
The AI Act and your website: what does an SME really need to do?
What changed in August
"I'd like the messages from my contact form sorted automatically: prospect, existing customer, supplier or spam. Is that possible, and is it even allowed now that the AI Act is here?"
A client asked me that recently, and they're not the only one. Since 2 August 2026, the transparency rules in Article 50 of the European AI Act apply. The European Commission's questions and answers state it in black and white. Ever since, it keeps coming down to the same thing: do I need to do anything with my website?
The honest answer is that it depends on what happens on that website. A site with texts and photos made by people, and no AI otherwise, does not fall under the AI Act as far as I can see. The obligations are about AI systems, you see: systems that talk to people, systems that generate content, and what you do with that generated content.
That client's contact form falls largely outside it. Sorting like that doesn't talk to visitors and doesn't produce anything you publish, so technically it's perfectly possible, and the AI Act asks little more of you than knowing what the system does. Whether it's allowed depends mostly on the GDPR, because a form like that holds personal data. At the very least, check where that data goes, and whether you have a data processing agreement with the supplier of the AI system. Beyond that, this piece isn't about the GDPR; that law still applies, of course.
If you do use AI on or for your site, there are three situations to watch. Separately there is a fourth, which is about your people rather than your website.
A chatbot on your site
If an AI system talks directly to your visitors, they need to know they are dealing with AI from the start of the first conversation. Only when that is obvious to a reasonably well-informed visitor is it not required. The Commission explicitly adds that this exception is read narrowly. A chat window with a human name and a photo is therefore not "obvious".
Formally, this duty lies with the provider of the system: whoever develops it or has it developed and offers it under their own name. If you use an off-the-shelf chatbot from a supplier, law firm Stibbe sees you as the deployer, and the supplier should take care of the notice. If you have the chatbot built and put it online under your own name, you may be the provider yourself. That borderline case is not described equally sharply everywhere.
In my opinion, you do not need to think about it for long. Put a clear line at the top of the chat window, or check that your supplier does. It costs nothing and you are covered, whatever your role.
Images and video made with AI
Here the duty does lie with you. If you publish a deepfake, you have to disclose it visibly or audibly, at the latest when someone first sees or hears it. An invisible watermark added by the image generator does not count.
What is a deepfake? According to the Commission, it is AI-generated image, audio or video that resembles existing or plausibly existing people, objects, places or events, and that falsely gives someone the impression it is real. Context matters, and so do your audience's expectations. A realistic photo of a "happy customer" who never existed falls squarely under it, in my opinion. An obvious fantasy image or a satirical drawing gets a lighter rule: an appropriate disclosure that does not spoil the work.
Two practical points. Marking AI images in a machine-readable way is the job of whoever makes the generator, not yours; existing systems have until 2 December 2026 to do so. And content generated before 2 August 2026 does not need to be labelled retroactively, according to the Commission, although it does encourage it where possible.
Text written with AI
For most SME sites this is the shortest part: product pages and service descriptions don't fall under it. The labelling duty for text is narrow, you see. It only concerns AI-generated text you publish to inform the public on matters of public interest, and the Commission gives politics, public health and consumer safety as examples.
Even then there is an exception. If a person has reviewed the text on substance, and someone holds final responsibility for publishing it, the label is not needed. A spell check or a quick glance does not count as a review.
Do you write about health or safety, in a blog for instance? Then have such a text checked by someone who is genuinely responsible for it, or add a label.
AI literacy of your people
This duty has been around longer: since 2 February 2025, Article 4 of the AI Act asks organisations that use AI to do something about their people's AI knowledge. With the Digital Omnibus, Regulation (EU) 2026/1744, that text was amended as of 27 July 2026. You now have to take measures to support the AI literacy of your staff, but you do not have to guarantee a specific level of knowledge per person (Law & Technology, Cooley).
If you work as a freelancer, the Commission considers you a deployer as soon as you use AI professionally. The duty covers your staff and people who work with AI on your behalf, such as subcontractors. If you work entirely on your own, there is little to arrange: make sure you know what your AI tools do and what you use them for; a short note on that is a sensible step.
If you do have staff or outsource work, write down briefly what you do: a training, an internal agreement, an overview of which tools you use and what for. The Dutch Data Protection Authority also recommends recording your approach and progress (an article on its guidance, in Dutch).
What it costs if you don't
Breaching Article 50 can lead to a fine of up to 15 million euros or 3% of worldwide annual turnover. For SMEs and start-ups, the lower of the two applies (Article 99). Article 4 is not on that list of fines. That doesn't mean it can never carry a penalty, though: Member States set their own penalties for breaches of the regulation, so they can enforce Article 4 too. What the Netherlands does with that depends on the implementation bill below.
Who supervises this in the Netherlands is not fully settled yet. In April 2026 the Dutch government put the AI Act implementation bill (in Dutch) out for consultation. Existing regulators will watch AI within their own fields, the Dutch Data Protection Authority takes the areas without a clear regulator, and it coordinates together with the national digital infrastructure inspectorate (RDI). In October 2026, according to Dutch IT Channel (in Dutch), that bill was still being worked on.
In short: how to approach it
- Make a list of all AI on and behind your site: chatbot, search, generated texts, images and video.
- If you have a chatbot, make sure every conversation starts with a clear AI notice.
- Label realistic AI images of people, places or events visibly.
- Have AI texts on matters of public interest reviewed on substance, or label them.
- Record what you do about AI literacy and keep it up to date.
Next week I'll run through this list for my own site.
This is not legal advice; for that, a lawyer remains the right address. But for most SME websites it is, I think, less work than the stories suggest. In doubt about your own site? Feel free to send me a message.
Sources
- Transparency obligations under Article 50 of the AI Act, European Commission, updated 24 July 2026
- Guidelines on transparency obligations for providers and deployers of AI systems, European Commission, 20 July 2026
- Article 4: AI literacy and Article 99: Penalties, rendering of the legal text
- AI literacy: the Digital Omnibus rewrites Article 4 of the AI Act, Law & Technology, 30 July 2026
- Digital AI Omnibus Delays Key Deadlines, Introduces New Rules, Cooley
- Salaris Vanmorgen on the Dutch Data Protection Authority's AI literacy guidance (in Dutch), 23 October 2025
- The AI Act's Transparency Obligations: Rules, Scope and Timeline, Stibbe, 27 July 2026
- Dutch government announcement on supervising the European AI rules (in Dutch), 20 April 2026
- Dutch IT Channel on the next step in AI supervision (in Dutch), 13 October 2026
"company_name": "Bas de Kort",
"legal_form": "sole proprietorship",
"trade_names": ["Bas de Kort", "LengthOfRope"],
"address": "Kerkeakker 3, 6674 EM Herveld, NL",
"chamber_of_commerce": "42038027",
"vat_id": "NL004476145B32",
"contact": { "form": "basdekort.nl/en/contact", "telephone": "+31628456862" },
"documents": [ "privacy", "terms and conditions", "data processing agreement", "colophon" ]
}