Book a demo
Topics · REGULATORY

Credit unions Model Risk Controls | Moring

Policy documents do not reduce model risk; controls do. Credit unions need mechanisms that observe AI systems, constrain them, and record what happened: a living inventory, risk classification by use case, pre-use testing, human review, ongoing monitoring, third-party oversight, and evidence generation that holds up in examination.

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

Short answer

Policy documents do not reduce model risk; controls do. Credit unions need mechanisms that observe AI systems, constrain them, and record what happened: a living inventory, risk classification by use case, pre-use testing, human review, ongoing monitoring, third-party oversight, and evidence generation that holds up in examination.

How it works

Model risk management for AI succeeds or fails at the level of the control, not the policy. Industry commentary on AI governance controls makes the point bluntly: every institution has an AI policy, but far fewer can point to the mechanism that makes any sentence of it true. A control observes a system, constrains it, and records what happened. Those three properties are what separate a governable AI portfolio from a documented aspiration.

The foundational control is a living inventory. For a credit union, that means a continuously refreshed record of internal models, vendor-embedded AI, agents, and the datasets behind them, not a spreadsheet updated once a year. Discovery based on self-reporting undercounts, because staff experiments and vendor feature releases introduce AI outside formal channels. Every subsequent control is scoped by this inventory, which is why it comes first.

Risk classification tied to use, not to model, is the second control. The same model is low risk when summarizing internal meeting notes and high risk when influencing a member loan decision. Credit unions that classify by use case can align validation depth, human review thresholds, and monitoring intensity to actual member impact, which is the proportionality that revised model risk guidance, SR 26-2, expects from institutions of different size and complexity.

The execution controls follow from the tier. Pre-use testing verifies conceptual soundness before first use; human review defines which decisions stay with a lending officer or member service supervisor; ongoing monitoring watches for drift against tolerance. Practitioner guidance for credit unions organizes the same work into six steps, each paired with the record that proves the step was done: an inventory entry, a risk rating, a test report, a review log, monitoring results, and vendor oversight documentation.

Evidence generation is the control that makes the others defensible. Examiners ask to see the inventory entry, the materiality rationale, the validation report, and the monitoring history for each model. Credit unions that generate these records automatically, as a byproduct of the controls themselves, avoid the examination scramble of reconstructing evidence after the fact. Controls that leave evidence reduce model risk; controls that merely describe intent leave findings.

Frequently asked questions

Why is a living inventory the foundational control for credit union AI?

Every other control is scoped by the inventory, so an incomplete one silently exempts systems from testing, review, and monitoring. Self-reported inventories undercount because staff experiments and vendor feature releases introduce AI outside formal channels. A continuously refreshed record of internal models, vendor-embedded AI, agents, and datasets is therefore the first requirement of model risk management, and it is the artifact examiners ask to see first.

How does risk classification by use case improve control proportionality?

The same model carries very different risk depending on what it does: summarizing internal notes is low impact, while influencing a member loan decision is high impact. Classifying by use lets a credit union match validation depth, human review thresholds, and monitoring intensity to actual member consequences. This proportionality is what revised model risk guidance, SR 26-2, expects, and it prevents governance effort from concentrating on the wrong systems.

What evidence should each control produce for examination?

Each control should generate its own record as a byproduct: an inventory entry with owners and limitations, a documented risk rating, a pre-use test report with sign-off, a human review log, ongoing monitoring results, and third-party oversight documentation. When records are produced by the controls themselves rather than reconstructed before an exam, the credit union can demonstrate that model risk management operated continuously rather than only on paper.

Which control catches a model that worked at validation but degrades later?

Ongoing monitoring against defined tolerances catches degradation that validation cannot predict. Drift past tolerance triggers escalation, and the model returns for revalidation before continuing to affect member decisions. This closes the loop between testing and operation, which is where model risk actually materializes: a conceptually sound model can still fail in production as member behavior, data quality, or vendor behavior changes over time.

Sources

  1. ncua.gov
  2. crosscheckcompliance.com
  3. cyberriskinstitute.org
  4. csiweb.com
  5. linkedin.com
  6. 360factors.com
  7. go.ai
  8. - federalreserve.gov
  9. - federalreserve.gov
  10. - ithandbook.ffiec.gov
  11. - artificialintelligenceact.eu