Since 2 August 2026, a chatbot on your website or on WhatsApp has to make clear, from the first exchange, that the person is talking to an artificial intelligence system. That is the transparency rule in Article 50 of Regulation (EU) 2024/1689, the EU AI Act. In practice it calls for a visible, accessible notice, and it is a good moment to review conversation logs, hand-over to a person and staff training too. Below we go through each point the way we handle it in the assistants we build. This is a practical guide, not legal advice.
What Article 50 says
The AI Act sorts AI systems by risk. An assistant that answers questions about opening hours, products or bookings is not a high-risk system, but it does fall under the transparency obligations. The core rule is short: AI systems designed to interact directly with people must inform them that they are interacting with an AI.
There is one exception, for cases where this is already obvious to a reasonably well-informed, observant person given the context. A chat bubble in the corner of a website, with a first name and fluent replies, rarely qualifies. An assistant that introduces itself as "Sophie" reads like a person. So we treat the notice as mandatory in every case.
The same article sets out how the information must be given: clearly, distinguishably, at the latest at the first interaction, and in line with accessibility requirements. A notice that only exists inside an image, or that a screen reader never announces, does not meet that bar.
Who is responsible for the notice
The regulation puts this duty on the provider of the system. A provider is not only whoever builds the language model. It is also anyone who has an AI system developed and puts it into service under their own name or trademark. A business that commissions an assistant for its website and runs it under its own brand may well count as the provider of that assistant. If in doubt, check with a lawyer.
If you use an off-the-shelf tool from another company, find out how it handles the notice and whether you can adjust it. Either way, your website and your WhatsApp number are where customers will look for that information, and your brand is the one exposed if it is missing.
What the notice looks like on a website chatbot
On a website AI chat assistant, the notice shows up at three points:
- Before the conversation. The chat button or panel says it is an AI assistant, for example "Virtual assistant (AI)".
- In the first message. The assistant introduces itself as an AI before answering anything, and says it can pass the conversation to the team.
- Throughout the conversation. A small label in the chat header stays visible for anyone who opens the panel later or comes back another day.
Keep the wording short and free of legal language. "Hi! I'm the shop's virtual assistant, an AI system. I can pass you to the team if you'd rather talk to a person." works far better than a disclaimer in the footer.
In the assistants we build, the notice is part of the chat component itself rather than a setting someone could switch off by mistake. The header and first message are ordinary text, so screen readers announce them like any other content on the page.
On WhatsApp and social media
On WhatsApp, Messenger or Instagram you do not control how the conversation looks, so the notice has to live in the text. In an AI assistant for WhatsApp and social media, we do it like this:
- the first reply in every conversation says it comes from an automated AI assistant;
- the business profile on the channel states that first replies are handled by an AI assistant;
- when a customer returns to a conversation that has been idle for a long time, the assistant reminds them who they are talking to.
These channels need one extra bit of care: the assistant and your staff reply from the same number. When a person takes over, the customer has to notice the switch. A plain message such as "From here on, you're talking to a member of our team" does the job.
Handing over to a person
Article 50 does not require someone to be available. But an assistant that cannot step aside frustrates people, and frustrated customers stop trusting the brand. So we always design the hand-over to your team:
- the customer can ask for a person at any time, in their own words, such as "can I talk to someone?";
- the assistant hands over when it has no reliable answer, instead of making one up;
- outside working hours, it takes the request and states when the team is available;
- whoever takes over sees the full conversation, so the customer never has to repeat themselves.
On sensitive subjects, such as health, payments or complaints, the rule is to hand over early.
Conversation logs
For assistants like these, the AI Act does not require the automatic logging it imposes on high-risk systems. Even so, keeping conversations is the only way to know what the assistant actually said, fix wrong answers and respond to a complaint with facts.
Conversations are personal data, so the GDPR applies in full:
- the privacy policy explains that conversations are stored, why and for how long;
- the retention period is fixed, with automatic deletion at the end;
- access is limited to the people who need it, and data sent to the AI model provider is covered by a data processing agreement;
- the assistant never asks for data it does not need, such as ID numbers or health details.
For an assistant that is already live, a security and GDPR technical review checks these points and lists what is missing.
AI literacy for your staff
There is a second duty many businesses have not heard of. Since 2 February 2025, Article 4 of the AI Act asks providers and deployers of AI systems to take measures so that the people working with those systems have a sufficient level of AI literacy. The European Commission has published questions and answers on this duty.
For a team running a chatbot, that comes down to concrete skills: knowing the assistant can be confidently wrong, reading and reviewing conversations, knowing when to update the information it draws on, and knowing who handles a difficult case. It does not take a long course or a certificate. A hands-on session with the assistant itself, an internal page with the rules and a record of who attended cover the essentials. That is the format we use in AI consulting, training and AI Act review.
Checklist
- The assistant says it is an AI before its first answer.
- The notice stays visible during the conversation and is announced by screen readers.
- On WhatsApp and social media, the notice is in the first message and in the profile.
- The assistant never presents itself as a member of staff.
- Customers can ask for a person at any time, and the hand-over carries the history.
- Conversations are kept for a fixed period and covered by the privacy policy.
- The team running the assistant has had practical training, and it is on record.
If your assistant is already live
Start with two quick tests. Open the chat in a private window, as a new customer would, and check that it is clear you are talking to an AI before any answer appears. Then ask to speak to a person and see what happens. Those two tests show the gaps fast.
If the notice is missing, changing the welcome message and the chat header is usually enough. If the hand-over or conversation logs are missing, there is more work involved, but it can be done without replacing the assistant.
If you would like help reviewing or rebuilding your assistant, request a quote. Every project is quoted individually, based on what your assistant already does and what needs to change.