Does SR 26-2 Apply to Credit Unions? How It Compares with NCUA AI Expectations

Start with jurisdiction, not the document number

SR letters are a Federal Reserve instrument, and the Federal Reserve Board is a different agency from the NCUA. NCUA describes itself as responsible for safety and soundness and consumer protection at federal credit unions. That distinction is the starting point for any applicability analysis. A credit union that uses AI for automated loan underwriting, member-service virtual assistants or fraud detection, all of which NCUA's Acting Director cited in March 2026 testimony, answers to NCUA examiners for that use.

The applicability question therefore reduces to three cases. A credit union with no direct Federal Reserve relationship and no Federal Reserve-supervised counterparty in its AI supply chain faces NCUA expectations. A credit union that depends on a Federal Reserve-supervised partner bank or vendor may meet that counterparty's expectations through contracts, due diligence requests or shared processes. A credit union with its own direct Federal Reserve relationship is a different case that its counsel should assess. Use the text of SR 26-2 and your legal advisers to place your institution in one of these cases. This page is not legal advice.

The practical consequence is to record, for each control, which standard it is built to. Teams that treat every bank-regulator document as binding tend to over-build for the wrong examiner. Teams that ignore them can miss obligations that arrive by contract.

NCUA expectations and Federal Reserve-side interagency materials by control domain

The table compares NCUA's stated supervisory approach with the Board, FDIC and OCC interagency materials that bear on AI. It reflects what those materials say as of November 2026 and does not summarize SR 26-2 itself.

• Model risk management: NCUA is not a party to the Model Risk Management Guidance issued by the OCC, Board and FDIC. It monitors credit unions' model risk management through its supervisory approach, and in its automated valuation model rule it looks for governance and controls appropriate to size, complexity, transaction risk and capabilities.
• Governance and policy: NCUA states that supervisory guidance does not have the force and effect of law and that it does not take enforcement actions based on supervisory guidance. The 2023 third-party guidance from the Board, FDIC and OCC says the same and adds that it imposes no new requirements.
• Third-party oversight: The 2023 interagency guidance names the Board, FDIC and OCC. The proposed guidance published September 15, 2026 adds NCUA as an issuing agency, and it is not final. Both treat third-party use as leaving institutional responsibility in place.
• Member and customer data: For credit unions, Part 748 Appendix A applies to member information maintained by or on behalf of federally insured credit unions. It aims at security, confidentiality and protection against unauthorized access or use that could cause substantial harm or inconvenience to a member.
• Incident reporting: Part 748 requires a federally insured credit union to report a reportable cyber incident no later than 72 hours after it reasonably believes one has occurred.
• Voluntary frameworks: NIST AI RMF Govern 1.1 calls for legal and regulatory requirements involving AI to be understood, managed and documented. It is a voluntary reference, and neither agency is described as requiring it.

Where overlap and indirect exposure arise

The strongest overlap is in third-party oversight. The proposed interagency guidance says an institution's use of a third party does not diminish its responsibility to meet requirements as if it performed the activity itself. That holds for affiliates, highly regulated service providers and subcontractors, and where a third party performs elements of risk management. For a credit union using an AI vendor, a vendor's own model providers and tool integrations extend that responsibility down the chain.

The second overlap is the expectation that governance matches the institution. NCUA's described approach scales controls to size, complexity and risk. A small credit union is not expected to replicate a large bank's program, but it is expected to know what its AI does and to show that its policies are carried out.

Indirect exposure is a contract and relationship question rather than a regulatory one. A partner bank may pass its own supervisory expectations to the credit union through agreements, questionnaires or service terms. Whether any such flow-down exists for a given standard is a fact about your contracts, so review vendor and partner-bank agreements and note the standard each requirement comes from.

Because supervisory guidance is not enforceable law, present internal control statements as supervisory expectations. Reserve the word requirement for obligations such as Part 748. This keeps examination responses accurate and keeps remediation priorities in order.

What runtime enforcement can and cannot answer

Several expectations in both frameworks come down to one demonstration: what the AI was permitted to do, and what it did. Policy documents and periodic reviews describe intent. Examiners also ask whether the policy was implemented. Enforcing policy in the execution path produces a record of that implementation for each interaction, rather than reconstructing it after the fact.

A runtime control sits between the AI system and the things it touches: prompts, responses, tools and data. For each interaction or tool call it applies a policy and takes one of a small set of actions. It allows the call, redacts sensitive fields such as member information, blocks the call, or escalates it to a named person for human approval. Each decision is recorded with the policy version applied, a timestamp, the acting user or system, and any human approver.

This addresses member-information protection under Part 748 and gives a concrete answer to the question of which data an AI system could reach. Trussed enforces policy at this layer across models, agents, tools and MCP servers, and the enforcement itself generates audit-ready evidence. The scope is individual interactions and tool calls.

Runtime control does not replace model validation, vendor due diligence, board-level governance or the incident response process. It also does not make a framework's requirements satisfied by itself, and neither NCUA nor the Federal Reserve is described as mandating specific tooling. Extend the same controls and logging to third-party AI services where you can, because responsibility stays with the credit union.