Model risk management

SR 26-2 explained: what it means for AI in banking

The new interagency model risk guidance replaces most of SR 11-7. Here is what changed, who it applies to, and what your model-risk team will ask about AI.

SR 26-2 is the interagency model risk management guidance issued on April 17, 2026 by the Federal Reserve, the OCC (Bulletin 2026-13) and the FDIC (FIL-15-2026). It revises the framework banks have used since SR 11-7 in 2011. It is now the guidance your model-risk team will apply to any AI system that touches a customer decision.

In short

SR 26-2 keeps the core of SR 11-7: sound development, independent validation and governance. It adds a dedicated section on vendor and third-party models. Expectations scale with an institution’s size and complexity. For AI, it raises one practical question: which part of the system is the model, and can you validate it?

What SR 26-2 is

SR 26-2 is supervisory guidance on model risk management, published jointly by the three federal banking agencies. Advisory firms describe it as replacing most of SR 11-7 while keeping its structure: a model inventory, development standards, independent validation, ongoing monitoring and board-level governance.

Sources differ on how completely it supersedes the older letters. Read the agencies’ text alongside your counsel’s view before you update your policies. The summaries below follow the published commentary listed at the end of this article.

What changed from SR 11-7

Vendor and third-party models
A dedicated section covers models you buy rather than build. That includes most AI platforms a bank deploys.
Proportionality
Expectations matter most for larger Federal Reserve–supervised organizations, and scale with size, complexity and risk profile. Community banks and credit unions still face the same questions, at a lighter weight.
Guidance, not rules
SR 26-2 sets supervisory expectations, not prescriptive, enforceable standards. That leaves room in how you meet them, and makes your own documentation the evidence.

Who SR 26-2 applies to

The guidance is aimed most directly at banking organizations with more than $30 billion in assets supervised by the Federal Reserve. Smaller banks and credit unions are not exempt from model risk discipline. Proportionality means the same principles, scaled to the risk of each model.

An AI system that decides fee waivers or Reg E disputes is a customer-facing decision system at any asset size.

What SR 26-2 means for AI systems

This section is our reading of how the guidance applies to AI agent platforms. It is not the agencies’ text.

Define where the model is
An AI system is not one model. Separate the language model from the rules, retrieval and workflow around it. Put each in the inventory with its role.
Narrow what the model decides
The more decisions a language model makes, the broader and harder validation becomes. If rules make the decision and the model only explains, validation can focus on classification and explanation quality.
Get vendor documentation early
Ask vendors for architecture, data flow, change control and testing evidence before the pilot, not after.
Prove outcomes, not intentions
Keep a record of each decision that shows the inputs, the rule that applied and the outcome, so validation and audit can test real behavior.
Control changes
Treat a policy change, a prompt change and a model upgrade as model changes, each with review and sign-off.

A checklist for your next AI review

QuestionWhat good looks like
Is the AI system in the model inventory?Each component listed with an owner and a role
Which component makes the decision?A named, testable component, ideally rules rather than a language model
Can you replay a past decision?Inputs, rule and outcome reproduced step by step
What does the vendor provide?Architecture, data flow, testing and change records
How are changes approved?Versioned changes with review and sign-off
How is performance monitored?Ongoing review of outcomes and escalations

How Palladium applies this

Palladium is built so the language model never makes the decision. The Rulebook decides eligibility, amounts and escalation with versioned rules in code. The model classifies the request and explains the outcome. The Ledger seals every step in a record your validators can replay.

That keeps the model’s role narrow and documented. The due-diligence kit gives your model-risk team the vendor documentation to review, and the technical overview walks through the architecture. Palladium maps its controls to SR 26-2 alongside Reg E, ECOA/Reg B, UDAAP, GLBA and interagency TPRM guidance. Regulatory mapping is not a certification by, or endorsement from, any regulator.

Sources: Federal Reserve SR letters; commentary from Abrigo, Baker Tilly and Sia Partners. This article is for information and is not legal advice.

Walk your model-risk team through it

We will show how rules in code keep the model’s role narrow, with a sealed record your validators can test on one of your own request types.