Skip to content
Implementa.
AI AgentsOpinion··10 min

Telling customers they're talking to an AI: how to say it without them hanging up

Telling customers they're talking to an AI has been mandatory in the EU since August 2, 2026, and the data is scary: in one field experiment, revealing the bot before the conversation cut purchases by more than 79.7%. The thesis: the disclosure isn't what scares people off. The bare disclosure is. Say what it is, what it can do for the customer and how to reach a person.

Senior AI Operations Implementer

AI Operations Pod

Illustrative scenario. A dental clinic sets up a voice agent to book appointments after hours. Someone in the meeting asks whether it has to say it's an AI, and the room gives the usual answer: “better if nobody can tell, people hang up.” They give it a human name, a warm voice and a greeting that mentions nothing. It works for three weeks. Then a patient asks to be put through to “the girl from before” and finds out, all at once, that there was no girl and that nobody had told her. She isn't angry about the AI. She's angry about the trick.

The thesis in one sentence: disclosure doesn't kill conversion; the bare disclosure does, the one that says “I'm a bot” and nothing else. If, along with what it is, you say what it can do for the customer right now and how to reach a person, the disclosure stops being an apology and becomes a service promise. A note before we go on: this describes the EU AI Act as it applies today and isn't legal advice; for a specific case, talk to your lawyer. Outside the EU the rules differ by country, and in the US by state.

Do you have to tell customers they're talking to an AI?

In the European Union, yes. Article 50(1) of Regulation (EU) 2024/1689 requires AI systems that interact directly with people to be designed so those people know they're dealing with an AI. It applies from August 2, 2026 and, according to the European Commission's FAQ on Article 50, the notice has to arrive from the start of the first interaction, in a clear, distinguishable and accessible way. Fines can reach €15 million or 3% of worldwide turnover, with proportionality for SMEs. If you sell to EU customers from the US, this is your problem too.

Formally it's an obligation of the system's provider, not of whoever uses it. But careful: if your company built its own agent on top of a model's API, the provider of that system is very likely you. The line is drawn in provider or deployer under the AI Act, and the full calendar of what was postponed and what wasn't is in what the AI Act postponed in August 2026 and why your chatbot has to disclose. We won't repeat it here. We'll go to what nobody covers: how to actually say it.

There's an exception: you don't need to disclose if it's obvious to a reasonably well-informed, observant person that it's an AI. The Commission wants it read narrowly, and its May 2026 draft guidelines, per Covington's summary, give as an example where the exception probably doesn't apply exactly the one most companies have: a chatbot embedded in a website's help desk. If your audience is the end customer, assume it isn't obvious.

The scary data: disclosing before the conversation sank purchases

The fear of disclosing isn't superstition. There's a study behind it. In 2019, Luo, Tong, Fang and Qu published in Marketing Science a field experiment with more than 6,200 customers of a financial services company, randomly assigned to sales calls made by a bot or by a person. Undisclosed, the bot sold as well as an experienced rep and about four times better than a rookie. When customers were told before the conversation that they were talking to a machine, the purchase rate fell by more than 79.7% and calls got shorter.

The authors' explanation is the interesting part: the bot wasn't worse for being known as a bot. Customers perceived it as less competent and less empathetic, turned curt and bought less. It was a bias against the machine, not an assessment of the service. And the study found two things that softened the blow: customers who already had AI experience, and revealing the bot late, after the conversation.

In Europe, that second way out is gone. If the notice has to come from the first interaction, the “tell them at the end” play is off the table for any system serving people in the EU. So the useful question is no longer whether to disclose, or when. It's what else you say in the same moment you disclose.

Disclosure isn't what scares people off: the bare disclosure is

Enter the second study, less cited and more useful. Mozafari, Weiger and Hammerschmidt presented at the HICSS conference in 2021 an experiment with 257 participants in a simulated insurance chat. Disclosure on its own lowered trust compared with no disclosure. But when the disclosure came with a sentence about what the bot could do, or even about what it couldn't, trust came back to a level not significantly different from no disclosure at all.

Read it with the fine print: it's a lab scenario, not a live chat, and the differences are statistically modest. It won't tell you how much you'll convert. It tells you something simpler and sturdier: the penalty comes from knowing only that it's a machine. If the next sentence gives the customer a reason to stay, the penalty shrinks.

The third data point is from this year and points the same way. In a Five9 survey of 3,000 consumers released in June 2026, reported by CX Dive, 55% trust AI customer service when there's a clear path to a person, versus roughly a quarter when there isn't. And 41% say they're less likely to use a company that serves customers with AI, rising to 53% if there's no way to reach a human. The exit door isn't a design detail. It's half the trust.

The three parts of a disclosure that doesn't scare people off

Put the three data points together and you get a short formula. A disclosure that works says three things in the first message, in this order:

  1. What it is, no euphemisms. “I'm the clinic's AI assistant.” Not “Laura, your virtual assistant” with a photo of a person. A human name isn't banned, but a human name plus a human face plus zero mention of AI is exactly what disclosure exists to prevent.
  2. What it can do for you right now. “I can book, move your appointment or tell you what your insurance covers.” Concrete and short. This is the sentence that, in the 2021 experiment, brings trust back. If you also say what it doesn't do (“I don't give diagnoses”), even better: you set expectations before you miss them.
  3. How to reach a person, and when. “If you'd rather talk to the team, type PERSON and we'll call you before noon.” A real door, with a deadline, not a “someone may get back to you.”

Notice what's missing: no apology (“unfortunately I'm a bot”), no technical justification (“I use an advanced language model”) and no three paragraphs of legalese. The legal notice fits in five words. The rest of the message is service.

Channel by channel: where the disclosure goes and what usually breaks

ChannelWhere the disclosure goesTypical mistake
Website chatIn the chat's first message, visible before the customer typesHiding it in the legal notice or behind a “learn more” link
WhatsAppIn the first message of every new conversationDisclosing once in January and assuming the customer remembers in June
PhoneIn the first sentence, before asking for any dataDisclosing only if the customer asks “are you a machine?”
Email sent by an agent with no reviewIn the message itself, next to the signatureSigning with the name of a person who never read it
Handoff to a personAt the moment of the switch, both waysThe customer can't tell whether a person has taken over or the bot is still talking

Two nuances that aren't in the law but do show up in operations. First, email: if an agent drafts and sends replies to customers with nobody looking at them, it's interacting directly with those people, and the prudent move is to treat it like a chat. If a person reviews and sends under their own name, the AI is a drafting tool; that last part is our reading, not a line from the regulation. Second, the handoff: the disclosure doesn't end with the first message. The customer also needs to know when they stop talking to the AI, and when they go back to it.

What doesn't count as disclosing

There are three shortcuts we see often, and none survives a serious reading. The first is putting the line in the terms of use: the Commission's draft guidelines say explicitly that a notice buried in terms, manuals or multi-layer menus doesn't meet the “clear and distinguishable” standard. The second is disclosing only when asked: the obligation is to inform, not to confess once you're caught. The third is the ambiguous label. “Virtual assistant” has spent twenty years also naming people who work remotely; if you want zero debate, the word is “AI.”

And a fourth that's technical and almost nobody checks: letting the model improvise the greeting. If the disclosure lives inside the agent's instructions as just another request, one day the model paraphrases it, the next it shortens it, and the day after it skips it because the customer opened with a direct question. The disclosure has to be fixed text in the flow, always sent, not a suggestion in the prompt. How to separate what the model decides from what's set in stone is in how to write an AI agent's instructions.

If disclosure is killing your conversion, the disclosure isn't the problem

When a company tells us people drop off since they started disclosing, the culprit is almost never the sentence. It's what happens right after. A customer who knows they're talking to an AI is less patient with it: they give it one shot, maybe two. If in those two tries the agent neither solves nor hands off, they leave. Before the disclosure, that same bad agent could hide behind the doubt. Now it has nowhere to hide.

That's why the real lever is the design of the exit: what the agent does when it doesn't know, who it hands the conversation to and with what context. It's laid out in what an agent does when it doesn't know the answer and, for deciding which decisions always need a person, in human in the loop in AI automation. It's the same conclusion we reached in why your AI chatbot makes customers angry: it doesn't anger people for being AI, it angers them for not knowing how to hand off.

And if you want to know whether your disclosure works, don't debate it in a meeting: measure it. Three numbers are enough. How many conversations die at the first message, before and after you change the text. How many get resolved without a person. And how long it takes a person to step in when the customer asks. If the first goes up and the other two look fine, try different wording. If the other two look bad, no wording will save you.

What to take away

  • In the EU, telling customers they're talking to an AI has been mandatory since August 2, 2026, from the first interaction and clearly; the “it's obvious” exception is read narrowly.
  • Revealing the bot before the conversation can sink purchases: in Luo et al. (2019), by more than 79.7%. Telling them at the end softened the blow, but in Europe that option is gone.
  • The penalty comes from the bare disclosure. Paired with what the agent can do, trust recovers; with a clear door to a person, it roughly doubles.
  • Formula: what it is, what it can do for you now and how to reach a person. Fixed text in the flow, not model improvisation.
  • If disclosure tanks conversion, fix the agent's exit before the sentence.

If your support agent is already in production and you don't know what its first message says, start there: it's the most-read sentence in the whole system. Shipping an AI support agent that discloses, solves and hands off with context is what we do, and we close the inventory of everywhere else Article 50 touches you in AI Act compliance by operating AI. What we don't do is hide the machine. There's no need.

Shall we get it shipping?

If this resonated, 30-minute conversation with no commitment. We tell you what fits, what doesn't and the approximate price.

See cases
Telling customers they're talking to an AI: how to say it without them hanging up · Implementa