Credit Union Model Risk Checklist | Moring
A model risk checklist for credit union AI has six working items, each paired with the record that proves it was done: inventory, risk rating, pre-use testing, human review, ongoing monitoring, and third-party oversight. The checklist survives the recent guidance changes, because the obligations moved between instruments without shrinking.
Short answer
A model risk checklist for credit union AI has six working items, each paired with the record that proves it was done: inventory, risk rating, pre-use testing, human review, ongoing monitoring, and third-party oversight. The checklist survives the recent guidance changes, because the obligations moved between instruments without shrinking.
How it works
Begin with the mapping, because the checklist changed shape in April 2026. The revised interagency model risk guidance issued on April 17 replaced the 2011 framework, rescinding the prior supervisory letter, and placed generative and agentic AI outside its scope, leaving those systems to the institution's own AI policy and existing risk frameworks. A checklist written today therefore maps each AI use to its governing instrument: traditional and non-generative models to the revised guidance with its familiar evidence set, and generative or agentic systems to the institution's policy records. Neither event removed a single control the credit union owes; they scattered the map.
Item one is the inventory, and it earns first position because everything else tests against it. Maintain a register of internally built models, vendor AI tools, generative applications, AI agents, and AI embedded in SaaS systems, each with a named owner, a tier, and a lifecycle state. Sample the register against actual usage, vendor notices, and spending, because AI arrives through platform updates and staff tools faster than registers cycle. The record an examiner inspects is the inventory entry with its materiality rationale.
Item two is risk rating, tiering each use case by member impact, data sensitivity, autonomy, and regulatory exposure. The tier is not decoration: it determines which of the later items apply at full depth. A lending decision-support workflow, touching fair lending obligations and member outcomes, warrants the deepest testing and monitoring; an internal summarization assistant warrants proportionally less. The record is the classification itself, with its rationale, dated and owned.
Item three is pre-use testing before any member-facing deployment: performance baselines, bias and error analysis appropriate to the use case, and documented acceptance against defined criteria. For generative systems, testing includes prompt and behavior evaluation rather than only statistical back-testing. The record is the testing document that predates the go-live, which is the artifact examiners use to distinguish a governed deployment from an enthusiastic one.
Item four is human review, defined precisely enough to audit: which decisions a person reviews before they take effect, what the reviewer actually checks, and where the review log lives. The distinction between meaningful review and confirmation-by-click is the difference between a control and a ritual, and governance commentary is increasingly blunt that human-in-the-loop labels can conceal hollow supervision. The record is the review log showing what the reviewer saw and decided.
Item five is ongoing monitoring with tolerances and escalation: performance tracking, drift detection, and a defined path from a breached tolerance to the board. Item six is third-party oversight: due diligence, contract controls covering data use and log delivery, ongoing monitoring of vendor behavior, and exit provisions for credentials and member data. The checklist closes with the meta-item that makes the rest durable: schedule the whole set on a cycle, review it after any significant vendor change, and keep the records where internal audit can find them without a request.
Frequently asked questions
What are the core items on a credit union model risk checklist?
Six, each paired with its evidence: an inventory of models, agents, and embedded AI with owners and materiality rationale; risk tiering by member impact and data sensitivity; pre-use testing with documented acceptance before deployment; defined human review of consequential decisions with logs; ongoing monitoring with tolerances and escalation to the board; and third-party oversight covering due diligence, contracts, and exit provisions. The pairing matters, because a step without its record is indistinguishable at examination from a step that never happened.
Did the April 2026 guidance change eliminate any checklist items?
No. The revised interagency guidance replaced the 2011 framework and placed generative and agentic AI outside its scope, but the obligations migrated into the institution's own AI policy and existing risk frameworks rather than disappearing. A credit union should update the mapping on its checklist so traditional models cite the revised guidance with its evidence set, and generative or agentic systems cite the institution's policy records. The controls, inventory, testing, review, monitoring, and vendor oversight, remain owed either way.
How should the checklist treat AI agents specifically?
As a distinct inventory class with their own records: per-agent identity, approved scope, human approval points for consequential actions, monitoring of tool-call behavior, and a tested way to shut them off. Agents act rather than recommend, so the human review item concentrates on the actions they may take unaided, and the monitoring item watches execution rather than only output quality. A checklist that treats an agent like a scoring model misses the execution risk, which is where the damage from a misbehaving agent actually lands.
How often should the checklist be run?
At a credit union, the inventory and tiering should update continuously as AI arrives through vendors and staff, while the full pass, testing currency, review logs, monitoring results, and vendor files, runs at least quarterly, with high-impact uses examined more often. Re-run the relevant items after any significant vendor update or scope change, since platform changes can alter model behavior without any internal action. Scheduling the set on a cycle and keeping records current is what converts a checklist from a document into an operating control.