AI receptionist disclosure: should it say it's an AI?

What callers do when they're told it's an AI

The fear is simple. Say "AI" and they hang up. There's some truth in it, but it comes from the wrong kind of call.

The study everyone quotes is a 2019 field experiment where a lender's bot rang customers to sell loan renewals. When it announced itself first, more than half of the people who answered hung up within five seconds. I went through it in the voice post. Those were sales calls nobody asked for. Your caller rang you, with water on the floor.

Closer to your situation, researchers at the University of Göttingen ran two online experiments in which about 200 people each read through a customer-service chat with an energy company. The bot introduced itself as "Leon". At the end, half were told Leon was a chatbot. When the request was high-stakes and the bot handled it, being told cost a little trust: 6.10 against 6.56 on a seven-point scale. When the bot failed, being told raised trust from 3.55 to 4.06, and with it their intention to stay a customer.

Read that as a tradesman. Your AI will get something wrong. It will mishear a street name or a letter in a postcode. A caller who knows it's a machine blames the machine and waits for your callback. A caller who thought "Leon" was a person, and finds out later, has two complaints instead of one.

People also want to be told. In a 2024 Salesforce survey of 15,015 people in 18 countries, including the US and the UK, nearly 75% said they want to know if they're talking to an AI agent, and 45% said a clear escalation path makes them more likely to use one. For a trade firm, the escalation path is you, ringing back.

Here's what AI can't do. It can't be sure the caller heard the disclosure. An 80-year-old on a crackly landline may miss "AI assistant" entirely, and the AI carries on as if they heard it. So the words go first and plain, not tucked into a long sentence, and a human calls back fast.

The four honest moments on every call

Disclosure isn't one line in the greeting. On a real call, honesty gets tested four times. I call them the four honest moments, and I script every one.

1. The greeting. Who they've reached, that it's the AI assistant, what happens next: the who-what-next greeting, under eight seconds, set as a fixed message so it never changes. If you record calls, the recording notice goes here too; the UK and US versions are in the UK plumbers post.
2. The question. "Am I talking to a real person?" The answer is one sentence of truth, then back to the job: "No, I'm the AI assistant for Dave's Plumbing. I'll get this to Dave now. What's the address?" No joke, no "I'm here to help", no change of subject. Utah requires a straight answer to a clear question, and its safe harbor asks for disclosure at the outset and throughout. I read that as: never contradict the greeting. Check that reading with your advisor.
3. The limit. The caller asks for something only a person can give. A price, a time, "can you come now?" This is where AIs drift into promises. The script says what it can't do and who can: "I can't give times, I'm the AI assistant. Dave will ring you back with that." The CMA puts it as don't overstate what the AI can or cannot do.
4. The callback. You ring back and own it: "Hi, it's Dave. You spoke to my AI assistant a few minutes ago, and I've got your notes." Don't pretend the AI was a colleague called Sarah. The caller hears that nothing was hidden and that a real person read what they said.

One more for US firms: keep the AI on the answering side. The FCC ruled in 2024 that AI-generated voices count as an "artificial" voice under the TCPA, so a call the AI makes needs the called person's prior express consent, absent an emergency or an exemption. Mine never rings out. The callback is a person, in their own voice. Texts and consent are covered in the TCPA post.

What disclosure costs, worked out

Owners worry that saying "AI" costs jobs. Run it both ways.

Say 60 calls a month reach your AI: after hours, plus the ones you can't grab on a job. "I'm the AI assistant" takes about a second and a half to say. That's 90 seconds of airtime a month.

Now the other way. Say one caller in ten asks whether it's a real person, and the AI dodges. That's six dodges a month. In Utah, the consumer protection division can fine up to $2,500 per violation. If each dodge counted as one, that's up to $15,000 a month on paper.

I don't expect a regulator to start with a plumber's phone line. The point is the ratio: ninety seconds of airtime against a five-figure risk, plus six callers who now doubt every word the AI told them. In the UK, the CMA's guidance reminds businesses that consumer law breaches can bring fines of up to 10% of worldwide turnover.

The bigger cost was never the fine. It's the job. For the other half of the sum, what a missed call is worth to you, the calculator runs it with your own numbers.

Do this this week

1. Ring your own line from a phone it doesn't know. Count the seconds until it says "AI". If it never says it, that's your first fix.
2. Ask whether it's a real person three ways: "Is this a robot?", "Am I talking to a human?", "Are you real?" Every answer should say plainly that it's the AI assistant, then go back to the job.
3. Ask "can someone be here by four?" It should say it can't promise times and that a person will ring back. If it says yes, it just made a promise in your name.
4. If it has a human name, swap it for a job title.
5. Write your callback opener on a card in the van: "You spoke to my AI assistant a few minutes ago."
6. Ask your vendor one question: does it ever make outbound calls? If yes, ask how consent is handled before you sign anything.

If you'd like a second pair of ears on your script, book a growth call. Thirty minutes, free. If your phone rings twice a day and you answer both, I'll tell you to keep your money. What I build for trade firms is on the AI phone receptionist page.