of the last-reviewed date of this page, there is no dedicated interagency model-risk framework for generative AI: the old guidance is gone, the new one excludes it, and the RFI has not produced successor guidance. Teams building today are building against their own governance, informed by the enduring disciplines mapped below.
The control map
The disciplines below are the revised guidance's own — its sections on development and use, validation, monitoring, governance, and vendor products — translated to generative-AI deployments. The framing is deliberate and honest: generative AI sits outside the guidance's scope, so nothing here is a cited obligation. It is the map an organization adopts as its own governance, which is precisely what the guidance's footnote 3 says should happen for tools it does not cover.
| Discipline | What the guidance describes (in-scope models) | The generative-AI translation |
|---|---|---|
| Model inventory and documentation (Governance and Controls) | Maintaining a comprehensive set of information for models under development or in use is described as common industry practice — an effective inventory carries sufficient information to understand model risks at the individual and aggregate levels, supported by documentation that sustains continuity of operations. | Inventory every generative model in use — including vendor-hosted ones — with the exact model and version, its deployment surfaces, and its owner. Per-request logs that record model identity at request time are what keep the inventory honest after model swaps. |
| Model development and model use | Development begins with a clear statement of purpose aligned to intended use; testing evaluates whether the model performs as intended; using a model beyond its intended purpose introduces additional uncertainty and calls for additional analysis and review of controls. | A written purpose per deployment (what this assistant is for, and not for), evaluation sets run before rollout, and system prompts treated as development artifacts — versioned, reviewed, and re-tested when they change. Use outside the stated purpose is a change, not a shortcut. |
| Validation and effective challenge | Critical analysis by objective experts with expertise, sufficient independence, and the organizational standing to effect change; in validating conceptual soundness, evaluating theoretical construction may be important for some models, while other assessments — such as interpretability measures or benchmarking to other models — may be more practical for others. | Pre-deployment evaluation and adversarial testing — including prompt-injection and data-exfiltration probes — run by someone other than the team that built the deployment, with authority to block rollout. The guidance's own recognition that benchmarking may be more practical than theoretical evaluation fits generative systems, whose internals resist direct inspection. |
| Outcomes analysis and ongoing monitoring | Comparing outputs to real-world outcomes against established performance thresholds; monitoring whether the model performs as expected given changes in products, activities, data relevance, or conditions; deterioration may warrant overlays, adjustment, or redevelopment. | Continuous output-quality review against the evaluation set, drift checks when the underlying model or its version changes, and per-request logging as the evidence base. Which logging mode the records use — full, redacted, or hash-only — is its own decision with its own failure modes. |
| Governance, roles, and accountability | Clear roles and responsibilities with well-defined accountability, attention to conflicts of interest between development and validation, and policies that define expectations commensurate with model risk. | A named owner per generative deployment, separation between whoever ships prompts and whoever reviews outputs, and policy decisions (allowed, blocked, flagged) recorded in the same trail as the requests they governed. |
| Vendor and other third-party products | Vendor products present unique challenges because organizations may not receive the underlying code, data, or methodology; sound practice includes understanding the vendor model's conceptual soundness and performance, ongoing monitoring, and documenting adjustments made in customization. | This is the generative-AI situation in miniature — nearly every deployment rides a model whose weights, data, and training you cannot inspect. The response is contractual and evidentiary: vendor due diligence on the subprocessor chain, retention and no-training terms in writing, and monitoring that detects when the vendor changes the model underneath you. |
Where the evidence layer comes from
Four of the six rows land on the same operational floor: per-request records. Inventory honesty, outcomes monitoring, governance accountability, and vendor change detection all draw on a log that says who sent what, to which model and version, and what came back. The record shape is the audit log schema template — its model_id field captured at request time is what makes model swaps visible — and whether those records carry full content, redacted fields, or hashes is the three-mode decision the zero retention and audit logs guide maps. For the vendor row, the AI vendor due-diligence questionnaire is the due-diligence instrument: subprocessor chain, retention terms, and the contractual commitments that substitute for the code and data a vendor model never shows you.
How PrivateStack fits this map
PrivateStack supplies the evidence layer several rows of the map rest on: every request is logged with user identity, timestamp, and the model that actually served it; audit logs export date-ranged and machine-readable; the public model catalog documents the lineup; and the Enterprise deployment runs the data plane inside the customer's own AWS account. The honest limits: model risk management is a program your organization runs — evaluation, challenge, governance, and the judgment calls between them are yours, no platform performs them for you, and nothing on this page is supervisory guidance or legal advice. The full control set is on the security page, and the audit trail requirements checklist covers the ten controls the evidence layer should carry.
Questions people actually ask
- Is SR 11-7 still in effect?
- No. Federal Reserve SR letter 26-2 (April 17, 2026) states that it supersedes and replaces both SR letter 11-7 and SR letter 21-8 (the 2021 interagency statement on model risk management for Bank Secrecy Act/Anti-Money Laundering systems), and OCC Bulletin 2026-13 rescinded the OCC's parallel guidance the same day. The replacement is the joint Revised Guidance on Model Risk Management issued by the Board, the Federal Deposit Insurance Corporation, and the Office of the Comptroller of the Currency. Documents that cite SR 11-7 or SR 21-8 as operative guidance are out of date.
- Does the 2026 revised guidance cover generative AI?
- No — expressly not. Footnote 3 states that generative AI and agentic AI models are novel and rapidly evolving and are not within the scope of the guidance, while directing that an organization's own risk management and governance practices should guide controls for tools the document does not cover. The principles are stated to apply to traditional statistical and quantitative models and to non-generative, non-agentic AI models.
- So is generative AI simply outside model risk management now?
- No. Three things remain true: the guidance itself tells organizations to apply their own risk management and governance practices to uncovered tools; the agencies have announced a request for information on AI model risk, signaling that dedicated expectations are coming rather than absent; and every other obligation a firm carries — recordkeeping, privacy, vendor oversight — applies to generative AI use regardless of the model-risk carve-out. The defensible posture is the control map on this page, adopted as your own governance rather than cited to a rule that does not yet exist.
- We are under $30 billion in assets. Does any of this apply to us?
- The revised guidance says it is expected to be most relevant to banking organizations with over $30 billion in total assets, that smaller organizations are generally excluded consistent with tailored supervision — though it may still be relevant where model-risk exposure is significant — and that it sets no enforceable standards or prescriptive requirements. Two qualifiers matter. OCC Bulletin 2026-13 carries a Note for Community Banks stating the guidance is applicable to all community banks, subject to the limitations discussed in the guidance. And the no-enforceable-standards statement has its own carve-back: supervisory action may still result for violations of law or unsafe or unsound practices stemming from insufficient management of model risk. For a smaller organization, the disciplines are best read as a well-organized menu of sound practice, applied in proportion to what your models actually do.
- What should a model-risk team do about generative AI this quarter?
- Four moves, in order: put every generative deployment in the model inventory with its exact model, version, owner, and purpose; stand up pre-deployment evaluation with independent challenge; turn on per-request logging with a deliberate content mode so monitoring and incident work have an evidence base; and run vendor due diligence on the model chain you cannot inspect directly. Then watch for the announced request for information — the successor framework will land on programs that already have these four in place.
Weighing where a governed workspace versus a direct model API leaves each of these disciplines? The closed-API comparison covers both sides honestly.
Nothing on this page is supervisory guidance or legal advice. It is an implementation map an organization can adopt as its own governance — read the primary sources below before relying on any summary of them, including this one.
Primary sources
This page is anchored on the regulators' own publications — the supersession and the generative-AI carve-out are their words, not commentary. Read them first.
- Federal Reserve SR letter 26-2 (April 17, 2026) — Revised Guidance on Model Risk Management — states that it supersedes and replaces SR letter 11-7, Guidance on Model Risk Management (April 4, 2011), and SR letter 21-8, Interagency Statement on Model Risk Management for Bank Systems Supporting Bank Secrecy Act/Anti-Money Laundering Compliance (April 9, 2021), and notes the guidance is expected to be most relevant to banking organizations with over $30 billion in total assets.
- Supervisory Guidance on Model Risk Management (SR 26-2 attachment) — The joint Board / FDIC / OCC guidance text: model definition, materiality-based tailoring, development and use, validation and effective challenge, outcomes analysis and monitoring, governance and inventory, vendor products — and footnote 3, which places generative and agentic AI outside its scope.
- OCC Bulletin 2026-13 (April 17, 2026) — Model Risk Management: Revised Guidance — rescinds OCC Bulletin 2011-12, the Model Risk Management booklet of the Comptroller's Handbook, Bulletin 1997-24, and Bulletin 2021-19, and announces the planned request for information on banks' use of AI, including generative and agentic AI.