Credit Union Fraud and BSA/AML AI: Production Patterns That Survive NCUA Exams

Credit unions don't fail NCUA exams because they lack a dashboard. They fail when alert logic can't be explained, when BSA triage leaves no decision trail, or when a vendor model drifts and nobody can show what changed. AI that survives exams is boring on purpose: typed inputs, reproducible scores, human ownership, and documentation that matches production behavior.

This article covers production patterns we use across credit union AI engagements, with focus on card and ACH fraud and BSA/AML triage. For a related large-institution build pattern, see our federal fraud detection case study.

Exams care about process, not model brand

Examiners ask how you detect, how you investigate, how you escalate, and how you prove the program works. A transformer that can't answer those questions is a liability. Design every AI surface as evidence for the BSA officer and fraud manager, not as a black-box score for a slide deck.

• Model inventory: what runs where, who owns it, last validation date, and approved use.
• Change control: prompt, feature, and threshold changes logged with approver and rollback path.
• Sample testing: periodic pulls of alerts with documented outcomes, including false positives you willingly accept.
• Vendor oversight: if a third party scores transactions, you still own the program—contracts and monitoring must say so.

Fraud: separate real-time block from investigative assist

Card and ACH fraud systems that try to do everything in one agentic loop usually do nothing well. Split the stack:

1. Real-time decisioning: low-latency rules + scored features for decline/allow/step-up. Deterministic fallbacks when the model is unavailable.
2. Case assembly: enrichment that pulls member history, device, geo, and prior disputes into a coherent packet for analysts.
3. Feedback loop: confirmed fraud and confirmed legitimate labels flowing back into training and threshold review on a schedule, not ad hoc.

Latency budgets matter. A beautiful LLM narrative that arrives after the ACH window closed is research, not controls. Put generative assist on the investigation side; keep authorization-path models narrow, monitored, and fail-safe.

BSA/AML triage: kill the alert firehose without losing the story

Most credit unions drown in alerts that were never risk-ranked for their membership profile. Production BSA AI should:

• Prioritize by risk factors your board already approved in the BSA policy
• Group related alerts into cases instead of forcing one-by-one clicks
• Draft narrative stubs from structured facts—never invent member intent
• Require analyst disposition codes that map to SAR / no-SAR / escalate
• Retain the full decision packet for the retention period your counsel specifies

Generative drafting is useful only when grounded in retrieved transactions and known member attributes. Free-form storytelling that can't be traced to source rows will not survive examiner sampling.

Production patterns that hold up under scrutiny

• Human-in-the-loop by design: AI ranks and drafts; humans dispose. Autonomous SAR filing is a non-starter for most CUs.
• Explainability at the feature level:"why this alert" in language a BSA analyst can repeat, not SHAP theater for data scientists only.
• Shadow mode before cutover: run new scores beside existing rules for a defined window; compare lift and miss rates before touching member impact.
• Segment-aware thresholds: business accounts, shared branching, and seasonal agricultural members need different baselines than consumer debit.
• Operational SLAs: time-to-first-touch on high-risk alerts is a control metric. Track it like you track charge-off.

What to build vs buy for credit unions

Core processors and card networks already ship detection modules. Buy the commodity rails. Build (or tightly configure) the layers that encode your risk appetite, membership mix, and exam narrative:

1. Case management and disposition workflow tailored to your org chart
2. Risk scoring calibrated on your history, not a national average alone
3. Documentation packs examiners can walk without a vendor present
4. Integration to your core, card processor, and wire channels with audit logs

If you're prioritizing roadmap items this year, start with measurable alert reduction that preserves true-positive recall, then layer generative assist on investigation packets. For implementation patterns across fraud and BSA triage, or to pressure-test an existing vendor stack before exam season, start a discovery conversation.