Do you have to tell customers they are talking to AI?

Short answer: yes. If a customer is talking directly to an AI system, they have to know it. That rule sits in Article 50 of the AI Act (Regulation (EU) 2024/1689) and has applied since 2 August 2026. There is one exception: you do not have to say it if it is already obvious to a reasonably attentive customer. In practice that exception is narrower than most business owners hope, and it disappears fastest precisely when your AI assistant works well.

This page explains what the rule actually says, when you have to disclose, where the notice belongs, what applies to text and imagery generated by AI, and why this is a different obligation from the information duty under the GDPR. Those two get mixed up in conversation, but they are two separate duties and both apply to you.

Where does this duty sit, and since when does it apply?

The transparency duty sits in Article 50 of the AI Act. That is European legislation with direct effect: you do not have to wait for the Netherlands to turn it into national law, because the regulation applies of its own accord in every member state.

When the regulation became applicable is set out in Article 113. It says the regulation enters into force on the twentieth day following its publication and applies from 2 August 2026. Part of the law applied earlier — the prohibited AI practices and the rules on AI literacy applied from 2 February 2025 — and the heaviest rules for high-risk systems come later. But the transparency duty belongs to the main body of rules that started on 2 August 2026. Which means it applies right now.

The Dutch data protection authority explains it the same way. In its public guidance it names exactly the two situations most small businesses are in: calling a practice and chatting with a web shop. From 2 August it has to be clear whether you are dealing with an AI assistant or with a person. This is not a theoretical rule for large tech companies; it touches the chat widget on your site and the phone line you let answer calls.

What does Article 50 actually say?

Article 50 is not a single sentence but a series of paragraphs, each covering a different kind of AI. Three of them matter for a small business with customer contact.

Talking to a machine

The first paragraph covers AI systems intended to interact directly with people. They must be built and set up so the person concerned knows they are dealing with an AI system. That is the core of it: a customer asking a question through your chat, your WhatsApp number or your phone line must not be left assuming a member of staff is reading along when nobody is.

Generated text, imagery, audio and video

Another paragraph is not about the conversation but about the output: material created or manipulated by AI. Think of a generated photo, a generated voice or a generated video. That material has to be recognisable as artificial. Further down this page you will find what that means concretely for your marketing.

The moment and the form

The final paragraph of Article 50 is the most important one in practice. It says the information must be provided at the latest at the time of the first interaction or exposure, in a clear and distinguishable manner. That is a strict requirement, and it rules out two habits you see everywhere. The notice may not wait until the customer asks for it. And it may not be tucked away in your privacy statement or in a document behind a link, because that is not "at the time of the first interaction" and it is not "distinguishable".

When is it obvious, so that you do not have to say anything?

Article 50 makes an exception for situations where it is already clear by itself. The law does not measure that from where you sit, but from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use. That is legal language for: would an ordinary, moderately alert customer notice?

In practice that exception works against you as soon as your AI is good. An assistant that answers fluently and naturally, that uses a human name, that has a real voice on the phone or that responds to what the customer just said, is exactly the kind of thing an average customer does not recognise as a machine. The better it works, the less you can lean on "surely they could tell".

There is another reason not to gamble on the exception. In a dispute afterwards, you are the one who has to make it plausible that it was obvious. That is difficult to prove. One sentence at the first message costs you nothing and removes the whole question. So in doubtful cases the answer is always: disclose.

How and where do you put the notice?

The law does not prescribe a literal wording. It does prescribe the moment and the manner: at the first interaction, clearly and distinguishably. Translated to the channels small businesses actually use, that means the following.

• Chat on your website. Put it in the first message the visitor sees, not in a tooltip and not in the footer. An opening line such as "You are chatting with an AI assistant. Say the word if you would rather speak to a colleague" does the job.
• WhatsApp. Same principle, but note that a customer sometimes picks a conversation back up days later. Repeat the notice when a conversation restarts after a longer gap, because that is a first interaction again.
• Email. If AI sends replies by itself, the notice belongs in the message itself, at the top or in a fixed line near the sender. Not only in the signature at the bottom of a long email.
• Phone. Say it in the opening sentences, before the customer starts explaining their question. On the phone the customer cannot read anything back, so the notice has to be explicit and calm enough to be heard.

Two things you had better not do. Do not give the assistant a name and a profile picture that suggest a member of staff; that works directly against the duty. And do not switch the notice off because it does not "look nice" in the design. The requirement is that the information is distinguishable, which means grey on grey in four pixels does not count.

Also make sure there is a way through to a human. That is not a standalone requirement under Article 50, but it is what makes the notice useful: a customer who knows they are talking to AI wants to know what to do when the AI cannot help them further.

Which situations need a notice and which do not?

The distinction that matters is not "am I using AI?" but "is the customer communicating directly with the AI, or with a person who uses AI as a tool?". In the second case the customer is talking to you, and then there is no AI system interacting with them directly.

Common small-business situations, with whether you have to tell the customer AI is involved and why

Situation	Disclose?	Why
Chat widget on your site answers customer questions itself	Yes, in the first message	The customer is communicating directly with an AI system; Article 50 requires the notice at the latest at the first interaction
AI answers incoming email on its own	Yes, in the message itself	Email is direct communication too, and the customer cannot see who wrote the reply
AI drafts a reply, a member of staff reads it, edits it and sends it	No	The customer is communicating with a person who is responsible for the content; here AI is a tool behind the scenes
AI receptionist answers the phone and handles the call	Yes, in the opening sentences	On the phone there is no text to read back, so the notice has to be audible and early
AI calls customers itself in order to sell something	Yes, plus the statutory opening information	Alongside Article 50, distance selling triggers Article 6:230v of the Dutch Civil Code: state at the start of the call who is calling, on whose behalf and for what commercial purpose
Assistant with a human name and profile photo	Yes, and more emphatically	The setup creates the impression of a member of staff, so the "it was obvious anyway" exception is precisely what you cannot rely on
Phone menu with "press 1 for sales"	No	A fixed menu is not an AI system holding a conversation; there is nothing artificial to mistake for a person
Automatic acknowledgement of receipt from your mail system	No	A standard message without AI does not fall under Article 50, although the message obviously still has to be accurate
Photorealistic image invented by AI, used as a photo of your business	Yes, mark it as artificial	It looks like a real recording of a real place and can therefore pass for authentic

What applies to text and imagery made by AI?

Besides the conversation, Article 50 covers the output of generative AI. Here it helps to keep two layers apart.

The first layer sits with the builder of the model. Providers of AI that produces synthetic text, imagery, audio or video have to mark that output in a machine-readable format, so that it can be detected as artificially generated or manipulated. That is engineering you do not build yourself as a business owner; it sits in the tooling you use. What you do have to do with it: ask about it when you buy, and do not strip the marking out.

The second layer sits with you, as the party publishing the material. Two cases stand out.

• Imagery, audio and video that look real. Where material imitates existing persons, places or events and could wrongly pass for authentic, you have to disclose that it was artificially created. A generated "team photo" or a generated voice in an advertisement falls under this sooner than a clear illustration or an abstract background.
• Text published to inform the public on matters of public interest. For that kind of text there is a disclosure duty, unless a human has reviewed or edited the text and somebody carries editorial responsibility. An ordinary product description or a quote sent by email is not text informing the public on a matter of public interest, so no separate AI label hangs under it.

The Dutch government summarises it for business owners in the same way: if you offer systems people come into direct contact with, such as chatbots or generative AI, your customers have to know they are using an AI system. That is the shortest version of the whole rule, and it covers both the conversation and the generated content.

One more point that does not come from the AI Act but sits right next to this duty: everything you publish has to be factually correct. AI that states a wrong delivery time, a wrong property or a wrong offer amounts to a misleading commercial practice, even without any intent involved. Saying it is AI does not make incorrect information permissible.

Are you the one who has to disclose, or your supplier?

The AI Act works with roles. Article 3 defines the deployer: the party using an AI system under its own authority. If you buy an AI assistant and put it on your own channels, you are usually that deployer, not the provider. The provider is the party that builds the system and puts it on the market.

That distinction matters for who carries which duty. The requirement that a system be designed so the user knows they are talking to AI is addressed to the provider. That is not a free pass for you, for three reasons.

1. The notice has to actually reach your customer. If the feature exists in the product but you switched it off, or you overwrote it with your own greeting without a notice, your customer never sees it. That was your choice.
2. The obligations that rest on deployers — such as the disclosure duty for certain kinds of generated material — sit with you directly.
3. The division of roles can shift. Article 25 provides that a deployer who puts its own brand name on a high-risk system, or substantially modifies it, is regarded as a provider under the law. If you build something bespoke yourself, the odds of landing in the heavier role go up.

Practically: agree with your supplier who shows the notice, what it looks like, and that it does not quietly disappear in an update. That is a question for the contract, not for the support chat afterwards.

How is this different from the GDPR information duty?

This is the mistake people make most often. Business owners have something about AI in their privacy statement and assume they are done. They are not: these are two obligations from two different laws, with a different trigger, different content and a different moment.

Article 13 GDPR is about personal data. As soon as you collect data from the customer themselves, you have to tell them, among other things, the purposes of the processing and the legal basis for the processing. Your retention period, the recipients of the data and the customer's rights belong there too. That is an information package, and it belongs in a privacy statement you point to from your channel.

Article 50 of the AI Act is not about data but about confusion. The question is not what you do with the data, but whether the customer knows what kind of counterpart they are talking to. That notice is short, appears on screen or is spoken over the phone, and has to be there at the first interaction.

They can also apply independently of one another. An AI assistant that only gives general opening hours and product information and stores nothing about the customer falls under Article 50 but barely touches the GDPR. The other way round: a member of staff entering customer data into a system falls under the GDPR without any AI involved. As soon as you let AI answer customer questions and process data while doing so, you have both.

What does this mean on the phone?

The phone is the channel where the most rules apply at once and where there is the least room to read anything back. Alongside Article 50, consumer law comes into play there.

If you call a consumer with a view to concluding a distance contract, Article 6:230v of the Dutch Civil Code requires you to state at the start of the call your identity, where applicable the identity of the person on whose behalf you are calling, and the commercial purpose of the call. That is a statutory opening line, not a courtesy. So AI calling to sell something has to say three things in its opening sentences: that it is an AI assistant, which company it is calling for, and that the call is about an offer.

If you record calls, for example to improve the AI or to look an appointment back up, a separate question arises. That falls outside this page, but do not fold it into the transparency notice: that there is AI on the line and that the call is being recorded are two different things to say.

What can you check right now?

A short list that lets you establish in half an hour whether you are in order on this point.

1. Open every channel where AI answers, as a customer. Do you see or hear in the first message that it is AI, without clicking anything? If not, that is the first thing to fix.
2. Check that the notice is visible on a phone, not only on a wide screen.
3. Check names and photos. If the assistant suggests it is a member of staff, change that or make the notice more emphatic.
4. Go through your published imagery. If there is generated material that could pass for a real photograph, it needs a disclosure.
5. Agree with your supplier who shows the notice and that an update does not remove it.
6. Separately, check your privacy statement for purpose and legal basis. That is the GDPR side and it sits alongside Article 50, not instead of it.
7. Make sure the customer can always reach a human, and that the AI says so when it cannot work something out.

This is general explanation of legislation, not legal advice about your specific situation. The legal provisions are named so you can look them up yourself.