Does SR 26-2 Apply to Credit Unions?

Start with the regulatory perimeter

Applicability questions are answered by three facts: who issued the document, who it is addressed to, and what supervisory relationship defines its scope. Only after those are settled does it matter what the document says about AI.

SR 26-2 is Federal Reserve guidance. NCUA rules, by contrast, apply to federally insured credit unions, and the NCUA is the supervisor that examines them. That structure means a credit union should not assume the guidance binds it, and should not assume it can be ignored either. The first task is to read the addressee language in the text and check whether it reaches any institution type outside Federal Reserve supervision.

Interagency practice shows why this matters. The 2001 safeguarding guidelines define scope agency by agency. The Federal Reserve version covers state member banks and their nonbank subsidiaries, and the guidelines reach customer information held by or on behalf of covered entities. Scope follows the supervisory relationship. Where an interagency document names insured credit unions, as the 2024 automated valuation model quality control rule does for federally insured credit unions that rely on such models in residential mortgage lending, the obligation is explicit. Those two documents illustrate how scope works. They are older and say nothing about SR 26-2.

Direct, indirect, and reference-point relevance

Keeping the three tiers separate prevents over-scoping and under-scoping. Document which tier each expectation falls into before assigning budget or owners.

• Direct applicability: The text names the credit union or its supervisory relationship. Counsel confirms. The control set is mandatory and evidence must meet the stated expectations.
• Indirect exposure: A bank partner contract, vendor arrangement, or NCUA examination brings similar expectations to the credit union. Scope is set by the contract or by the examiner's questions.
• Reference point: No obligation attaches. The guidance is used as a benchmark for AI controls and documentation, at the credit union's discretion.

How expectations can reach a credit union indirectly

NCUA examination authority is generally limited to the credit union's own products, services, and operations, not to the vendors that supply them. The NCUA examines those products and services in relation to the credit union offering them. An AI feature delivered by a vendor would likely be examined as part of the credit union's offering, so the credit union needs to be able to show how that feature is controlled even when it does not operate the system.

Interagency frameworks also reach credit unions through examination. Under the interagency consumer compliance rating system, NCUA examiners fold their view of a credit union's compliance risk management into the Management component and the overall CAMEL rating. A weak control environment around AI can therefore surface in a rating without any Federal Reserve document applying directly.

The interagency proposal on third-party risk management, published in the Federal Register on September 15, 2026, defines banking organizations to include insured credit unions. As of November 2026 it is proposed guidance with no effective date, and its final form is unknown. It should not be treated as final, and it should not be treated as equivalent to SR 26-2.

Two further exposure routes depend on facts specific to each institution. A bank partner or sponsor may contractually require a credit union or its credit union service organization to follow particular expectations. Contract language, not the guidance itself, then defines the obligation. Separately, the NCUA stated in 2026 that it will not consider reputation risk in supervisory determinations. Justify AI controls on operational and compliance grounds, not as protection against reputation-based supervisory action.

What runtime enforcement covers and what it does not

Runtime policy enforcement addresses individual interactions and tool calls. It can show what an AI system was permitted to do under a defined policy and what it did in each case. That evidence supports an examination conversation under whatever framework applies, because it answers a basic question directly: were the stated controls operating at the time of the activity?

It does not replace the parts of a model risk program that sit outside the execution path. Model development documentation, independent validation, pre-deployment testing, and model lineage are separate disciplines with separate owners. A credit union that treats runtime enforcement as a complete answer to any model risk expectation will leave gaps.

Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, and generates audit-ready evidence from that enforcement. The evidence is mapped to frameworks including HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. Trussed has run in continuous production in a regulated enterprise for more than 18 months, with 99.95% uptime on live production traffic, and deploys self-managed in the customer's cloud with configurable retention and data residency. Those capabilities do not constitute a mapping to SR 26-2 or to NCUA expectations, and any such mapping would need to be established separately.