An AI voice agent for debt collection is easy to demo and hard to run safely. The voice is the simple part. The hard part is a set of rules about who you may call, when, how often and what you must say. A model cannot be trusted to remember those rules on its own.
This guide is for heads of collections, recovery operations and compliance at agencies, debt buyers, lenders and utilities. It covers why collections has to be designed compliance-first, a call matrix for what the AI resolves and what stays human, the architecture that keeps the rules out of the prompt, a rollout sequence and a vendor checklist. It focuses on U.S. rules. It is an engineering view, not legal advice: take every design decision below to your counsel.
Why collections is a compliance problem first
The U.S. Consumer Financial Protection Bureau's own data shows where collection programmes break, and much of it happens on the phone.
- Volume of complaints. The CFPB received about 207,800 debt collection complaints in 2024, seven percent of all the complaints it received that year (CFPB, FDCPA Annual Report 2025, November 2025).
- Call frequency is the main communication complaint. In the same report, 51% of complaints about communication tactics concerned frequent or repeated calls. A further 34% concerned continued contact after the consumer asked the collector to stop, and 5% concerned calls outside the 8:00 a.m. to 9:00 p.m. window.
- The safe harbour is not a target. CFPB examiners found collectors that stayed inside Regulation F's call-frequency limits but placed over 100 calls to consumers who had asked them to stop. The examiners ruled that the presumption of compliance had been overcome and the conduct was harassing.
- AI voices count as artificial voices. In a declaratory ruling adopted February 2 and released February 8, 2024, the FCC confirmed that AI-generated voices are “artificial” under the Telephone Consumer Protection Act (FCC 24-17). In practice, outbound AI calls generally need the called party's prior express consent.
Every one of those failures is a counting, timing or memory failure. These are exactly the things a language model does badly and a database does well.
A collections voice agent is not judged on how persuasive it sounds. It is judged on whether it can prove it called the right person at the right time, the right number of times.
What an AI voice agent for debt collection should handle
Rank call types by volume and by the damage a wrong answer does. This is the starting matrix we would bring to a collections operation.
| Call type | AI role | Why |
|---|---|---|
| Inbound balance, due date and “why did I get this letter?” | Resolve | A read-only lookup after right-party verification. |
| Payment in full or a one-off payment | Resolve within rules | Route card or bank details to a dedicated payment flow so they never enter the transcript or the model. |
| Payment plan or promise to pay | Resolve within rules | The AI can only offer plans a rules engine has pre-approved for that account. It never improvises an amount or a date. |
| “Stop calling me” or “don't call this number” | Resolve immediately | Record the request as a structured flag that every dialler and every human agent reads. No retention pitch. |
| “I don't owe this”: dispute intake | Assist | Capture the dispute and its reason, pause the automated workflow on the account and pass it to the dispute team. |
| Wrong person or third party answers | Resolve, say nothing | No mention of a debt. Leave a callback request and update the contact record. |
| Hardship, medical, bankruptcy, deceased account holder, attorney representation | Keep human (AI collects) | Specific legal handling and real judgement. The AI's job is fast identification and a complete hand-off. |
| Settlement requests outside approved bounds | Keep human | Money and precedent. A collector decides; the AI prepares the case. |
| Distressed or angry callers | Keep human | Transfer at the first clear signal, with full context, so the consumer does not repeat the story. |
A cease request is the one call where the AI must be more reliable than a person: the CFPB data above shows what happens when it does not propagate. And because collectors are restricted in what they may reveal to anyone other than the consumer, the AI's default is silence about the debt until verification succeeds.
Architecture: the rules live in code, not in the prompt
We hold one rule internally: if a behaviour matters, it is enforced outside the model. I made the general case in Confidently Wrong. In collections the argument is stronger, because the cost of a wrong answer is a regulatory finding rather than a disappointed customer. A collections voice agent needs four deterministic components around the conversational layer.
Three design choices that go beyond the minimum
- Count per person, not just per debt. The 7-in-7 presumption is set per debt, but a consumer with three accounts experiences calls per phone. A conservative ledger caps attempts per person across every debt and every caller, human or AI.
- Treat the cap as a ceiling, not a quota. Set your own lower frequency and a hard stop after any cease or channel request. The 100-call finding above happened inside the presumption.
- Make the AI fail closed. If the ledger, the time-window gate or the account system does not answer, the agent does not dial. On an inbound call, it takes a callback request instead of guessing a balance.
Underneath all of this sits security: encryption in transit and at rest, card data kept out of the conversation, and data-processing terms agreed before go-live. Our guide to AI CX security and privacy covers the platform questions in more depth.
Rollout: inbound first, outbound with legal in the room
The fastest safe path starts where the consumer is already calling you. Inbound callers have chosen to engage, the consent question does not arise in the same way, and the calls are repetitive.
- Week 1: call-flow audit. Classify 90 days of inbound calls by reason and outcome: balance, payment, plan, dispute, cease, hardship, wrong party. Map each reason to a row of the matrix above.
- Weeks 2–3: integrations and rules. Read access to the collection platform, the payment flow, the contact ledger and the offer rules. Write the disclosures and voicemail templates with compliance and lock them.
- Weeks 4–5: adversarial testing. Replay the awkward cases: a spouse who answers, a caller who mentions bankruptcy mid-call, a cease request phrased as “just leave me alone”, a payment-plan request one dollar outside bounds.
- Week 6: go live on overflow and after-hours inbound. Widen to first-ring answering once transcripts and event logs pass a compliance review.
- Later: outbound reminders. Only with documented consent, a separate legal review and the ledger already proven on inbound.
Unicall's done-for-you deployments typically go live in 4 to 6 weeks without replacing the existing phone system. The agent answers only from an approved knowledge base, routes low-confidence cases to a person and logs every call with a transcript. That timeline comes from our general customer-care deployments, not from a collections-specific benchmark, so treat it as a reference. The managed service starts from $1,000 a month, billed on ticket volume rather than seats (see pricing).
Buyer checklist: questions for any collections voice AI vendor
If you are comparing specialist voice platforms, our comparison hub applies the same ten buyer criteria to each vendor, using public information only.
FAQ
Can an AI voice agent make debt collection calls in the US?
It can, under the same rules as any other collection call, plus the rules for artificial voices. In February 2024 the FCC ruled that AI-generated voices are “artificial” under the Telephone Consumer Protection Act, so outbound AI calls generally need the called party's prior express consent. Debt collectors covered by the FDCPA must also respect Regulation F's call-frequency presumptions, calling hours, disclosures and cease-contact requests. Inbound calls, where the consumer rings you, avoid the consent question and are the safer place to start. This is not legal advice; confirm your programme with counsel.
How does Regulation F limit collection calls?
A debt collector is presumed to comply with the harassment rules if it calls a person about a particular debt no more than seven times within seven consecutive days, and not within seven days after having a telephone conversation with that person about the debt. Calls before 8:00 a.m. or after 9:00 p.m. local time at the consumer's location are presumed inconvenient. The presumption can be rebutted: the CFPB's 2025 FDCPA report describes collectors who stayed within the call limits but placed over 100 calls after consumers asked them to stop, which examiners found harassing.
Does an AI collections agent have to say it is a debt collector?
Under Regulation F, the initial communication with a consumer must disclose that the debt collector is attempting to collect a debt and that any information obtained will be used for that purpose, and later communications must disclose that they come from a debt collector. A voicemail that contains only the content Regulation F allows in a limited-content message is not a communication and does not carry those disclosures. Build both as fixed templates the AI cannot rephrase, and play the disclosure only after the right person has been verified.
Which collection calls should stay with a human?
Disputes beyond simple intake, hardship and medical situations, bankruptcy, deceased account holders, consumers represented by an attorney, settlement requests outside your pre-approved bounds, and any caller who is distressed or asks for a person. The AI should still answer, verify, capture the facts and stop the automated workflow, then hand the case to a trained collector with the context attached.
Should we start with inbound or outbound collection calls?
Inbound. Consumers calling back after a letter, a statement or a missed payment are already engaged, the consent question does not arise in the same way, and the calls are repetitive: balance, payment, payment plan, dispute intake. Prove identity verification, disclosures and logging on inbound first, then scope outbound reminders separately with your compliance team.
Unicall.ai builds and operates done-for-you AI voice agents that answer every call 24/7, take real actions in your systems, and hand over to your people, with full context, the moment they should.