Do you have to tell customers they're talking to an AI? What the law requires and how to get the introduction right

Same operation, two calls. In the first, the agent answers with a natural voice and a first name, and halfway through the conversation the customer starts to suspect: "sorry, are you a robot?". Evasive answer, awkward silence, customer hangs up and calls back demanding to speak "with someone real". In the second, the agent opens with one sentence: "I'm the company's virtual assistant; I can help you with your appointment, your invoice or the status of your case, and if you'd rather speak with a person I'll transfer you right now". The customer goes straight to what they need, the agent resolves it, and nobody spends a second working out who is speaking.

The difference between the two calls isn't the technology: it's a design decision that many operations still treat as a last-minute detail. And from 2 August 2026 it also stops being optional: the EU AI Act requires informing people that they are interacting with an AI.

This article explains what the rule actually demands, why hiding the agent was already a bad idea before it became illegal, and how to design the introduction —in voice and in email— so that transparency adds instead of subtracting. We're talking about agents that execute real operational work with customers, like the ones we compared in customer service bot vs AI agent, not laboratory experiments.

What the EU AI Act requires (and from when)

Regulation (EU) 2024/1689, the "AI Act", entered into force in August 2024 with a staggered application calendar. The part that directly affects anyone with an agent talking to customers is Article 50, the transparency obligations, which applies from 2 August 2026.

What it says, reduced to the operational core, is this: an AI system designed to interact directly with people must be built so that those people know they are interacting with an AI, unless it is obvious from the context to a reasonably well-informed and observant person. The obviousness exception is narrower than it looks: the fact that "that's obviously a bot" to you doesn't mean it is to your customer base, and the burden of having done it right falls on whoever deploys the system. For a voice agent that makes or answers calls, the prudent reading is simple: you say it, and you say it at the start.

Two more notes complete the picture. First, the same article requires marking synthetic content (generated audio, for example) in a machine-readable format, something the platform provider resolves in practice. Second, penalties for breaching these obligations reach 15 million euros or 3% of worldwide annual turnover; in Spain, the supervisory authority is AESIA. None of this replaces your legal counsel —review your specific case with them before the date—, but the underlying message needs no nuance: if an agent talks to your customers, your customers have the right to know.

It's worth remembering this doesn't come out of nowhere. The GDPR already required transparency about data processing —it's the question we answer in almost all our flow articles, from payment reminders to appointment confirmation calls—. The AI Act adds the missing piece: not just what is done with the data, but who (or what) is on the other side of the conversation.

Hiding it was already a bad idea before it was illegal

The temptation not to say it has an apparent logic: "if it sounds human, the customer cooperates more". The problem is that this bet only works while nobody suspects, and suspicion is a matter of time, not probability.

Customers detect patterns: the pause that always lasts the same, the too-perfect rephrasing, the question that ignores an obvious nuance. And the moment of detection is the worst possible place to manage the news: the customer no longer processes what the agent says, but the fact that "they were being deceived". A conversation that was going well becomes a complaint about the conversation itself. With today's voices, the paradox gets worse: the better the agent sounds, the more the detection resembles deliberate deception and the less it resembles a technical limitation.

There's a second, less visible cost: the team inherits the ambiguity. If the agent doesn't identify itself, the person receiving an escalated case has to decide on the fly whether to maintain the fiction or break it, and either option leaves them in a bad place. And there's a third, strategic one: trust in the channel. An operation that automates calls and emails needs the customer to accept the automated channel as legitimate; every awkward discovery erodes that acceptance for good.

Against all that, what you gain by hiding it is, being generous, marginal: a few seconds of "smoother" conversation at the start. It's a bad deal even without the fine.

The introduction is design, not a formality

Once we accept that it gets said, the useful question is how. And here lies the mistake symmetrical to hiding it: treating the notice as a legal disclaimer recited quickly to move on to something else. The agent's introduction is the channel's first impression; it deserves the same care as the flow that comes after.

In voice, the formula that works has four pieces in the first sentence: what it is, on whose behalf, what it can resolve, and how to reach a person. For example: "Hi, I'm the virtual assistant for Example Clinic. I can confirm, change or cancel your Thursday appointment. If you'd rather speak with a person, just say so and I'll transfer you". Twenty words after pickup, the customer knows who they're talking to, what the call is for, and that they're not trapped. That third piece —the visible exit— is what changes the customer's attitude the most: people tolerate talking to a machine just fine; what they don't tolerate is feeling held hostage by one.

In email, identification goes in the signature and the tone: the agent signs as what it is ("Company virtual assistant — automated reply supervised by the team"), not with an invented first and last name. The practical distinction lies in who sends: if the agent handles the conversation autonomously in a shared inbox, it identifies itself; if it only prepares a draft that a person reviews and sends —the "suggest" zone we described when designing escalation—, the communication belongs to that person and there's nothing to declare.

Just as important is what not to do. No human first names that invite confusion ("Hi, I'm Marta") if Marta doesn't exist. No theatre: office noises, rehearsed filler words or fake "I'm typing" pauses are exactly the kind of simulation the rule wants to prevent and that customers punish when they discover it. And no preventive apologies: an agent that opens with "I'm just a virtual assistant, sorry I'm not a person" is apologising for existing instead of resolving. Well-executed transparency is neutral and brief: you say it, and you prove usefulness in the next two sentences.

"Are you a robot?": the question you design, not improvise

Even with an impeccable introduction, the question will come: customers who didn't hear the greeting, who want to check, or who are simply curious. That moment must be designed with the same precision as any other step in the flow, because it's a live credibility exam.

The right answer has three properties. It's immediate and direct: "Yes, I'm the company's virtual assistant". Any evasion —"I'm your assistant", "that doesn't matter right now"— turns a neutral question into confirmation of deceit. It offers a real choice: "Would you like me to continue with this, or would you prefer I transfer you to a person?". And it respects the choice the first time: if the customer asks for a person, the case escalates with all the captured context, without a second retention attempt. It's exactly the first trigger on the list we proposed when designing escalation to people: the customer who asks for a person gets one, always.

There's an extra use almost nobody makes of it: logging the question as a metric. If "are you a robot?" appears in a relevant percentage of conversations, you don't have a problem of distrustful customers; you have an opening introduction that isn't landing or isn't clear. It's a free tuning signal.

What changes in the results (and what doesn't)

The fear holding many operations back is measurable: "if we say it's an AI, people will hang up". Operational experience points the other way: what determines whether the customer stays on the call is not who is speaking, but whether the call is useful to them. A declared assistant that confirms the appointment in forty seconds retains more than an ambiguous voice that forces the customer to play detective. In transactional flows —appointment confirmations, case status, after-hours calls— customers value immediate resolution far above the nature of the speaker.

To get past opinions, four numbers watch over the transition:

• Early hang-ups: abandonment in the first seconds after the introduction. If they rise when the notice is introduced, the opening script needs work (too long, too legalistic); you adjust the wording, not the honesty.
• Person-request rate: what percentage asks for a human after learning they're talking to an agent. High at first and declining is normal; high and stable signals a poorly chosen scope: the agent doesn't resolve what people call to resolve.
• Resolution rate without intervention: the usual metric. If it holds with the transparent introduction, the debate is closed.
• Mid-conversation "are you a robot?" questions: they should trend towards zero with a good opening; if they persist, the notice isn't being understood.

The combined reading matters more than each figure: an operation that introduces its agent well, resolves fast and escalates without friction doesn't lose customers to transparency. It loses them —as always— by not resolving.

How BeeAgent fits in

In BeeAgent, transparency isn't a patch on the script: it's part of each agent's configuration. The introduction is defined per flow and per channel —what the agent says on pickup, how it signs the emails it sends on its own— and the answer to "are you a robot?" comes designed by default: direct confirmation and the offer of a person, connected to the escalation rules so that the request for a human is respected the first time, with all the context.

In addition, every interaction is logged —what was said, when and with which introduction—, which is exactly the kind of traceability that makes it easier to demonstrate compliance if anyone asks. And since it's a no-code platform, adjusting the opening after seeing the first week's metrics —shortening the greeting, reinforcing the offer of a person, changing the signature— is done by the operations team in minutes, with no development project.

Get your operation ready before 2 August

The adaptation fits in a calm week of work, not a project.

First, inventory the contact points: every flow where an AI converses with customers —inbound calls, outbound calls, autonomous emails— and what it says today in the first seconds. Second, rewrite the openings with the four-piece formula: what it is, on whose behalf, what it resolves, how to reach a person. Third, design the answer to the robot question and connect it to escalation with context. Fourth, align the team: whoever receives escalated cases must know the agent identifies itself, so nobody improvises. And fifth, run the result past your legal counsel —the fine-grained interpretation of Article 50 in your sector is their job— and measure the baseline of hang-ups and person requests to compare after the change.

Whoever has this done in July won't notice 2 August. Which is exactly the idea.

Conclusion

The question in the title has a short answer: yes, you have to say it, and from 2 August 2026 the EU AI Act requires it with serious fines behind it. But the long answer is more interesting: saying it well is not a toll, it's a design advantage. The agent that introduces itself clearly, proves useful within two sentences and offers a way out to a person builds the trust the automated channel needs to absorb real volume; the one that hides destroys it at the worst possible moment and in front of the wrong customer.

If you're preparing your flows for the new rule, in the use cases you can see how real agents introduce themselves in customer service and outbound calls, or write to us and we'll review your flows' openings together before the deadline.