A Stop-or-Continue Playbook for AI in Credit Union Collections

Credit unions should permit collections AI to continue only when the account data is complete, the next action is preapproved and no exception signal is present. The workflow should stop for human review when it detects a dispute, hardship, legal or protected status, fraud, a consequential credit decision, conflicting data, a channel restriction or an output the credit union cannot explain.

This is the difference between using AI as a queue assistant and allowing it to become an unbounded collector. A system can summarize notes, rank routine work, prepare an approved message or propose a next action. It should not improvise through a member's contested balance, bereavement, bankruptcy notice or request for accommodation.

The NCUA's current AI resource says existing requirements remain technology-neutral and identifies algorithmic decisions, fair lending, privacy, resilience, model risk and vendor oversight as issues credit unions must manage. Collections leaders therefore need a control that is more specific than “human in the loop”: a defined machine lane, named stop signals and an auditable handoff.

First, define the routine lane

Start with actions that do not change a member's rights or the credit union's position: assembling account facts from approved systems, checking whether a promised payment posted, drafting from a locked template, scheduling an employee's work or highlighting a missing document. For each action, name the data source, permitted output, approval owner and maximum time before the case must be refreshed.

The system should not infer permission from silence. If a required field is absent, an account status is stale or two systems disagree, the case leaves the routine lane. That simple rule prevents a fast workflow from turning incomplete data into false certainty.

Eight events that must stop automation

1. The member disputes the debt, amount or ownership

Words such as “not mine,” “already paid,” “wrong amount” or “identity theft” should open a dispute route, preserve the original communication and halt conflicting activity. Regulation F contains specific communication, validation and dispute requirements for covered debt collectors, including stopping collection in defined circumstances. A credit union collecting its own debt is not always a “debt collector” under the federal rule, but state law, product rules and service-provider arrangements can change the analysis. Counsel should set the applicable stop logic; the AI should not decide coverage.

The operative requirements are available in the current text of 12 CFR Part 1006, including communication limits and dispute handling. The workflow should capture the member's exact statement, the disputed portion, date, channel and all activity paused.

2. The member signals hardship or asks for help

Job loss, illness, disaster, reduced hours, caregiving, military orders and an inability to make the proposed payment require judgment. Route the case to an employee who can review approved assistance options and explain consequences. AI may prepare facts; it should not pressure a member, invent a concession or optimize solely for dollars collected.

3. A legal or protected status appears

Bankruptcy, representation by counsel, deceased-member status, active-duty protections, guardianship, language access or a disability-accommodation request should trigger the appropriate specialist queue. The handoff must carry the source document, effective date and activity already taken, not merely a generic risk flag.

4. Fraud or identity uncertainty is present

A changed device, compromised credentials, disputed transaction or synthetic-identity signal belongs in a coordinated fraud and collections case. Do not let separate models send contradictory messages while the institution is still determining who acted and what amount is valid. Our reporting on fraud-claim investigations after impersonation scams shows why credentials are evidence, not a substitute for reconstructing the event.

5. The next action changes credit terms or access

A decision to reduce a line, revoke credit, decline a modification or impose materially different terms may carry notice and fair-lending consequences. If the action constitutes adverse action under the Equal Credit Opportunity Act and Regulation B, the creditor must provide specific, accurate principal reasons. The CFPB's complex-algorithm circular says opacity is not an excuse. Legal and compliance teams should map exactly which workout and account-management decisions are covered.

6. Systems disagree or the record is incomplete

A balance mismatch, missing payment, outdated contact preference, unclear promise-to-pay note or inconsistent delinquency status should stop the workflow. The human reviewer needs both values, their timestamps and source systems. Do not silently choose the field that advances the collection sequence.

7. Contact rules or preferences are uncertain

Before any automated call, email, text or app message, confirm the allowed channel, time, recipient, consent and opt-out state. Regulation F restricts inconvenient communications, prohibited media and public social messages for covered collectors. Even when a particular provision does not apply, the credit union should avoid channel behavior it cannot defend as accurate, respectful and member-authorized.

8. The recommendation cannot be explained or reproduced

If the credit union cannot show the input data, rule or model version and approved policy behind a recommendation, the case stops. A confidence score alone is not an explanation. The employee should be able to reproduce the material account facts and identify why the proposed action is permitted.

Build the handoff packet before launch

Every exception should arrive with the member and account identifiers, triggering signal, source records, recent contacts, promises and payments, deadlines, prohibited actions and the last action taken. Assign an owner and service level by exception type. Test whether the employee can resolve the case without searching across several disconnected systems.

Then measure the control, not just the automation. Track false continuation, unnecessary escalation, time to qualified review, repeated member contacts, broken promises, corrected balances, complaints and outcomes by member segment. Sample cases the model labeled routine and compare them with cases employees escalated. That review connects the exception design to the broader AI workforce plan and the document-automation controls needed upstream.

The safest scaling sequence is narrow: automate preparation, prove the stop conditions, improve the handoff, and expand authority only when evidence shows that members and the credit union are better served.