Most companies that put an AI assistant in front of customers ask the question too late: do we have to tell them it isn't a person?
In the European Union the answer is yes, and it is written down. Article 50 of the EU AI Act sets transparency obligations for systems that interact directly with people. It is not a best-practice suggestion — it carries a penalty regime.
This is what it requires, who it applies to, and the part everyone actually asks about: how to comply without making the conversation feel like a consent form.
What it says, in plain English
The core obligation is simple: the person has to know they are interacting with an AI system, unless that is obvious from the context to a reasonably observant person.
Three things that matter in practice:
Before, not after. The notice has to arrive at first contact. A disclosure at the end of the conversation, or buried in a privacy policy, does not do the job — the person has already talked believing something else.
Clear, not buried. "This service may use automated technologies" informs nobody. The test is whether a customer reading that line would understand they are writing to a machine.
"Obvious from context" is narrower than it sounds. It is tempting to think that calling it "Virtual Assistant" settles it. But modern assistants write well and naturally — which is exactly why people get confused, and exactly why the article exists.
Who it applies to
If your company operates in the EU or serves people in the EU, it applies. It does not matter where the system is hosted or whose model it uses: what matters is where the customer is.
It also applies when a vendor built the assistant for you. The obligation is yours, not theirs. If your provider gives you no way to comply, the problem is still yours.
And a note for UK readers: the UK is not covered by the EU AI Act, but if you serve customers in Spain, Ireland or anywhere else in the bloc, you are.
How to comply without ruining the conversation
Here is the real fear: that saying "I'm an AI" scares customers off. In our experience the opposite happens — with one condition, that the assistant is good.
What works:
- A persistent, visible notice before the first message. On web chat, a line above the input. On WhatsApp, in the first message.
- The assistant admits it when asked. "Are you a real person?" has exactly one correct answer. We enforce that in code rather than in the prompt, so it cannot be talked out of it.
- A clear route to a human. Knowing you are talking to a machine bothers people far less when they can see there are humans behind it.
What does not work: ambiguous names, light-grey 10px disclaimers, or an assistant that dodges the question. That last one damages trust far more than being an AI ever did.
What Article 50 does not require
Worth bounding, because there is a lot of alarmism about:
- It does not require you to explain how the model works.
- It does not require consent to use AI — this is disclosure, not authorisation.
- It does not make your chatbot a "high-risk" system. A sales assistant that answers enquiries and books appointments is not in that category; the high-risk obligations are different and apply elsewhere.
And the part no law covers
Article 50 is the floor, not the ceiling. There is a second half that appears in no regulation and decides whether the customer comes back: the assistant must not make things up.
An assistant that declares itself an AI and then invents a price, a lead time or availability does more damage than one that passes as human and gets it right. Transparency without reliability is transparency about a problem.
Which is why, in our case, the guarantees that matter are not instructions to the model but rules in the code: it cannot quote a price that did not come from the real system, promise stock it has not checked, or deny a product that exists. If the model tries, the reply is corrected before it goes out.
Five minutes to check your own
- Open your chat as a customer. Is the notice visible before you type?
- Ask whether it is a person. What does it say?
- Ask for a human. Does it know how?
- Ask it a price. Where did that number come from?
The fourth is usually the one that hurts.
Ours ships this way by default: the notice appears before the first message, the assistant always admits what it is, and a human is one message away. See how Zatio works, or our security page for where conversations are hosted and who processes them.