Credit Union Model Risk Audit | Moring
Auditing AI model risk at a credit union means internal audit examines whether models and AI workflows operate inside their approved boundaries, with the records to prove it. The roadmap is familiar: recalibrate existing risk assessments and audit plans to cover AI, then test six governance steps against the records each one should produce.
Short answer
Auditing AI model risk at a credit union means internal audit examines whether models and AI workflows operate inside their approved boundaries, with the records to prove it. The roadmap is familiar: recalibrate existing risk assessments and audit plans to cover AI, then test six governance steps against the records each one should produce.
How it works
Internal audit is well positioned for this, and the industry guidance says so plainly: credit unions are equipped to audit AI because audit teams already manage credit, compliance, and operational risk, and the failure mode is viewing AI-related risk in a vacuum rather than as it interacts with categories the institution already manages. An AI model is a credit risk object when it scores loans, an operational risk object when it acts in workflows, and a third-party risk object when a vendor supplies it. The audit plan does not need a new universe; it needs the existing one extended to a new class of system.
The structure of the audit follows the governance steps, each paired with its record. Governance checklists for financial institutions define six steps, inventory, risk rating, pre-use testing, human review, ongoing monitoring, and third-party oversight, each paired with the record that proves the step was done. An audit tests exactly that pairing: does the inventory entry exist and match reality, does the risk rating reflect member impact and data sensitivity, does the testing record predate deployment, do human review logs show meaningful review rather than rubber stamps, do monitoring results exist between reviews, and do vendor files contain current due diligence. Each missing record is a finding with a clear owner.
Inventory accuracy is the first audit target because everything else depends on it. AI enters credit unions through obvious channels like copilots and fraud tooling and through quiet ones: staff experimenting with public models, line-of-business teams using embedded AI features in SaaS products, and vendors adding generative capabilities to existing platforms. An audit that samples the inventory against actual system usage, vendor notices, and expense lines usually finds the inventory understated. The gap matters because ungoverned systems cannot produce the records the rest of the audit tests, and a breached organization lacking AI governance policies is a documented industry pattern, with sixty-three percent of breached organizations in one report lacking such policies.
Validation and monitoring records deserve specific audit attention for AI. For traditional and non-generative models, supervisory guidance pairs the inventory entry and materiality rationale with a validation report and monitoring results; generative and agentic AI sit outside that guidance and are governed through the institution's own AI policy and existing risk frameworks, which makes the institution's own records the only evidence there is. The audit should test whether performance monitoring catches drift past tolerance and whether the escalation path from monitoring to board reporting actually operates, by tracing one or two real events end to end rather than reviewing the policy alone.
Vendor AI extends the audit to contracts and exit paths. Third-party oversight expectations cover due diligence, contract controls, and ongoing monitoring, and AI-specific questions attach naturally: what data the vendor's model trains on, what the credit union can reconstruct about vendor-touched decisions, and how the institution would disentangle from the vendor. Audit can test whether the vendor file contains the answers and whether termination provisions for AI credentials and data have ever been exercised. The board-facing summary should state residual risk in the categories the board already governs, not in AI vocabulary the board has not adopted.
Frequency and follow-through make the audit program real. AI changes faster than annual cycles, so audit should re-examine high-impact uses, lending decisions and money movement, on a shorter cadence, and track findings to closure with the same discipline as any operational finding. Credit unions that built AI governance into audit programs report the payoff in examination posture: the institution can show what was tested, what was found, and what was fixed. That record, more than any framework document, is what converts AI adoption from an examiner concern into a managed program.
Frequently asked questions
How should a credit union structure an audit of AI models?
A credit union should extend the existing audit universe rather than inventing a new one: classify each AI system into the risk categories the institution already manages, credit, operational, compliance, and third-party, then test six governance steps against their records: inventory, risk rating, pre-use testing, human review, ongoing monitoring, and third-party oversight. Sample the inventory against actual usage to catch understatement, trace real events through the monitoring and escalation path, and report findings with owners and due dates. Recalibrating routine risk assessments this way is the roadmap audit teams are already equipped to run.
What records should the audit expect to find for AI systems?
An inventory entry with a materiality rationale, a risk rating reflecting member impact and data sensitivity, a pre-use testing record that predates deployment, human review logs showing substantive review, ongoing monitoring results between reviews, and vendor due diligence with current contract controls. For generative and agentic AI, which sit outside the revised model risk guidance, the institution's own AI policy records are the only evidence available, so their completeness is the audit finding. Every missing artifact is a finding with a named owner.
What is the first thing internal audit should test?
Inventory accuracy at the credit union, because every downstream test depends on it. AI enters through copilots, embedded features in SaaS products, vendor platform updates, and staff experimentation, and inventories are almost always understated on first sampling. The audit should reconcile the inventory against system usage, vendor notices, and spending, then flag every AI system operating without an owner or an approved use case. Sixty-three percent of breached organizations in one report lacked AI governance policies, and ungoverned systems are where that statistic starts.
How should audit handle vendor-supplied AI?
As third-party risk with AI-specific depth: verify due diligence is current, test whether contracts secure log delivery, audit rights, data-use commitments, and termination behavior for credentials and member data, and check whether ongoing monitoring actually consumes the logs received. Audit should also ask how the institution would disentangle from the vendor, since exit paths are part of examination expectations. The credit union owns the risk of vendor models regardless of who built them, so the audit tests the institution's oversight, not the vendor's internals.