A Hardship-Treatment Record for AI-Assisted Credit Union Collections

A collections model can make a delinquency queue look orderly. It can predict contact, cure or loss; recommend a channel; summarize prior conversations; and suggest which account should receive attention first. The danger is that operational efficiency can obscure whether members with similar circumstances received comparable options, whether a hardship treatment actually helped and whether a bad recommendation was corrected.

Credit unions need more than a model log. They need a treatment record that follows a member from the first material recommendation through contact, assistance, performance and remediation. The record connects what the system knew, which policy applied, what a person decided, what the member was told and what happened next.

Appendix B to NCUA Part 741 directs federally insured credit unions to adopt written loan-workout policies and controls that are commensurate with the scope of activity. NCUA’s loan-workout examination guidance emphasizes sound controls, accurate reporting and board-approved policies. Those requirements do not prescribe an AI design, but they establish the operating baseline: assistance must remain governed by approved policy, documented judgment and reliable records.

The record begins before the recommendation

Start with the account and member facts the system was permitted to use, their source, effective date and known limitations. Identify the product, delinquency stage, prior contacts, promises, disputes, accommodations and active protections. Flag missing or conflicting data rather than allowing a model to infer certainty.

Then attach the approved policy version in force at the decision time. The system may map facts to eligible options, but eligibility rules, prohibited practices, contact limits and escalation paths must come from controlled policy—not from a generated summary or a stale configuration.

Preserve six linked decisions

1. Queue position

Record why the account entered a work queue, which model or rule assigned priority, the score version and the cutoff. Keep legally protected or operationally irrelevant variables out of ranking. If a person changes priority, capture the reason rather than overwriting the original recommendation.

2. Contact strategy

Store the approved channel, timing, language, consent state and contact history. If the activity is covered by the federal debt-collection rule, the CFPB’s Regulation F resources provide a useful boundary for communications and recordkeeping. Credit unions should determine applicability with counsel instead of assuming that every first-party collection falls inside—or outside—the same rule.

3. Hardship assessment

Separate observed facts from model inference. Income interruption, medical expense, disaster effects or other member circumstances should be recorded from a controlled source or a documented member statement. Do not let a probability score become a finding that the member can or cannot pay.

4. Treatment offer

List the options evaluated, the policy authority, material terms, fees, interest effects, credit reporting implications, expected payment path and reason an option was offered or withheld. The accountable employee approves the treatment and can explain it in plain language.

5. Member decision and communication

Preserve the offer, required disclosures, language and accessibility accommodations, the member’s acceptance or decline, and the date and channel. A generated call summary is a convenience; the retained communication and authoritative servicing record remain the evidence.

6. Performance and remediation

Follow the treatment through first payment, redefault, cure, charge-off, further assistance and complaint. If the system used bad data, misapplied policy or produced an inconsistent recommendation, record the correction, affected population, member remediation and control change.

Monitor treatments as cohorts, not anecdotes

A single successful workout does not validate a strategy. Compare similarly situated cohorts across treatment, channel, product and member segment. Show the number eligible, offered, accepted, completed, redefaulted and remediated. Keep denominators and observation windows visible so a small or immature cohort is not mistaken for proof.

At minimum, review contact success, promise kept, cure, redefault, fees, time in delinquency, complaints, disputes, repossession or foreclosure where relevant, charge-off and net loss. Add service measures such as repeat contacts and time to an authorized decision. The purpose is not to optimize for the highest short-term collection rate; it is to understand whether the treatment is effective, consistent, sustainable and member-safe.

Outcome monitoring should include overrides. A high override rate can signal weak model fit, bad inputs or policy ambiguity. A very low rate can signal automation bias. Sample both approved and declined assistance decisions, including cases that never generated a complaint.

Keep adverse-action reasoning separate

Some servicing or workout decisions can implicate the Equal Credit Opportunity Act and Regulation B. The CFPB has stated that creditors using complex algorithms still must provide specific principal reasons when an adverse-action notice is required. Its complex-algorithm circular is a useful reminder: a model’s opacity is not an acceptable substitute for a legally sufficient explanation.

Do not repurpose an attention score or a generated narrative as the reason. Map the actual decision factors to controlled reason codes, verify them against the account record and preserve the notice delivered. Compliance should determine when notice rules apply and test explanations separately from the collections model’s predictive performance.

Use stop conditions for the system and review triggers for the strategy

The existing collections exception playbook addresses real-time stop conditions: missing consent, active disputes, protected status, contradictory records and other cases that require a manual path. The treatment record solves a different problem. It lets leaders determine whether the strategy itself should continue, change or be suspended.

Set review triggers for material drift in approval, cure, redefault, complaint, hardship-access or remediation rates; unexpected subgroup gaps; stale policy mappings; unresolved data defects; vendor changes; and repeated unexplained overrides. The NIST AI Risk Management Framework supports monitoring systems in context and responding when risk or performance changes.

The minimum durable packet

For each material treatment, retain the authoritative inputs, model and rule versions, policy version, recommendation, alternatives considered, human decision and override reason, member communication, notices, performance outcome, complaints and remediation. For management review, aggregate those records with defined cohorts, denominators, exceptions, uncertainty and an accountable action owner.

The operating test is straightforward: months after a decision, can the credit union reconstruct what was known, why the treatment was authorized, what the member received and whether the outcome exposed a control problem? If the answer depends on a vendor dashboard, a generated summary or an employee’s memory, the collections program does not yet have a durable treatment record.