2026-08-11 · Compliance

What a compliant AI voice call in collections actually says

Voice is where AI agents in financial services go from interesting to consequential. The quality bar cleared some time ago, the latency is conversational, and the economics on outbound collections are compelling enough that nearly every lender has run a pilot. What stops most pilots reaching production is not the voice. It is that a collections call is a regulated speech act, and the obligations concentrate in the opening seconds where an improvising model is least predictable.

The disclosure position hardened over the past two years. The FCC's 2024 ruling brought AI-generated voices inside the definition of artificial voices under the TCPA. Federal rulemaking on a universal real-time AI disclosure has stayed at the proposal stage, and state law has filled the gap unevenly: Utah requires a clear disclosure that the consumer is interacting with generative AI, California requires it for automated systems used to influence a commercial or political decision, Florida amended its telemarketing statute in the same direction. If you dial nationally you are subject to the union of those, so build for the strictest and stop tracking the delta.

Three rulebooks arriving on the same call #

Teams tend to hold one of these in their head and get surprised by the others.

• TCPA and FCC rules govern the dialling: consent for artificial or prerecorded voice, revocation handling, identification requirements. This is the layer most voice AI vendors talk about.
• FDCPA and Regulation F govern the collecting, and they are stricter than the dialling rules on hours and frequency. The mini-Miranda on initial communication, the validation information within five days covering the amount, the creditor and dispute rights, and the restrictions on third-party contact and workplace calls all live here.
• State AI disclosure statutes govern the fact that the caller is a machine, independently of the other two.

Missing any one of them makes the call unlawful regardless of how well the other two were handled. A perfect consent record does not rescue a missing mini-Miranda.

Everything that can be decided before dialling, should be #

The single most useful architectural move in voice collections is to shrink what the model is responsible for. Almost every compliance obligation on the call is a function of data you already hold at dial time, which means it belongs in a preflight check that either returns a permitted call or blocks it.

1. Local time at the consumer's location. Not your time zone, and not the area code if you have a better signal. A call outside permitted hours is a violation that no in-call behaviour can undo.
2. Frequency, across channels. Regulation F's presumptions are about attempts, and your SMS campaign, your WhatsApp reminders and your voice dialler are usually three systems that do not know about each other. They need one ledger. This is the most common gap we see, and the contact caps playbook exists because of it.
3. Consent and revocation state, including revocations captured by a human agent on a previous call, which is where they most often fail to propagate.
4. Account state. Disputed, in hardship, in bankruptcy, represented by counsel, deceased. Each of these changes whether the call may happen at all, and none of them are things you want a model discovering conversationally.

Structured this way, a blocked call produces a logged decision with a rule name attached rather than a silent absence. That log is what turns "we honour hardship flags" into something you can demonstrate.

What has to hold once the call is live #

Three controls run per turn rather than per call.

Amounts are read, never computed. The outstanding balance, the settlement figure, the instalment amount all come from a tool call against the system of record and are spoken as returned. A model that arithmetics its way to a number will eventually state a wrong one with total confidence, and on a recorded line that is a misrepresentation.

Offers stay inside numeric bounds. Settlement percentages, term lengths and fee waivers are configuration with hard limits. Anything outside routes to a human. This matters more on voice than on chat because the pressure of a live conversation is exactly the condition under which a model becomes accommodating.

Handover triggers are deterministic. A stated dispute, a hardship declaration, a request for a human, any mention of counsel or bankruptcy, sustained distress. These end the automated portion of the call. On voice the transfer has to carry context so the customer does not repeat themselves, which is a warm-transfer problem worth solving properly rather than a callback promise. The voice playbook covers the wiring, and the handover desk playbook covers the receiving end.

Stop conditions deserve their own review #

Ask a team to demo their voice agent and you will see the happy path. Ask them to demonstrate the stop conditions and you learn what was actually built. Worth testing explicitly, with recorded synthetic calls:

• The consumer says stop calling. Does the revocation land in every channel's ledger, or just the dialler's?
• Somebody else answers. Does the agent avoid disclosing the debt to a third party?
• The consumer says they are at work and cannot take calls there. Is that captured as a durable restriction, or as a sentence in a transcript?
• The consumer disputes the amount. Does the agent stop negotiating?
• Voicemail picks up. Does the message meet the requirements for a message left for a consumer, and does it avoid disclosing the debt to anyone else who might hear it?

Each of these is a scripted synthetic conversation that asserts on an outcome, run on a schedule against production configuration. That is the difference between believing the controls hold and knowing when they stopped.

What you should be able to produce per call #

The complaint arrives months later and names a date. What resolves it quickly:

• The preflight decision, with the hours, consent and frequency inputs as they were evaluated at that moment.
• The disclosure and mini-Miranda as spoken, tied to the configuration revision they came from, so "which script was live in March" is a lookup.
• Every tool call and result, so a stated balance can be traced to the system of record rather than defended as plausible.
• Every guardrail decision including the blocks, because the blocks are the evidence the controls were running.
• The handover, if there was one, with the trigger that caused it.

Recording the call is table stakes and does not cover this. An audio file tells you what was said, not why the system decided to say it, and the second question is the one that arrives from the second line of defence. See what a defensible record contains for the general form.

Where AI voice does not belong yet #

Being straight about scope makes the rest more credible. Early-stage reminders, payment-intent capture, arrangement confirmations and inbound balance queries are good fits: bounded, high volume, and the failure modes are recoverable. Hardship assessment, dispute resolution and anything involving a vulnerable customer are not, and the right design has the agent recognise them and hand over rather than attempt them well.

The teams that get voice into production are not the ones with the best model. They are the ones that narrowed the job until the compliance surface was small enough to enforce outside the model, then proved the enforcement with synthetics on a schedule.

Related: the collections use case for the control set end to end, and SMS collections for the channel most deployments start with.

← All posts