Support chats have become ordinary furniture on a company website. Since 2 August 2026 they carry a legal duty that most of the companies running one have never checked: the visitor has to know they are talking to an AI system.
Article 50(1) of the AI Act is the shortest of the transparency duties and the easiest to fail, because failing it looks exactly like normal product design. A friendly first-name greeting, a small avatar, an instant reply. Nothing in that sequence tells a first-time visitor what is on the other side.
What Article 50(1) actually requires
An AI system intended to interact directly with people must be designed so that those people are informed they are interacting with an AI system. The information has to reach them at the latest at the time of the first interaction, and it has to be given in a clear and distinguishable manner.
Two details are easy to miss. The duty is a design duty, not a disclaimer duty: it attaches to how the system is built, not to what your terms say. And the notice must meet applicable accessibility requirements, so a purely visual cue in a small grey font will not carry it for every visitor.
When is it obvious enough not to say?
The article exempts situations where the AI involvement is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect. The Commission’s guidance sets out a two-step assessment: identify the audience the system is aimed at, then ask whether an average member of that audience would understand that AI is involved.
The test is about your visitors, not your team. Everyone internally knows the chat is automated, and that knowledge has no bearing on the assessment. A widget plainly titled as an AI assistant will usually pass. A chat that opens with a human-sounding name and no other signal usually will not, however familiar the pattern has become.
The guidance also warns against leaning on the exception. Where the assessment is close, the cheaper course is to say it.
Where the notice has to appear, and where it must not
Clear and distinguishable rules out three placements the guidance names directly: buried in terms and conditions, rendered as faint low-contrast text, or flashed briefly and then removed. A notice the visitor has to go looking for has not been given.
What works is unremarkable. A line in the opening message. A label on the widget header that stays visible. Wording a visitor reads without effort, in the language of the page, present from the first exchange rather than introduced when the conversation turns difficult.
Who owes the duty when you bought the chatbot
Article 50(1) is written as a provider duty. If your company builds or configures the assistant, it is yours. If you licensed it from a vendor, the design duty sits with them — but the notice appears on your website, under your name, and it is your visitors who are affected when it is missing.
There is a further question worth resolving rather than assuming. A provider under the AI Act is a party that places a system on the market or puts it into service under its own name or trademark. A company that takes a licensed assistant, brands it as its own and puts it in front of customers may sit closer to that definition than it expects. The answer turns on the specific arrangement and the contract behind it, which is exactly why it belongs in the vendor conversation rather than in an assumption.
The duties that travel with a chatbot
The disclosure is rarely the only obligation in play. If the assistant generates text for the visitor, the provider also owes machine-readable marking under Article 50(2), with a transition to 2 December 2026 for systems already on the market before 2 August 2026.
Staff who operate and oversee the assistant fall within the AI literacy duty in Article 4, which national authorities have supervised since 3 August 2026. Chat transcripts are personal data, so retention, lawful basis and the information you give about them sit under the GDPR rather than the AI Act. And any feature that infers a visitor’s emotional state brings Article 50(3) into play, with emotion recognition prohibited outright in workplace and education settings.
Practical example: the Gothenburg e-commerce company
An online retailer in Gothenburg runs a chat widget licensed from a vendor. It opens with “Hi, I’m Ellen — how can I help?” and hands over to a human when the query is complex. The team calls it the bot internally. Nothing on the widget says so.
A first-time visitor has no way to tell Ellen apart from a member of staff, so the obvious exception does not carry it. The fix is one sentence in the opening message and a persistent label on the widget header. Alongside that, the retailer asks the vendor two questions in writing: how the assistant identifies itself by default, and how its generated text is marked. It also checks how long transcripts are kept and what customers are told about them. None of this is a project.
Common mistakes
Treating the notice as a legal disclaimer and moving it into the terms, where the article says it does not belong.
Relying on the obvious exception because everyone internally knows the chat is automated.
Giving the assistant a human first name and no other signal, which makes the exception harder to rely on rather than easier.
Assuming a licensed chatbot makes the duty entirely the vendor’s, without a contractual term or a written answer confirming it.
Disclosing only when the conversation escalates, rather than at the first interaction as the article requires.
Recommended actions
Open your own site in a private window and start a chat as a first-time visitor would. What you see there is what the assessment is about.
Make the notice part of the opening message, and keep a visible label on the widget itself.
Check the notice against your accessibility requirements: contrast, screen reader output, and the language of the page.
Ask the vendor in writing how the assistant identifies itself and how its output is marked under Article 50(2).
Confirm who in your organisation oversees the assistant, and that the AI literacy duty is covered for them.
Frequently asked questions
Does a rule-based chat widget count as an AI system?
Not necessarily. Article 3(1) defines an AI system by its capacity to infer, and the Commission’s guidelines place a range of ordinary business software outside the definition. A decision-tree widget that only matches keywords to prewritten answers may fall outside the AI Act entirely. A widget that generates its replies does not.
Is naming the widget “AI assistant” enough?
Usually yes, provided the label is visible from the first interaction and legible to every visitor. The test is whether a reasonably well-informed member of your audience would understand that AI is involved, and a plain label is the most direct way to satisfy it.
Which authority supervises this in Sweden?
On 12 June 2026 the government designated five national competent authorities under the AI Act: Post- och telestyrelsen, Integritetsskyddsmyndigheten, Finansinspektionen, Läkemedelsverket and Swedac. The assignment runs until 31 December 2026 and permanent legislation is still being prepared. Penalties for breaching Article 50 reach EUR 15 million or 3% of total worldwide annual turnover, with proportionality applied to SMEs and small mid-cap companies.
Conclusion
Of all the obligations the AI Act brought into force this August, this is the one that takes an afternoon. A sentence in the opening message, a label that stays put, and two written answers from the vendor. What makes it worth doing now is not the penalty; it is that the fix is trivial today and looks like an oversight once someone asks.
If you want a clear picture of which of your systems fall under Article 50, and what to prioritise, we will go through it with you. The first hour is on us.
