Book a demo
Topics · REGULATORY

Credit Union Model Risk Compliance | Moring

Compliance for credit union AI model risk runs through existing frameworks rather than a dedicated AI rulebook: supervisory expectations apply technology-neutrally, the revised model risk guidance covers traditional models, state statutes add obligations where members live, and every question is answered with records rather than attestations.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Compliance for credit union AI model risk runs through existing frameworks rather than a dedicated AI rulebook: supervisory expectations apply technology-neutrally, the revised model risk guidance covers traditional models, state statutes add obligations where members live, and every question is answered with records rather than attestations.

How it works

There is no separate AI examination, and pretending otherwise wastes preparation. The supervisory posture treats AI through the frameworks institutions already know: third-party oversight, information security, consumer protection, and model risk where models are traditional. AI in origination is a fair-lending and screening question first, meaning that member identification, sanctions screening, and adverse-action obligations work exactly as they would without AI, and that decisions remain explainable. An AI vendor is a third-party vendor, covered with the same rigor as the core processor plus AI-specific items such as model governance and data-use commitments. The credit union that prepares for an AI examination is really preparing its existing programs to cover new systems.

The model risk layer has a specific current shape. The revised interagency guidance issued April 17, 2026 replaced the 2011 framework and covers traditional and non-generative models with an evidence set of inventory entries, materiality rationales, validation reports, and monitoring results, while stating plainly that generative and agentic AI are outside its scope. Compliance for those systems therefore rests on the institution's own AI policy and existing risk frameworks, which raises the bar on policy quality: the policy the credit union wrote is the rule its examiners will measure its generative systems against. Institutions that deferred policy work lost their only applicable standard.

State law adds a layer the federal posture does not cover. Several states have passed AI legislation, with California and Colorado leading, and a credit union serving members in those states will likely need to comply even if federally chartered, so monitoring legislative developments and evaluating their impact on the AI governance program becomes an ongoing compliance activity. The practical translation is a use-case map: which AI uses touch members in which states, and which state statutes attach to each. For most community institutions the list is short today, but it is growing, and the mapping exercise is far cheaper annually than reactively.

Examination itself is evidence-driven. Examiners ask where AI is used across workflows, whether each AI-touched decision can be reconstructed, whether systems can be shut down, what data they can access beyond authorized limits, how vendors are governed, and where humans sit in the loop, and those questions are answered with records rather than assertions. The supervisory resource hub the credit union agency maintains frames its AI materials as tools for due diligence and risk management rather than new regulatory requirements, which is a consistent signal: the regulator is mapping AI risk onto existing frameworks, and institutions should map their evidence the same way.

The compliance program therefore reduces to a documentation discipline over familiar controls. Maintain the AI inventory with tiering; keep validation or testing records current per tier; log human review of consequential decisions; monitor performance and drift between reviews; carry current vendor due diligence with contract controls; and rehearse shutdown of agentic systems. Supervisory guidance confirms AI remains subject to existing, technology-neutral regulations, and the institutions passing examination cleanly are those treating AI compliance as a present-tense obligation rather than a future project. The gap between the two postures is visible in every examination cycle, and it is closed with records, not with new frameworks.

Frequently asked questions

What compliance frameworks apply to AI at a credit union?

The existing ones, applied technology-neutrally: third-party oversight, information security, consumer protection including fair lending and adverse-action obligations, and model risk guidance for traditional models. The revised interagency guidance covers non-generative models and explicitly places generative and agentic AI under the institution's own AI policy and existing risk frameworks. State AI statutes, such as those in California and Colorado, add obligations where members live, potentially even for federally chartered institutions. No separate AI examination exists; preparation means extending the programs already in place.

How should a credit union prepare for AI questions in an examination?

By assembling model risk records that answer the known question set: where AI is used across workflows, how each AI-touched decision can be reconstructed, whether systems can be shut down, what data they can access, how vendors are governed, and where humans sit in the loop. Each answer should come from an artifact, an inventory entry, a testing record, a review log, a monitoring result, a vendor file, rather than an assertion. Institutions that treat AI compliance as a present-tense obligation produce these from operation; others reconstruct them under deadline.

Do state AI laws affect federally chartered credit unions?

Likely yes, where members live in states that passed AI legislation, with California and Colorado the early leaders. State applicability generally follows the member relationship rather than the charter, so a credit union serving members in those states should evaluate the statutes' impact on its governance program and monitor legislative developments continuously. The practical control is a use-case map connecting AI uses to the states they touch, reviewed annually; for most community institutions the mapped obligations are few today, but the list grows and the mapping is cheaper than retrofitting.

What does compliance mean for generative and agentic AI specifically?

Compliance through the institution's own AI policy, because the revised model risk guidance excludes those systems from its scope. That makes the policy the operative standard: it should define approved uses, testing requirements, human review points, monitoring tolerances, and shutdown capability for agentic systems, and its records, approved-use records, testing records, human review logs, become the evidence of compliance. A credit union with a thin policy in this area has no applicable standard for its fastest-growing systems, which is the single most exposed position in the current posture.

Sources

  1. ncua.gov
  2. crosscheckcompliance.com
  3. cyberriskinstitute.org
  4. americascreditunions.org
  5. torchlight.io
  6. linkedin.com
  7. csiweb.com
  8. ituring.ai
  9. - federalreserve.gov
  10. - federalreserve.gov
  11. - ithandbook.ffiec.gov
  12. - artificialintelligenceact.eu