AI voice agents for debt collection: a buyer’s guide

What to evaluate before you put an AI voice agent on collections calls — from the “principal purpose” trap to the consent, disclosure, and frequency records that make or break an audit.

An AI voice agent that answers and places collections calls can lift contact rates and free your licensed collectors for the conversations that need a human. But collections is one of the most heavily regulated places to deploy voice AI, so the buying decision is less about how natural the voice sounds and more about what the system records, discloses, and refuses to do. This guide walks the questions that actually matter.

Start with the structural question: who is the collector?

The safest architecture is simple. The AI is software operated by a licensed collector — it is never the collector itself. If a vendor’s own agents place calls and negotiate on debts, that vendor risks being treated as a debt collector whose “principal purpose” is collection, which pulls it under the FDCPA and state licensing regimes such as California’s Debt Collection Licensing Act.

Ask any vendor to state plainly where they sit. You want a tool you configure and control, running under your license and your name in the mini-Miranda — not a service that inserts itself between you and the consumer.

Consent and disclosure are not optional features

The FCC has ruled that AI and cloned voices are “artificial voice” under the TCPA, so an artificial-voice call to a cell phone needs prior express consent. There is no personal or informational carve-out that removes that requirement. On top of that, Reg F requires the mini-Miranda in the language of the call, and honesty about what the caller is — an undisclosed human-sounding bot stresses the ban on false or misleading representations.

• Does the agent disclose that it is AI, in plain language, near the top of the call?
• Does it capture and store proof of prior express consent for artificial-voice calls to cells?
• Does it deliver the mini-Miranda in the language the conversation is being conducted in?
• Does it default to all-party recording consent with an audible notice?
• Can it produce these records per call, on demand, for an examiner?

Frequency limits demand a real ledger

Reg F creates a rebuttable presumption of a violation once you exceed seven calls per debt in seven days. NYC goes further, capping three communications per account per seven days across all channels combined, effective January 2027 — a cross-channel count most dialers were never built to keep. The buying test is whether the system maintains one ledger that unifies calls, texts, and other contacts per account, not a per-channel counter that quietly double-books the consumer.

What good looks like in practice

Teltura is built around these constraints rather than bolting them on. It discloses that it is AI, records with the right consent, and keeps a compliance ledger tracking consent, disclosure, recording, and cross-channel contact frequency. It warm-transfers to a human on request and reports when the human didn’t answer, edits are self-serve — “No ticket. Ever.” — and it never takes a card number by voice.

The real deliverable in AI collections is the audit trail. Buy for the consent records, the disclosure logs, and the frequency ledger first — the natural voice is table stakes, and compliance is the product, not the overhead.