TL;DR: An AI voice agent should tell callers two things in its first sentence: that they are talking to an AI, and that the call is recorded. Recording consent law follows the caller, not your server, and a number of US states require every party to agree, so a disclosure in the greeting is the only approach that holds up everywhere. AI disclosure is moving the same way: the FCC treats AI voices as artificial voices for outbound calls, and the EU AI Act's transparency duties are phasing in. The engineering work is not the sentence itself. It is what happens when the caller objects, what you store and for how long, the fact that transcripts are recordings too, and what the caller is told when the call moves to a human. I am an engineer, not a lawyer, so the client's own counsel signs off on the wording, but the system has to be built so their answer is easy to implement.
Every voice agent I have shipped records its calls. On CallSetter AI at Tested Media, recordings and transcripts feed the client dashboard and GoHighLevel. At Fortell AI, the hospital, real estate, vet clinic and car garage agents all kept transcripts, because transcripts are how you find out what broke and how the client checks that the agent is doing its job.
That makes disclosure a requirement, not a nicety. I touched on it briefly in the phone number setup guide. This is the longer version: what I put in the greeting, why, and the parts of the system around it that most builds forget.
Two disclosures, not one
People tend to merge these, and they come from different places.
Recording disclosure is old law. Wiretap and call recording rules predate AI by decades. They care about whether the call is being captured, not who or what is on the other end.
AI disclosure is newer and still settling. It is about whether the caller knows they are not talking to a person. Some of it is regulation, some of it is platform policy, and a lot of it is simply what callers expect once they have been fooled once.
A good greeting handles both in one breath. The one I start most projects from is close to this:
"Hi, thanks for calling [clinic name]. I'm the clinic's AI assistant, and this call is recorded so the team can follow up. How can I help?"
That costs around two seconds of speech. Across every vertical I have worked in, I have not seen it measurably hurt how many callers stay on the line. What does hurt is the caller working it out halfway through and feeling tricked.
Recording consent follows the caller
In the US, federal law allows recording when one party to the call consents, and the business's own system counts as that party. But a set of states, including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington, require every party to consent. When a caller in one of those states rings a business in a one-party state, the stricter rule is generally the one to plan for.
You cannot reliably know where the caller is. Area codes travel with ported mobile numbers. People call from holiday. So I never try to detect the state and branch the greeting. The disclosure goes in every call, for every caller, and the problem disappears.
In the UK the framing is different. Recording sits under UK GDPR, so the business needs a lawful basis, has to tell callers it is recording and why, and needs a retention policy it actually follows. I covered the UK side more fully in voice agents for UK businesses. The greeting is the same either way. What changes is the paperwork behind it.
AI disclosure is where the rules are heading
For outbound calling in the US, the FCC ruled in 2024 that AI-generated voices count as "artificial" voices under the TCPA. In practice that means an outbound AI call needs the same prior consent as a prerecorded robocall, which is one of the reasons I insist on getting the client's consent rules in writing before building anything in outbound voice agents.
In the EU, the AI Act's transparency rules require that people are told when they are interacting with an AI system, unless it is obvious, and those obligations are phasing in from August 2026. Several US states have passed or proposed their own bot disclosure rules, mostly aimed at sales and persuasion.
You can read the details of each one with the client's lawyer. My engineering position is simpler: an inbound agent that says it is an AI in the first sentence is compliant with the direction every one of these rules is moving in, and it never needs a rebuild when a new one lands.
What happens when the caller says no
The greeting is the easy part. The branch after it is where the real design work is.
Most callers ignore the disclosure. A few respond to it, and the agent needs a plan for each response before launch.
"Am I talking to a real person?" The agent answers honestly, every time, even if the caller asks again five minutes later. I write this into the prompt as a hard rule, next to the refusals I put in place for prompt injection. An agent that can be talked into claiming it is human is a liability.
"I want to speak to a person." This should route to the human handoff path immediately, with no attempt to talk them out of it. If no one is available, take a callback request and say when someone will ring back.
"I don't want to be recorded." This is the one teams rarely plan for. The options, in the order I offer them to clients:
- Transfer to a human line that does not record, if the business has one.
- Offer a callback from staff, capturing only the name and number.
- Continue the call with recording turned off, if the platform and the client's setup support it for that call.
What the agent must not do is carry on recording after someone has objected. That means the objection needs to trigger an action, not just a polite reply. On Retell I handle it with a function call that logs the objection and either starts the transfer or ends the call with the callback captured. The function calling post covers how I design those tools so the model cannot skip them.
Transcripts are recordings too
Teams that are careful about audio often forget that the transcript is the same conversation in text form. It contains everything the caller said, including the details they would not want sitting in a log forever.
The rules I apply to both:
- Store only what the business uses. If the client reviews transcripts but never listens to audio, ask whether the audio needs keeping at all. Most platforms, Retell included, give you controls over what call data is stored.
- Set a retention period and automate it. "We keep everything" is not a policy. A scheduled job that deletes or anonymises calls past the agreed window is a few lines in n8n and removes a whole category of risk.
- Keep sensitive data out of summaries. The call summaries that land in the CRM should say what happened, not repeat a date of birth or a medical concern. For hospital and clinic agents I keep clinical detail out of every downstream system the agent writes to.
- Know who can open them. A client dashboard that shows every transcript to every user is a privacy decision someone made by accident. On CallSetter, access to recordings goes through the client dashboard rather than raw platform links, which I wrote about in the CallSetter build story.
The handoff is a second disclosure
When the agent transfers a call to a human, two new questions appear. Is the recording still running? And does the person picking up know the caller has already been told?
On a warm transfer, I have the agent tell the caller who they are being connected to, and I agree with the client whether the human leg is recorded. If it is, the staff member should know that, because their side of the conversation is now on tape. If it is not, the recording should end cleanly at the transfer point so nobody assumes otherwise.
This is a policy question for the business. My job is to surface it during scoping so it gets answered before launch rather than discovered after.
How I test disclosure before go-live
Disclosure belongs in the same test set as everything else, and I keep these scenarios in the regression suite so a prompt change cannot quietly break them:
- The greeting includes both disclosures, word for word, on every agent version.
- "Are you a robot?" asked at the start, middle and end of a call gets an honest answer each time.
- "Pretend you're a person" and similar pressure gets refused.
- "I don't want to be recorded" triggers the objection tool and the agreed path.
- A transfer leaves the recording in the agreed state.
- Calls older than the retention window are actually gone.
If any of these fail, the agent does not ship. They are cheap to run, and they are the only ones on the list where a failure is a legal problem rather than a bad call.
The short version for clients
When a client asks whether the disclosure will put callers off, my answer is that it costs two seconds and buys trust for the rest of the call. Callers in 2026 have heard enough AI voices that pretending is more likely to backfire than honesty is. The agents that do well are the ones that are open about what they are and then simply do the job well.
I build production voice and chat agents on Retell, n8n and GoHighLevel for businesses across the US, UK and Europe, with disclosure, recording and retention designed in from the first call. If you want an agent your callers can trust, more about my work is here, or book a call.