Key Takeaways
- SR 26-2 superseded SR 11-7 and SR 21-8 on April 17, 2026, so an existing SR 11-7 program needs re-papering against the new text rather than retirement.
- Footnote 3 of the guidance places generative and agentic AI outside its scope, which moves the control decision for LLM applications and agents to the bank’s own governance program.
- Model purpose and model exposure set materiality, and the guidance mandates no tier scheme, so the bank has to document its own.
- A vendor model, including a vendor LLM, stays inside the bank’s own validation and monitoring responsibility even when the vendor withholds code, data or methodology, so the bank cannot rest on the vendor’s assurance alone and should validate and monitor the model itself or through an outside party.
- Generative AI and agents need controls from the bank’s own AI governance program, starting with a registry and continuous monitoring.
Consider a bank whose customer-service agent clears its pre-launch review in March. The validator confirms the retrieval corpus, checks a sample of fee-dispute answers against policy, and signs off. In June the foundation-model provider ships an update, the agent begins quoting a waiver rule the bank retired a year earlier, and the bank does not recheck because the review was a one-time event.
The scenario illustrates a documented behavior-change risk. In the published study, Chen et al. measured GPT-4’s accuracy on a prime-number task falling from 84.0% to 51.1% between its March and June 2023 versions [1].
SR 26-2, Revised Guidance on Model Risk Management (MRM), is supervisory guidance issued by the Federal Reserve, the Office of the Comptroller of the Currency (OCC) and the Federal Deposit Insurance Corporation (FDIC) on April 17, 2026. It supersedes and replaces SR 11-7, the 2011 MRM guidance the OCC issued as Bulletin 2011-12 [2], and SR 21-8, the 2021 interagency statement on MRM for Bank Secrecy Act and anti-money laundering (BSA/AML) systems [3].
What SR 26-2 Changes for Model Risk Management
SR 26-2 keeps the lifecycle structure of SR 11-7, strips out most of the prescriptive detail, and ties the rigor of each practice to model materiality. The Federal Reserve letter describes the guidance as emphasizing “a risk-based approach to model risk management that is tailored to a banking organization's model risk profile and the size and complexity of its operations” [3]. The attached guidance adds a statement SR 11-7 never carried. It says the guidance “does not set forth enforceable standards or prescriptive requirements,” and non-compliance with it will not draw supervisory criticism, though a footnote preserves supervisory action for “violations of law or unsafe or unsound practices stemming from insufficient management of model risk” [4].
The OCC issued the same text as Bulletin 2026-13 and rescinded Bulletin 2011-12 and the Model Risk Management booklet of the Comptroller’s Handbook [5]. The FDIC’s letter states the guidance generally does not apply to institutions at or below $30 billion in total assets unless they have significant model risk exposure [6].
The letter is “expected to be most relevant to banking organizations with over $30 billion in total assets regulated by the Federal Reserve” [3]. The guidance says models at smaller institutions typically fall under internal practices suited to their size, but it “also may be relevant” to organizations whose models are prevalent and complex or whose activities fall outside traditional community banking [4].
The following table compares the two documents on the dimensions a program owner has to re-paper:
SR 11-7 organized model risk around development, implementation and use; validation built on conceptual soundness, ongoing monitoring and outcomes analysis; and governance with a firm-wide model inventory and documentation [2]. SR 26-2 keeps each component as a section heading and removes the techniques, cadences and role lists SR 11-7 attached to them [4]. Existing inventories, documentation standards, validation reports and monitoring reports remain the right artifacts. What changes is the stated rationale for how much of each a given model gets.
How SR 26-2 Narrows the Model Definition to Exclude LLMs and Agents
SR 26-2 narrows the model definition and then carves generative and agentic AI out of it entirely. Section II defines the term: “the term 'model' refers to a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates.” The same section excludes “simple arithmetic calculations, such as those found within spreadsheets, as well as deterministic rule based processes and software where there are no statistical, economic, or financial theories underpinning their design or use” [4].
Footnote 3 attaches to that definition and settles the LLM question in four sentences: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance. Nonetheless, a banking organization's risk management and governance practices should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document. However, the principles described in this guidance apply to traditional statistical and quantitative models and non-generative, non-agentic AI models” [4].
The footnote sorts a bank’s AI systems into two groups:
- In scope: a gradient-boosted credit decision model, which as non-generative AI is still expected to undergo effective challenge.
- Out of scope by default: a foundation-model customer-service agent or a document-summarization LLM application.
Federal Reserve Vice Chair for Supervision Bowman described the result on May 1, 2026: the guidance “now applies narrowly to traditional models and basic AI applications,” and “Today, banks are relying on existing risk-management frameworks to guide their use of AI” [7]. OCC Bulletin 2026-13 says the agencies “plan to issue in the near future a request for information” on MRM and banks’ use of AI, including generative and agentic AI [5]. Until that request is issued, the footnote’s instruction to let the bank’s own practices “guide the determination of appropriate governance and controls” is the operative text.
Many banks will assign the residual risk for out-of-scope systems to the AI governance program. Either way, the MRM policy and the AI governance policy need a written boundary stating which systems each covers and who signs off when a system moves between them.
Core Expectations Under SR 26-2
SR 26-2 organizes its expectations into lifecycle sections IV through VII and scales each by materiality rather than by a fixed checklist [4].
Development, Implementation and Use
Under Section IV, a banking organization builds and documents a model on sound theory, then uses it for the purpose it was designed for. A bank can limit testing based on risk: “For models that have relatively less material impact or for banking organizations with relatively less complex operations, model testing may be more limited in scope” [4]. Documentation supports continuity of operations, “the tracking of recommendations, responses, and exceptions,” and “any model remediation efforts.”
Validation and Effective Challenge
Under Section V, validation covers conceptual soundness, meaning model design, key assumptions, data selection, construction and “developmental testing.” It also covers ongoing monitoring of whether a model is “performing as expected” as conditions change, along with outcomes analysis that “compares model outputs to corresponding real-world outcomes,” such as back-testing or outlier analysis [4]. For effective challenge, SR 26-2 replaces SR 11-7’s structural independence language with a quality test, stating that “The quality of validation process depends on the rigor and effectiveness of the review rather than on organizational structure.” SR 11-7’s “at least annually” cadence is gone, and validation “generally occurs prior to a model's first use.” Timing then varies with model purpose, methodology, and the frequency and scope of model changes.
Governance, Inventory and Materiality
Under Section III, bank staff assess materiality using the guidance’s governing relationship: “Model purpose, together with model exposure, determines model materiality” [4]. Purpose concerns what the model is for. Models that support regulatory requirements or manage financial risk exposures generally carry more risk. Exposure concerns how much the output drives business decisions, which a bank can measure by portfolio size. Higher-materiality models warrant more rigorous oversight. The guidance prescribes no tier names, so any existing tiering scheme the bank uses needs a written purpose-and-exposure rationale. Under Section VI, banks should keep “sufficient information to understand model risks” in the inventory for models under development or in use [4].
Third-Party and Vendor Models
Section VII keeps vendor models inside the program: “the principles of model risk management remain applicable” even when a bank does not receive the vendor’s underlying code, data or methodology [4]. Sound practice includes understanding the vendor model’s conceptual soundness, design, development data and performance, and validating it through internal or outside parties. Banks are also expected to run ongoing monitoring and outcome analysis to confirm vendor models stay accurate and fit for purpose, and to document, justify and validate any customizations [4]. A foundation-model API sits outside SR 26-2’s scope, yet these four expectations are a defensible minimum for any vendor LLM the bank places in front of customers.
Governing Generative and Agentic AI Outside SR 26-2
Because footnote 3 hands the control decision to the bank, the bank’s own program has to supply what MRM would otherwise have supplied, and it has to supply it for systems whose behavior changes between reviews. For the customer-service agent in this scenario, those controls are:
- Inventory: The bank should give every LLM application and agent a registry entry with an owner, a lifecycle status and the model version, applying its own governance to systems outside SR 26-2 [4]; a retired agent’s tool credentials should be revoked.
- Ongoing monitoring: Drift means a provider update changed answers to the same prompts, the pattern Chen et al. measured [1]. Hallucination means the agent stated a fee or policy the retrieved context does not support. The monitoring record captures which tool the agent called, with what arguments, and whether that matched policy. The bank should record prompt provenance, including the system-prompt version, retrieved documents and any tool-returned content. The bank should tie each signal to a threshold that triggers an alert or an enforcement rule.
- Inline policy enforcement: The bank can check the request before invoking the model to detect personally identifiable or protected health information (PII/PHI) or a prompt-injection attempt. It can check the response before it reaches the customer to score groundedness against the retrieved context. The bank should start with deterministic rules and apply model-based scoring where a rule cannot assess the behavior.
- Audit trail: The bank should retain contextual metadata for every enforcement decision, including who triggered it and what policy applied, so it can reconstruct what the agent did on a given date for an examiner or a customer complaint.
Fair lending and privacy obligations continue to apply to these systems. BSA/AML obligations also continue to apply. The Consumer Financial Protection Bureau’s 2023 report on chatbots in consumer finance states that chatbots “must comply with all applicable federal consumer financial laws” and that institutions have “an obligation to keep personally identifiable information safe, regardless of the technology used” [8]. The NIST AI Risk Management Framework 1.0, which NIST says is being revised, gives banks a reference structure for mapping these four controls [9].
The Fiddler AI Observability and Security Platform produces a registry and monitoring record for a bank’s first-party, third-party and coding agents and LLM applications.
- Reliable Evaluation: Scores prompts and responses against built-in and custom evaluators before deployment and in production.
- Continuous Monitoring: Tracks hallucination, drift and tool-use behavior across the agentic hierarchy of application, session, agent, trace and span.
- Enforceable Policy: Enforces policy inline on the agent request and response path, with allow, block or redact actions depending on the policy and use case.
- Auditable Governance: Provides an AI registry with a single view of live, testing and retired models and applications, with ownership and lifecycle status. It also records each enforcement decision with who triggered it and what policy applied.
The platform supports compliance evidence for the bank’s own AI governance program over the systems SR 26-2 leaves out. Alignment and evidence generation are not a guarantee of legal compliance.
Request a demo to review how a registry and continuous monitoring for LLM applications and agents would fit the bank’s existing model risk program and deployment environment.
References
[1] L. Chen, M. Zaharia, and J. Zou, “How Is ChatGPT’s Behavior Changing over Time?” Harvard Data Science Review, 2024. Accessed: Oct. 13, 2026. [Online]. Available: https://hdsr.mitpress.mit.edu/pub/y95zitmz
[2] Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation, SR 11-7: Guidance on Model Risk Management. Apr. 4, 2011. Accessed: Oct. 13, 2026. [Online]. Available: https://www.federalreserve.gov/boarddocs/srletters/2011/sr1107a1.pdf
[3] Board of Governors of the Federal Reserve System, SR 26-2: Revised Guidance on Model Risk Management. Apr. 17, 2026. Accessed: Oct. 13, 2026. [Online]. Available: https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm
[4] Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation, Supervisory Guidance on Model Risk Management. Apr. 17, 2026. Accessed: Oct. 13, 2026. [Online]. Available: https://www.federalreserve.gov/supervisionreg/srletters/SR2602a1.pdf
[5] Office of the Comptroller of the Currency, OCC Bulletin 2026-13: Model Risk Management: Revised Guidance. Apr. 17, 2026. Accessed: Oct. 13, 2026. [Online]. Available: https://www.occ.gov/news-issuances/bulletins/2026/bulletin-2026-13.html
[6] Federal Deposit Insurance Corporation, FIL-15-2026: Agencies Revise the Interagency Model Risk Management Guidance. Apr. 17, 2026. Accessed: Oct. 13, 2026. [Online]. Available: https://www.fdic.gov/news/financial-institution-letters/2026/agencies-revise-interagency-model-risk-management-guidance
[7] M. W. Bowman, “Artificial Intelligence in the Financial System,” speech, Board of Governors of the Federal Reserve System, May 1, 2026. Accessed: Oct. 13, 2026. [Online]. Available: https://www.federalreserve.gov/newsevents/speech/bowman20260501a.htm
[8] Consumer Financial Protection Bureau, Chatbots in Consumer Finance. Jun. 6, 2023. Accessed: Oct. 13, 2026. [Online]. Available: https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/
[9] National Institute of Standards and Technology, AI Risk Management Framework (AI RMF 1.0). Jan. 26, 2023. Accessed: Oct. 13, 2026. [Online]. Available: https://www.nist.gov/itl/ai-risk-management-framework
[10] Board of Governors of the Federal Reserve System, Supervision and Regulation Letters. Accessed: Oct. 13, 2026. [Online]. Available: https://www.federalreserve.gov/supervisionreg/srletters/srletters.htm
Frequently Asked Questions
Is SR 11-7 still in effect?
The Federal Reserve letter states that SR 26-2 “supersedes and replaces SR letter 11-7, Guidance on Model Risk Management (issued April 4, 2011) and SR letter 21-8” [3]. The OCC rescinded Bulletin 2011-12 [5].
Does SR 26-2 apply to smaller banks?
The FDIC states that the guidance generally does not apply to institutions with $30 billion or less in total assets unless they have significant exposure to model risk [6]. The OCC bulletin says the guidance applies to all community banks subject to its limitations, and it cites OCC Bulletin 2025-26 for the point that community banks are not required to perform annual model validation [5].
Are LLMs and AI agents models under SR 26-2?
Footnote 3 of the attached guidance states that generative AI and agentic AI models “are not within the scope of this guidance,” and that the bank’s own risk management and governance practices “should guide the determination of appropriate governance and controls.” Non-generative, non-agentic AI models remain in scope [4].
What does the 26-2 in the letter number mean?
The Federal Reserve numbers SR letters sequentially by year, and its index gives the example that “the first letter issued in 2005 is numbered SR 05-1” [10]. On that convention, SR 26-2 is the second supervisory letter of 2026 and SR 11-7 was the seventh of 2011.
What are the principles of the OCC's model risk management guidance?
OCC Bulletin 2026-13 adopts the same interagency text as SR 26-2 to clarify model risk management principles and set out a risk-based approach [5]. The principles cover materiality based on model purpose and exposure, conceptual soundness and documentation, validation with effective challenge and outcomes analysis, ongoing monitoring, a model inventory with clear accountability, and the same expectations for vendor models [4].