AI-Assisted Collections Playbook for Consumer Lenders

What is pushing AI into collections

Three pressures arrive at once. The first is arithmetic: a person costs roughly the same to put on a $200 balance as on a $2,000 one, so every staffed operation has a balance below which an attempt costs more than it can return. Those accounts are sold, at market rates around $3.25 per $100 of principal, or written off. The second is consistency: knowing the right time, channel and offer for an account is a solvable problem, but executing it identically across tens of thousands of accounts through people with their own schedules is not. The third is the record: when a borrower discloses a bankruptcy filing or military status on a call, the obligation triggers in that moment, and a sampled QA program may never find the miss.

AI-assisted collections is attractive because it addresses all three at the level of architecture rather than effort. The marginal cost of an account becomes compute. Execution is the same on the fifty-thousandth account as on the first. And because the system generates the conversation, it can record, transcribe and review every one of them rather than a sample.

Three non-negotiables before the first contact

1. Compliance is enforced at execution, not reviewed afterwards

Before any contact the system should check consent for the channel, the time-of-day window, contact-frequency limits, opt-outs, open disputes and your own policies — and if a check fails, the contact does not happen. A program that contacts first and audits later has the same exposure as a floor with a QA sample, only faster.

2. A human is reachable, and judgment is escalated

A borrower must be able to ask for a person at any time and get one. Hardship, disputes, anything outside policy and anything the system is not confident about should route to your team with the full transcript attached. The program's escalation paths are yours to set and should be written down before launch.

3. Every contact is recorded and reviewable

Ask for call recordings and transcripts, message logs, flags for regulated events, and an export path. Review a sample against the account record before relying on a vendor's coverage claim.

Six use cases that hold up

Early-delinquency outreach

A few days past due, most borrowers need a reminder and a simple way to pay. The manual reality is that this outreach is the first thing a stretched team drops. An automated program contacts every eligible account, in the borrower's preferred channel, with payment in the same conversation.

Payment plans and hardship options

Plans within your policy limits can be offered and set up in the conversation, by voice, text, chat or email, with the borrower's choices logged. Anything outside policy — or any hardship disclosure — routes to your team.

Right-party contact and verification

Build identity checks and any required disclosures into the contact flow. Have counsel confirm the wording and timing for each channel and program.

Multi-channel coordination

Voice, text, email, chat and direct mail coordinated around one account record, so a borrower who replied by text is not called the next morning as though nothing happened. Contact-frequency limits apply across channels, not per channel.

Dispute and cease-contact handling

Outreach on the account stops immediately, the dispute is logged with its timestamp and transcript, and the account routes to your team. Nothing resumes until it is cleared.

Small-balance and charged-off accounts

The accounts a staffed queue never reaches. Because the cost of an attempt is compute, the balance stops deciding whether an account is worth contacting. See the small-balance recovery playbook for the arithmetic.

Regulatory questions for the program

Using software does not remove applicable collection rules. For debt collectors covered by the FDCPA, Regulation F addresses contact timing, call frequency, validation notices, disputes and misleading conduct. Its telephone-call rule generally presumes a violation after more than seven calls about a particular debt in seven consecutive days, or a call within seven days after a conversation about that debt; the rule has exceptions and the presumption can be rebutted. The TCPA can impose consent and other limits on certain calls and texts. Federal unfair, deceptive or abusive practices standards and state requirements may also apply. Counsel should map the rules to the collector, channel, technology, location and program before launch.

These are useful questions for a lender's risk review of an AI program:

1. How do you know consent existed for this channel at the moment of contact?
2. How do you enforce time-of-day and frequency limits, across channels, per debt?
3. How are disputes, cease requests and hardship disclosures captured and acted on?
4. What does the borrower hear or read about who is contacting them and why?
5. Show me the record for this account.
6. For AI: how does the system decide what to say, who reviews that, and how do you verify what it said?

Oversight, validation and ongoing monitoring

Treat the program like any other third-party or model risk. Before launch: review the program rules line by line with servicing, risk and legal; listen to sample conversations; walk the escalation paths end to end; confirm what is logged and how you export it. After launch: review flagged exceptions on a cadence, sample transcripts by outcome type, reconcile payments to your ledger, and track disputes, opt-outs and complaints as first-class metrics alongside recoveries.

• Program rules documented and signed by servicing, risk and legal
• Disclosure wording reviewed by counsel for each channel
• Escalation paths tested with a real handoff to a named person
• Export of conversation records and flags confirmed in your format
• Exception review cadence set and owner named
• Complaint, dispute and opt-out tracking in place from day one
• Change management: how you are told when the program's behaviour changes

What AI cannot do in collections today

• Decide your policy. Plan limits, settlement authority, hardship criteria and tone are yours to set; the system operates inside them.
• Substitute for licensing. A third-party program needs the collector's licences in the relevant states; a first-party program runs under yours.
• Resolve a genuine dispute. It stops, records and routes; a person decides.
• Guarantee an outcome. Any vendor quoting a recovery rate before seeing your book is quoting someone else's book.
• Replace the ledger. The system of record stays where it is.

Where programs fail

• Buying the agent without the program. A capable conversational agent with no consent, frequency, dispute and payment machinery around it is a liability with a nice voice.
• Auditing after the fact. If compliance is a review step rather than a gate, the program has the exposure of a floor at the speed of software.
• Tooling the constraint. If headcount caps how many accounts get worked, better tooling for the headcount does not lift the cap.
• Starting with the whole book. A pilot should be a defined batch with a defined comparison, not a servicing transfer.
• Measuring the wrong thing. Gross recoveries flatter; cash after fees on the same accounts over the same window is the number.
• No named owner for exceptions. Flags nobody reviews are decoration.

From first batch to production: a 30/60/90-day path

Risk review checklist

• Can you produce the full record for any account — every contact, every channel, every flag — on request?
• Can you show that consent was checked before each contact, per channel?
• Can you show frequency limits enforced across channels, per debt?
• Can you show what the borrower was told about who was contacting them, and when a person was offered?
• Can you show that a dispute stopped outreach the same day, and who reviewed it?
• Can you show who set each program rule and when it last changed?
• Can you show the escalation path was exercised, not just documented?
• Can you show complaint, dispute and opt-out trends alongside recoveries?
• Can you show what the program cannot do, and how that boundary is enforced?