Telling Customers When AI Is Involved

Why this matters

The moment a customer realizes, halfway through a call, that the friendly voice booking their furnace repair is not a person, the shop loses something it does not get back on that call. They start testing it, repeating themselves, asking for a manager. The same customer told up front, in one plain sentence, usually just carries on. Disclosure is cheaper at the start than at the discovery.

It is also, in some places and some channels, the law. This card covers where a disclosure is required, where it is simply good practice, and the exact lines that work at each point a customer meets an automated tool: the phone greeting, the "am I talking to a real person?" moment, texts, website chat, AI-drafted emails and outbound calls. The broader privacy notice is a separate document. See related: The Customer Privacy Promise When You Use AI.

Step 1: Know which rules can reach you

There is no single federal AI disclosure law as of 2026. The layers that exist:

• Federal floor, deception. The FTC Act's ban on deceptive practices (Section 5) applies to every business. Presenting a bot as a human, or letting a customer believe they are talking to a person when the shop knows they are not and the belief matters to the sale, is the kind of thing it reaches. No statute names "AI" here; the general rule does the work.
• Federal, outbound calls only. In 2024 the FCC ruled that AI-generated voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act. An automated outbound call using one generally needs the called party's prior express consent, and prior express written consent where the call is telemarketing. This reaches reminder calls and sales calls the shop places, not calls customers place to you.
• California. The bot disclosure law (Business and Professions Code 17940 and following, in effect since 2019) makes it unlawful to use a bot to communicate with a person in California online with intent to mislead them about its artificial identity in order to encourage a sale. A clear, conspicuous disclosure is the safe harbor. Its text is written for online interactions (websites, apps, social media); whether it reaches a voice agent on a phone line is not settled, so treat the phone the same way rather than relying on the gap.
• Utah. The Artificial Intelligence Policy Act (2024, amended in 2025) requires a business using generative AI with a consumer to disclose that it is AI when the consumer clearly asks. People in occupations the state licenses through its Department of Commerce, which in Utah includes many contractors, carry a stronger duty to disclose up front in interactions the amended law treats as high-risk, such as collecting sensitive information or giving advice someone could rely on for a significant decision. The law carries a sunset date that has already been extended once.
• Colorado and others. Colorado's AI Act includes a duty to tell consumers they are interacting with an AI system unless that is obvious, with an effective date that was pushed into mid-2026; check whether it has taken effect and in what form. Other states have bills moving every session.

This list is current as of 2026 and it will be wrong in pieces within a year. Re-check it each January with your state attorney general's consumer protection pages or a trade association's legislative tracker, and check the state of every customer you serve, not just the state where the shop sits. This is not legal advice; the point where you need a lawyer is named in step 6.

Step 2: Write the phone greeting so the disclosure is in the first sentence

The simplest policy that satisfies every rule above, and the most honest one, is to disclose at the start of every automated interaction in every state. One sentence, before anything else, in plain words, with the way out offered immediately.

AGENT:    Thanks for calling Ridgeline Heating and
          Air. I'm the automated assistant. I can
          book a visit or take a message, or I can
          get you to a person - just say "person"
          any time.
CUSTOMER: Yeah, my furnace is blowing cold air.
AGENT:    I can help set that up. What's the
          address for the visit?
...
CUSTOMER: Wait, am I talking to a real person?
AGENT:    No, I'm the automated assistant. Want me
          to connect you to someone in the office,
          or keep going here?
CUSTOMER: Just keep going.

What makes it work: "automated assistant" rather than a human name, so nobody is misled by a persona; the exit ("say person") in the first breath, so the customer knows they are not trapped; and a straight answer to the direct question, every time, with the same offer repeated. Never script the agent to dodge "are you a real person?" or to answer it with a joke. Under Utah's rule that dodge is the violation, and everywhere else it is the moment trust breaks.

What breaks if you skip it: customers who discover the agent mid-call, and complaints that read "they tried to pass a robot off as a person", which is a worse review than any you would have earned by saying so.

Step 3: Make the handoff real

A disclosure promising a person is only honest if a person is reachable. Decide what "person" does at each hour:

• Business hours: transfer to the office, with a target pickup. A common starting point is within 60 seconds; if nobody picks up, the agent says so and takes a callback number.
• After hours: the agent says plainly that the office is closed, gives the callback time ("by 8 tomorrow morning"), and for anything that sounds like an emergency, follows the emergency script rather than taking a message. A caller describing a gas smell is told to leave the building and call the gas utility's emergency line or 911 from outside, not offered a callback.

What breaks if you skip it: a disclosure that turns into a broken promise. "Say person any time" followed by a transfer that rings out is worse than no offer at all.

Step 4: Disclose in text, chat and email, scaled to how much the tool did

The rule of thumb that holds across channels: disclose in proportion to how much the customer is dealing with a machine rather than with you.

• Website chat run by an assistant: a label at the top of the window ("Automated assistant - a person can take over") and a first message that repeats it. This is the channel California's law was written for, so the disclosure must be visible before the conversation starts.
• Automated texts (confirmations, after-hours replies): sign them as automated ("Automated message from Ridgeline - reply PERSON to reach the office"). A customer who replies to a text expects a person to read it, and is entitled to know when nobody will until morning.
• Emails a person drafted with AI help and then reviewed and sent: no disclosure is needed, any more than you would disclose a spellchecker. The person who sent it is the author and is responsible for every word. What matters there is the review. See related: The AI-Drafted Customer Message Review SOP.
• Emails or messages sent with no human review: treat them like automated texts. Sign them as automated.

Step 5: Handle outbound calls separately

If the shop uses an AI voice for outbound calls (appointment reminders, maintenance-due calls, follow-ups), the TCPA consent rule in step 1 applies before the disclosure question even arises. Confirm you hold the right consent for that number and that type of call. Then open the call the same way as the greeting: who is calling, that it is automated, and how to reach a person or opt out. A reminder call that sounds human and is not, placed without consent, stacks two problems on one call.

Step 6: Know the fork where you stop and ask a lawyer

Run the steps above yourself. Stop and talk to an attorney before acting if: you serve customers in Utah and your agent does anything beyond booking (answering whether something is safe, discussing financing, collecting health or financial details); you want to give the agent a human name or persona; you plan outbound AI voice calls for marketing; or a customer or regulator complains that they were misled. Walk in with the agent's exact greeting script, a transcript of a typical call, the list of states your customers are in, and how a caller reaches a person at each hour.

Worked example: a shop measures what disclosure costs it

A five-truck HVAC shop in a metro area that spans two states adds an AI phone agent for overflow and after-hours calls. The owner worries that saying "automated" up front will drive callers away, so he runs the first month without the disclosure sentence (the agent simply says the company name and asks how it can help) and the second month with the greeting from step 2.

His numbers, from the agent's call log:

Read it carefully, with each rate on its own month's calls. Early hang-ups went from 14 of 240 (about 5.8 percent) to 17 of 252 (about 6.7 percent), roughly one more point. Requests for a person rose from 22 of 240 (about 9.2 percent) to 29 of 252 (about 11.5 percent), because the greeting told people they could. Bookings went from 118 of 240 (about 49 percent) to 127 of 252 (about 50 percent), essentially flat. And the "is this a real person?" interruptions, each one a stalled call, fell from 31 to 4, while complaints about the agent went from 3 to none.

What it tells him. The disclosure cost about one point of early hang-ups and a few more transfers, and it bought shorter, calmer calls and an end to complaints about being misled. On these numbers he keeps it, and staffs the transfer line for the extra requests. The months are not a controlled experiment: month two was later in the heating season, which could move any of these figures, so treat the comparison as directional. But nothing in it suggests disclosure hurts bookings, and it removes a legal exposure in at least one of the two states his customers live in.

What would change the answer. Nothing in these numbers would justify dropping the disclosure, but they could justify changing its wording. If early hang-ups had jumped by several points rather than one, the right move would be to test a shorter greeting ("You've reached Ridgeline's automated assistant - say person for the office"), not to remove the sentence.

References

• FTC Act Section 5, 15 U.S.C. 45, deceptive practices
• FCC Declaratory Ruling (February 2024) applying the TCPA, 47 U.S.C. 227, to AI-generated voices
• California Business and Professions Code 17940 to 17943 (bot disclosure); Utah Artificial Intelligence Policy Act (2024, amended 2025); Colorado AI Act (SB 24-205), as currently in effect
• See related: Setting Up an AI Phone Answering Service and Testing It
• See related: The Customer Privacy Promise When You Use AI