The EU AI Act’s Chatbot Disclosure Rule, for Support Teams: What Article 50 Asks of Your AI Agent
September 15, 2026 · 9 min read · by the hiy team
The short answer
Article 50 of the EU AI Act applies from 2 August 2026. For a support team the relevant sentence is short: an AI system designed to interact directly with people has to be built so that the people concerned are informed they are interacting with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person — and the Commission's own guidance says the information has to come "from the start of the first interaction in a clear and distinguishable manner". The duty to design the disclosure falls on the provider of the system; the practical exposure for a support team using someone else's widget is not hiding, removing or contradicting it. The easy way through is a chat surface that says it is an AI in its own header, says so again when asked, keeps saying so on every plan, and names a person as a person when one takes over. This post explains what the rule asks in plain language; it is not legal advice, and the Commission's FAQ is the source to read before your lawyer does.
What does Article 50 require of a chatbot?
Four things, read from the Commission's FAQ on the article:
- The date. "Article 50 of the AI Act applies as from 2 August 2026." (A limited grace period, to 2 December 2026, covers marking obligations for systems placed on the market before that date — that is about labelling generated content, not about telling a customer they are talking to a machine.)
- The duty. Systems designed for direct interaction with people must be developed so that "individuals concerned are informed that they are interacting with an AI system, unless this is obvious."
- The exception. "Obvious" is judged by an average person who is "reasonably well-informed, circumspect, and observant". A chat bubble labelled with a product's name and a person's avatar is not obviously a machine to that person. A box that says AI assistant in its header is.
- The manner. The disclosure has to come "from the start of the first interaction in a clear and distinguishable manner and in accordance with accessibility requirements". Not after the third message, and not in a tooltip.
The penalty regime for breaching Article 50 reaches "up to 15 million euros or 3% of total worldwide turnover for the preceding financial year", with proportionality for smaller companies. That number is worth knowing and not worth panicking about: the obligation is one of the cheapest in the Act to meet.
Who carries the duty — the vendor or the support team?
The Act splits responsibility between the provider (who develops the system and puts it on the market) and the deployer (who uses it). For the chatbot disclosure, the Commission's text is explicit that providers "must design and develop those systems" so that people are informed. If you run a support widget built by someone else, the designed disclosure is the vendor's job.
What that does not mean is that the support team is out of the picture. A deployer who strips the label off, restyles the widget so the disclosure is invisible, names the agent "Sarah" with a stock portrait, or instructs it to deny being an AI when asked has defeated the provider's design — and the reasonably-observant-person standard is judged on what the customer actually sees. The practical question for a support team is therefore not "have we written a disclosure?" but "does our configuration keep the vendor's disclosure intact on every surface, on every plan, in every state of the conversation?"
The header says it is the product’s support and that it is an AI, before the first message.
first interaction
Nothing in the reply claims to be a person; the header stays.
clear and distinguishable
“Are you a bot?” gets a plain yes, whatever the rest of the rules say.
cannot be overruled
A person taking over is named as a person; the switch is visible.
no blurred edge
Removing the vendor’s name swaps a noun; it never removes the AI label.
on every plan
What a compliant support chat looks like
| Surface | What the customer should see | Where teams get it wrong |
|---|---|---|
| The widget's header, on open | A line that says this is the product's support and that it is an AI, plus that it may make mistakes | A person's name and portrait; "Chat with us"; the disclosure in a settings tooltip |
| The first message | Nothing that undoes the header — no "Hi, I'm Sarah from the team" | A persona greeting that reads as a human even though the header said otherwise |
| The direct question | "Are you a real person?" answered with a plain no | A configured persona instructed to stay in character, or to deflect the question |
| The handoff | The customer told a person is taking over, and who; the agent's messages visually distinct from the person's | Agent and human replies in one indistinguishable thread; the customer never learns which was which |
| A white-label or de-branded tier | The vendor's name gone, the AI label still present | "Remove branding" implemented as "remove everything", label included |
The last row is the one to check on your own vendor, because it is the one that fails silently. De-branding is a legitimate commercial feature; it should swap the brand noun for the category noun — "AI support" instead of "Powered by Vendor" — and never leave the label empty.
How hiy meets it, in the code rather than the copy
Because we sell a support agent, here is what ours does, stated as mechanisms you can check rather than assurances:
- The header says what it is. A Support Agent says it is the product's support, in its own header, and that it may make mistakes. It says it is an AI when asked, and the handover sentence you write on the Escalation page cannot overrule that — the rules that never bend include still saying it is an AI.
- There is no persona to hide behind. A Support Agent has no voice layer: the part of the system that carries a person's phrasing and rhythm is never built for this variant, and its rules are one shared, name-free block rather than a per-agent string anyone can edit to say "you are Sarah". It cannot pretend to be a person because the machinery for pretending is absent.
- De-branding is a swap, never a deletion. The plan resolves at render to one of two nouns — the brand one or the category one — and there is no third branch that returns nothing. The de-branded wording is the more disclosing of the two, which is why it is safe to sell, and why no plan removes the AI label, the honest gap, or the report link.
- A person is a person. When a handoff happens, the customer reads that the team has their question and their email — not that "Sarah" will be right back.
- Names that could be a person are reviewed by a person. A Support Agent called "Acme Support" clears identity review quickly; one called "Ask Sarah" always waits for a human to look at it, and that is deliberate.
The full identity policy — you may only twin yourself or a product you own, and every page carries a report link — is on the trust page, and what a Support Agent is has the rest.
Beyond the letter
It would be a mistake to treat Article 50 as a compliance chore, because everything it asks for is also what makes a support agent trusted. A customer who knows they are talking to a machine forgives a refusal and reads the citation; a customer who thought they were talking to Sarah and finds out otherwise forgives nothing. The disclosure is not a cost of the agent. It is the reason the honest "I don't have that" lands as honesty rather than evasion.
The same logic is why we do not decorate the product with pictures of people who do not exist: an agent sold on not pretending to be a person cannot be marketed by pretending either.
The honest version
This is a plain-language reading of one article, by a vendor, and it is not legal advice. Article 50 also covers things this post does not — labelling AI-generated text and images, deepfakes, emotion recognition — and the Commission's guidelines and code of practice on generated content are separate documents with their own timelines. National authorities enforce it, and how they read "obvious" will be settled in cases, not FAQs.
And two things hiy does not yet do: voice is an upcoming mode, not a shipped one, so nothing here speaks to spoken disclosure; and the agent sends no email to the customer after a handoff, so there is no AI-written reply to label — when that changes, the label will have to travel with it.
Where to start
Open your own support widget in a private window, as a customer would, and read the first screen. If it does not say it is an AI before the first message, that is the fix, and it is a small one. Then ask it whether it is a person. Then hand off and see what the customer is told. If any of the three fails, talk to your vendor this week — 2 August has already passed. If you are choosing a vendor, the Support Agent demo shows the header, the answer and the refusal on a sample product's docs, so you can check the three yourself.
Questions people ask
Does the EU AI Act require chatbots to say they are AI?
Yes. Article 50, applying from 2 August 2026, requires AI systems that interact directly with people to be designed so that those people are informed they are interacting with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person. The Commission’s guidance says the information must come from the start of the first interaction, in a clear and distinguishable manner.
Is the support team or the chatbot vendor responsible for the AI disclosure?
The Act places the design duty on the provider — the vendor who builds and markets the system. A support team deploying it should not remove, hide or contradict that disclosure: a persona name and portrait, a widget restyled so the label disappears, or an instruction to deny being an AI would defeat the provider’s design on the surface the customer actually sees.
Does removing vendor branding from a chat widget breach Article 50?
Not if the AI label survives. De-branding should swap the vendor’s name for a category word — “AI support” rather than “Powered by Vendor” — and never leave the disclosure empty. Check the de-branded state of your widget specifically, because it is the one that fails silently.