Do I need consent to record calls handled by an AI voice agent?

It depends where the caller is, which is exactly why most businesses should stop trying to work it out per call. United States federal law sets a one-party consent baseline, and several states go further and require every party to consent, California, Florida, Illinois, Pennsylvania and Washington among them. Calls that cross state lines are where the neat table stops helping, because you rarely know where someone is when they dial. The practical answer is to disclose recording at the top of every call, in every state, and stop maintaining a map. A short line costs nothing in conversion when it is delivered as information rather than as a warning, and building state-aware recording logic is real engineering that can be wrong silently. Three operational details matter more than the legal geography. The disclosure has to come before the capture, which makes your greeting a compliance artifact rather than branding. The disclosure should be logged as a structured event with a timestamp next to the call, because saying it and being able to prove it are different systems. And there has to be a tested path for the caller who objects, either stopping the recording and continuing or routing to a human line that is not recorded. This is the shape of the problem rather than legal advice, so confirm the specifics for your jurisdiction with counsel.

Can an AI voice agent make outbound calls?

Yes, but outbound is where most of the enforcement risk in voice AI actually lives, and it should be scoped separately from inbound. Inbound is forgiving because the caller chose to dial you. In the United States, calls placed with an artificial or prerecorded voice fall under the TCPA and generally require prior express consent, and the FCC confirmed in February 2024 that AI-generated voices count as artificial for this purpose. Several states add their own rules on top. In practice that means four things. Only call numbers that were given to you for this purpose, and keep the consent record with a timestamp and its source, because "it was in the CRM" is not a consent record. Honour opt-outs across channels, since someone who asked to stop receiving texts has not agreed to keep receiving calls. Identify the business in the first seconds and give a working way to stop. And watch your number's reputation, because carrier spam-labelling damages your answer rate long before anyone complains. Not all outbound is equal either: calling back someone who just rang you and got no answer is a very different proposition from working a cold list, and appointment reminders to existing customers sit somewhere in between. The cleanest way to build this is to keep consent state as a field on the contact record that the system checks before dialling.

Recording, Consent and Disclosure: What Your AI Voice Agent Has to Tell Callers

Q: Does an AI receptionist have to tell callers it is AI?

In a growing number of places, yes, and everywhere else it is still the right call. California's bot disclosure law covers certain commercial interactions, Utah requires disclosure on request in regulated occupations, and the EU AI Act carries transparency duties for systems that interact with people. That list has only ever grown. Rules aside, concealment does not work: callers usually work it out, and the ones who feel misled complain publicly rather than to you. Two things make disclosure cheap. The first is wording. Leading with capability rather than apology, by naming the assistant, saying what it can do and offering a person on request, reads as service, while an anxious "I am only an automated system" reads as a warning label. The second is a specific build requirement that most deployments get wrong: when a caller asks directly whether they are speaking to a person, the agent has to answer plainly, with no deflection and no charming non-answer. That behaviour regresses silently after unrelated prompt edits, so it belongs in a regression set that is re-run on every change, including platform updates you did not schedule. One extra detail for multilingual deployments: the disclosure has to be delivered in the language the caller is actually speaking, not only in English.

TL;DR: Three separate obligations get collapsed into one question and then handled badly. May you record the call? Must you say the voice is AI? What are you allowed to do with what the call captured? They come from different bodies of law and each has a different fix. Outbound calling is a fourth animal, and the strictest. Here is how the three differ, why a disclosure line on every call is cheaper than a routing table, where outbound rules bite, what your calls leave behind in four places, and how to build so compliance is configuration rather than a rewrite.

We build voice systems, we are not anybody's lawyers, and these rules differ by state, country and industry and are changing quickly. Treat this as the operational shape of the problem, the part a vendor should be able to explain, then confirm the specifics with counsel who knows your jurisdiction.

The pattern we see most often is a business that asked one question, got one answer, and assumed it was covered. It usually was not, because the question they asked was the easy one.

The three questions are not the same question

Can you record? This is wiretap and eavesdropping law, and it is decided before the conversation starts. It is about capture.

Do you have to say it is AI? This is bot disclosure and consumer protection law, and it is about the caller's understanding of who they are speaking to.

What can you keep and do with it? This is privacy, sector regulation and your vendor contracts. It applies for as long as the data exists, which is usually much longer than anyone intended.

They are independent. You can be entitled to record a call and still be non-compliant because the agent implied it was human, or disclose perfectly and still have a problem because a transcript containing health information sat with a vendor who never agreed to hold it. Naming the three separately is the cheapest thing on this page.

Recording: one-party, all-party, and the rule that saves you the thinking

In the United States, federal law sets a one-party consent baseline, and a number of states go further and require every party to consent. California, Florida, Illinois, Pennsylvania and Washington are among them. Calls that cross state lines are where the neat table stops helping, because you rarely know where the caller is when they dial.

That leads to the rule that saves most businesses a lot of thinking: disclose on every call, in every state, and stop maintaining a map. A short line at the top costs nothing in conversion when it is delivered as information rather than as a warning. Building state-aware recording logic is real engineering, it can be wrong silently, and almost nobody needs it.

Three operational details matter more than the legal geography:

The disclosure has to come before the capture. That makes your greeting a compliance artifact, not just branding, and nobody should edit it casually.
Log the disclosure with the recording. A structured event with a timestamp is what turns "we always disclose" into something you can show.
Have a path for someone who objects. Either stop recording and continue, or route to a human line that is not recorded. An agent with no branch for that request will improvise, which is the last thing you want here.

If you operate in more than one language, the disclosure has to be in the language the caller is speaking, a detail that quietly breaks on multilingual deployments.

"Am I talking to a robot?"

Two forces push the same way here.

The first is law. Some jurisdictions already require it: California's bot disclosure law covers certain commercial interactions, Utah requires disclosure on request in regulated occupations, and the EU AI Act carries transparency duties for systems that interact with people. The list gets longer, not shorter.

Concealment does not work anyway. Callers usually clock it, and the ones who feel misled complain publicly rather than to you.

The build requirement that follows is small and specific: if a caller asks whether they are speaking to a person, the agent answers plainly. No deflection, no charming non-answer, no changing the subject. This regresses silently after unrelated prompt edits, which is why it belongs in the regression set we argue for in AI voice agent monitoring, tested on every change rather than checked at launch.

Two things to get right in the wording. Lead with capability rather than apology: naming the assistant, saying what it can do and offering a person on request reads as service, while "I am only an automated system" reads as a warning label. And be careful with voice cloning. Using a synthetic voice is ordinary, but cloning a specific real person's voice without their written permission, or implying the caller is speaking to a named employee, stops being a disclosure question and becomes a different kind of problem.

Outbound is a different animal

Inbound is the forgiving direction, because the caller chose to dial you. Outbound is where most of the enforcement risk in voice AI actually lives.

In the US, calls placed with an artificial or prerecorded voice fall under the TCPA and generally require prior express consent, and the FCC confirmed in February 2024 that AI-generated voices count as artificial for this purpose. Several states layer their own rules on top. The practical consequences:

Only call numbers that were given to you for this purpose, and keep the consent record with a timestamp and its source. "It was in the CRM" is not a consent record.
Honour opt-outs across channels. Someone who asks to stop hearing from you by text has not asked to keep receiving calls.
Identify the business in the first seconds, and give a working way to stop.
Watch your number's reputation. Carrier spam-labelling is an operational problem before it is a legal one, and it takes your answer rate down long before anyone complains.

Not all outbound is equal. A callback to someone who just rang you and got no answer is a different proposition from a cold list, which is one reason the missed-call revenue leak is such an easy first project. Reminders to existing customers sit in between, and the booking work in the AI appointment setter guide is worth mapping against what each contact agreed to. A vendor happy to point an outbound agent at a purchased list is telling you about the rest of their practice.

What the call leaves behind

One call produces four artifacts, in four places, with four retention answers: the audio, the transcript, the structured fields written into your CRM or calendar, and whatever logs sit with your telephony and model providers. Most businesses have thought about the first and none of the others.

Regulated data. If calls can contain health information, every processor in the chain needs an agreement covering it, including the telephony layer and the model provider, and not all will sign one. Ask before you build. The vertical version is in our AI receptionist for dental practices guide.
Card numbers. Do not take them on a recorded AI call. A recording that captured a card number pulls your whole recording store into a compliance scope you did not plan for. Send a payment link instead.
Retention. Pick a period per artifact and enforce it with a job that actually runs. Deletion has to reach vendor storage too, not only your own database.
Training. Read the terms on whether your calls can be used to improve someone's models. Zero retention arrangements are available from several providers. Ask for it in writing rather than assuming.
Access. Decide who on your side can play a recording back. Call audio is some of the most sensitive material a small business holds, and it is usually the least access-controlled.

Ownership belongs in the same conversation. If you cannot export your recordings, transcripts and agent configuration, you do not really own the deployment, which is the argument we make in software ownership and handover. The same discipline for a different sensor is in XR data and privacy.

Build it as configuration, not as a rewrite

All of this is cheap designed in and expensive retrofitted. Six things to insist on:

Disclosure text as configuration per number and per campaign, versioned, not buried in a prompt where an unrelated edit can drop it.
A logged consent event stored as structured data next to the call, so proof is a query rather than an archaeology project.
A stop-recording branch that is tested, not theoretical.
Honesty and refusal cases in the regression set, re-run on every change, including platform updates you did not schedule.
A retention job with a report, because an unenforced retention policy is just a document.
Consent state on the contact record, so outbound eligibility is a field the system checks, not a judgement someone makes.

None of that is difficult at build time. All of it is painful to add to a live deployment with six months of recordings behind it.

Questions to ask before you sign

Where is the disclosure text stored, and who can change it?
What does the agent say when a caller asks if it is a real person, and can you show me that call?
Where are recordings and transcripts stored, for how long, and who can play them?
Will you sign the agreements my industry requires, and which subprocessors are in the chain?
Are my calls used to train anyone's models, and can I get that in writing?
For outbound, where does consent come from and how is it recorded against each contact?
If I leave, do I get the recordings, transcripts, number and agent configuration?

Where we fit

We build voice systems where the call is not a low-stakes interaction. CallGuard AI answers, books and captures revenue around the clock. CallSetter AI qualifies and books for an agency's client base, so the disclosure and consent posture has to survive being resold under someone else's brand, a burden we break down in white-label voice AI for agencies. Fortell AI handles intake in over 100 languages for Community Action Agencies, where the caller is often in a difficult situation and the disclosure has to land in their language, not only in English.

Evaluating a first deployment rather than fixing an existing one? Start with the AI receptionist buyer's guide.

Not sure what your voice agent is telling callers, or what it is keeping? Book a demo and we will walk through your disclosure, consent and retention setup honestly, including telling you where you are already fine and do not need to spend anything. See our work: CallGuard AI, CallSetter AI, Fortell AI and more, shipped in days, not months.