Model Risk Directory emblemModel Risk Directory
AI modelsmachine learninggenerative AIagentic AI

AI model risk under SR 26-2: what is in scope and what is not

SR 26-2 applies its model risk principles to traditional statistical, quantitative, and non-generative, non-agentic AI models. Generative and agentic AI are outside this guidance's model scope, but they are not outside bank risk management.

Published: Last updated: Last reviewed by: Model Risk Directory editorial team

Start with the definition, not the label

SR 26-2 defines a model as a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to input data to produce quantitative estimates. A tool does not move into or out of scope merely because a vendor calls it AI. The institution must examine the method, output, intended use, materiality, exposure, and business risk.

The guidance excludes simple arithmetic and deterministic rule-based processes that do not rely on statistical, economic, or financial theory. That boundary is useful, but it is not a free pass for weak tools. A system outside the guidance's model definition can still create operational, compliance, consumer, third-party, information-security, or safety-and-soundness risk.

Traditional and non-generative AI can be in scope

The Federal Reserve's footnote is unusually direct: the principles apply to traditional statistical and quantitative models and to non-generative, non-agentic AI models. A machine-learning credit model, fraud model, forecasting model, or classification model can therefore be governed under SR 26-2 when it meets the definition and supports a banking use.

The risk-based approach matters. Validation depth, monitoring, documentation, and governance should reflect inherent risk, model purpose, exposure, materiality, changes, data constraints, and the complexity of the institution. 'AI' is not a substitute risk tier. A simple but high-impact model may warrant more control than a complex experiment with no material decision role.

  • Document intended purpose, users, affected decisions, exposure, and prohibited uses.
  • Assess data quality, representativeness, assumptions, methodology, performance, and limitations.
  • Validate before first use when practical, with compensating controls for justified exceptions.
  • Monitor performance, data relevance, limitations, thresholds, changes, and outcomes.

Generative and agentic AI are outside this guidance's scope

SR 26-2 states that generative and agentic AI models are novel and rapidly evolving and are not within the scope of this guidance. That sentence should be read precisely. It does not say banks may deploy those systems without governance, inventory, testing, monitoring, or accountability.

The same footnote says a banking organization's risk management and governance practices should determine appropriate governance and controls for tools, processes, or systems not covered by the document. The practical result is a two-step assessment: decide whether the system is an SR 26-2 model, then identify the other governance obligations and risks that apply whether or not the answer is yes.

Vendor AI remains the bank's risk problem

SR 26-2 gives vendor and third-party products their own section. Proprietary code, restricted training information, closed methodologies, and vendor-supplied parameters can make validation harder, but the model risk principles still apply to in-scope vendor models.

A procurement team should treat missing information as a control constraint to manage, not as evidence that a product is reliable. Due diligence should test conceptual understanding, intended use, data, performance evidence, limitations, change notices, monitoring, audit rights, continuity, incident handling, and exit. When a vendor model is customized, the institution should document and evaluate the adjustment.

  • Require enough evidence to understand design, development data, performance, and limitations.
  • Define validation access, monitoring data, change notification, issue support, and exit rights in the contract.
  • Maintain institution-specific outcomes analysis rather than relying only on vendor-wide benchmarks.

A workable control decision

An inventory record should capture the system's method, purpose, owners, users, provider, decisions, model-definition conclusion, materiality, validation status, monitoring plan, limitations, dependencies, and applicable control regimes. The goal is a traceable decision, not a semantic argument about whether the product is 'really AI.'

Internal audit should evaluate whether the governance process is rigorous and effective rather than duplicating model development or validation. Model risk, compliance, legal, technology risk, information security, third-party risk, and business owners may all need defined roles. The mix depends on the system and use case.

What to update now

Institutions still using an SR 11-7-era policy should update the model definition, materiality logic, validation-frequency approach, third-party-model controls, and references to the current guidance. They should also add an explicit decision path for generative and agentic AI so the SR 26-2 scope exclusion does not become a governance gap.

The useful deliverable is a control map that shows which systems fall under SR 26-2, which do not, which other policies apply, who approved the determination, and what evidence will be reviewed over time. That map is more defensible than forcing every AI system into one policy or excluding every new AI system from model governance.

Primary sources

  1. Federal Reserve SR 26-2 supervisory letter
  2. Federal Reserve revised interagency guidance
  3. OCC Bulletin 2026-13

AI model risk under SR 26-2: what is in scope and what is not: FAQ

Does SR 26-2 apply to machine-learning models?

It can. The guidance says its principles apply to traditional statistical and quantitative models and non-generative, non-agentic AI models when they meet the guidance's model definition.

Does SR 26-2 apply to generative AI?

The guidance says generative and agentic AI models are outside its scope. It also says the banking organization's broader risk management and governance practices should determine appropriate controls for tools and systems not covered.

Are vendor AI models exempt from validation?

No. For in-scope vendor models, SR 26-2 says the model risk principles remain applicable even when proprietary components limit access. The institution should understand, validate, monitor, and govern the product using available evidence and appropriate controls.

Related content

GuideModel Risk ManagementModel risk management (MRM) is the discipline of identifying, measuring, and controlling the risk that a bank's quantitative models produce incorrect or misused results. An MRM advisory firm helps build or run that discipline: model inventory, tiering, governance, and independent validation, rooted in Federal Reserve SR 26-2 (formerly SR 11-7) and OCC guidance.GuideModel Risk ValidationModel risk validation is the independent testing of a specific model's conceptual soundness, ongoing performance, and outcomes, performed by staff or a third-party firm genuinely separate from that model's development. It is the 'effective challenge' function required under SR 26-2 and SR 11-7, distinct from broader model risk management program design.GuideModel Risk Management FrameworkA model risk management framework is the documented set of policies, roles, and processes an institution uses to identify, tier, govern, and validate its models. Under SR 26-2, a sound framework covers model development and use, independent validation, and governance and controls, scaled to the institution's size and model risk profile rather than applied uniformly.FrameworkSR 26-2SR 26-2, issued jointly by the Federal Reserve, OCC, and FDIC on April 17, 2026, is the current US interagency guidance on model risk management. It supersedes SR 11-7 (2011) and SR 21-8 (2021), preserving core disciplines like effective challenge and independent validation while shifting to a risk-based approach tailored to an institution's model risk profile, size, and complexity.FrameworkMAS AI MRMThe MAS AI MRM paper reports good practices observed in a thematic review of selected banks. It covers AI governance, inventories, materiality assessment, data, development, independent validation, deployment, monitoring, change, generative-AI risks, and third-party models. It is supervisory guidance and observed practice, not a general statutory MRM rule.FrameworkOSFI E-23OSFI E-23 is Canada's enterprise-wide model risk management guideline for federally regulated financial institutions. The final 2025 guideline takes effect on 1 May 2027 and applies a risk-based approach across model identification, inventory, risk ratings, lifecycle governance, independent review, monitoring, third-party models, and decommissioning.