From 2 August 2026, a company that lets visitors talk to a conversational agent must tell them clearly that they are talking to a machine. That is the core of Article 50 of the European regulation on artificial intelligence, the AI Act. The rule is about transparency, not prohibition: you keep the right to use AI anywhere in your business, provided you do not let anyone believe a human is involved when none is.
The subject is anything but theoretical for a Belgian SME. According to the Belgian Digital Economy Overview report published by the FPS Economy on 9 June 2026, 34.54 % of Belgian companies used at least one AI technology in 2025, up from 24.71 % a year earlier. The European average stands at 19.95 %, and Eurostat measured 20.0 % of user companies across the Union in 2025, against 13.5 % in 2024. In short: Belgian companies are in the European leading group, and therefore among the first concerned.
The detail that matters: 28.82 % of small Belgian companies already use AI, according to the same report, against 76.41 % of large ones. An SME that has plugged a chatbot into its website, that generates its social media visuals or that has drafts written by an assistant is fully within the scope of the regulation, without a legal department to point it out.
You keep the right to use AI anywhere in your business, provided you do not let anyone believe a human is involved when none is.
What Article 50 actually requires, and from whom
Are you a provider or a deployer?
The regulation distinguishes two roles, and confusing them is the first source of needless worry. The provider develops an AI system and places it on the market. The deployer uses it in the course of a professional activity.
An SME using an off-the-shelf tool is almost always a deployer, never a provider. That changes everything: the heavy technical obligations, such as machine-readable marking of generated content, fall on the tool's publisher, not on you. Your obligations are information obligations, and they are settled with a few lines of text in the right place.
The four situations covered
Article 50 covers four cases, and only one affects most SMEs day to day:
Direct interaction with an AI: a visitor chatting with a chatbot or a voice assistant must know they are dealing with an automated system, from first contact. Exception provided: cases where this is obvious to a reasonably well-informed person.
Marking of synthetic content: the outputs of generative AI systems must be detectable as artificially generated, in a machine-readable format. This obligation targets the provider.
Emotion recognition and biometric categorisation: the deployer must inform the people exposed. A rare case in an SME.
Deepfakes and certain published texts: the deployer must disclose the publication of image, audio or video content manipulated to appear authentic, as well as AI-generated text published to inform the public on matters of public interest. With one important reservation: the obligation falls away if the content has undergone human review and a person takes editorial responsibility for it.
No, you do not have to label every AI-generated image
This is the most widespread fear, and it is unfounded in the vast majority of cases.
The case of marketing visuals
An abstract illustration generated to accompany a blog article, a banner background, a pictogram: none of that constitutes a deepfake within the meaning of the regulation. The text targets content resembling real people, objects or events closely enough to be wrongly taken as authentic. A clearly illustrative visual does not tick that box.
A realistic photo of a team that does not exist, a fake customer video testimonial, an image of your worksite rebuilt from scratch: there you are squarely within scope, and the absence of a notice becomes a problem. The subject connects with the credibility questions we raised about AI-written content and what Google tolerates.
The case of the chatbot on your website
This is the most concrete obligation and the easiest to handle. If your site offers automated conversational support, the first message must state that this is an automated assistant and, ideally, how to reach a person. The information must be provided in a clear and distinguishable manner, at the latest at the time of the first exchange. One sentence is enough, provided it is visible and not buried in terms and conditions.
The same reasoning applies to automated email replies, request qualification systems and agents processing your forms. If you have already set up that kind of mechanism, for instance as part of automated lead follow-up, compliance is a question of wording, not of technical rebuild.
The real timeline, after the digital omnibus
The regulation was amended this summer. Regulation (EU) 2026/1744, known as the digital omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force a few days later. It postpones part of the deadlines, which changes how the calendar reads.
What does apply on 2 August 2026: the transparency obligations of Article 50, including disclosure of interaction with an AI and the flagging of deepfakes.
What is postponed to 2 December 2026: the marking obligations for synthetic content produced by general-purpose AI models, with four extra months. The new prohibition on applications generating intimate images without consent also takes effect on that date.
What is postponed further: the rules for high-risk systems, to 2 December 2027 for stand-alone systems and to 2 August 2028 for those built into regulated products. A typical SME is only affected by that part in specific cases, for instance if it uses an automated application-screening tool.
Worth noting: the Commission published guidelines on Article 50 on 20 July 2026, together with a code of practice on disclosing AI-generated content. That is the reference document for settling borderline cases.
What you risk, and what compliance costs
Article 99 of the regulation provides, for breaches other than prohibited practices, administrative fines of up to 15 million euros or 3 % of total worldwide annual turnover. The figure is frightening, but you have to read the next paragraph: for SMEs and start-ups, the lower of the two amounts applies. For a company with one million euros of turnover, the theoretical ceiling drops to 30,000 euros.
That ceiling remains dissuasive, and it says nothing about the reputational cost. In Belgium, the BIPT has been designated as the main market surveillance authority for the AI Act, while the FPS Economy centralises the information aimed at businesses. Belgium fell behind the European timetable for designating authorities, which left a grey area for several months, but that grey area is closing.
On the other side, the real compliance effort for an SME is counted in hours, not weeks: list the AI tools in service, add a notice to automated contact points, agree an internal rule on realistic visuals, document the decisions. It is light governance work, comparable to what bringing cookie banners and Consent Mode into line once was.
Frequently asked questions
Do I need a notice on every text written with AI assistance?
No. The disclosure obligation for texts targets those published to inform the public on matters of public interest, and it falls away if a person has reviewed the content and takes editorial responsibility for it. A service page or a product sheet you have reviewed yourself is not concerned.
Is my chatbot vendor responsible in my place?
They carry the design and technical marking obligations. Informing the visitor is on you, as the deployer. Check what the tool displays by default: many vendors have added an automatic notice, but not all of them, and not always in your language.
What if my tool was placed on the market before 2 August 2026?
Machine-readable marking obligations for synthetic content get an extension to 2 December 2026 for those systems. Your obligations to inform the visitor get no reprieve at all.
Is a Belgian company using an American tool concerned?
Yes. The regulation applies as soon as the system is used within the Union, whatever the country where the publisher is established. It is the place of use that counts, not the provider's head office.
Priority action plan
List: put on a single page every place where AI intervenes in front of a customer, chatbot, automated replies, visuals, published texts.
Display: add a clear notice at the first contact of every automated exchange point, in French and in the other languages of your site.
Settle the visuals: set a simple rule, a mandatory notice as soon as a generated image or video could pass for authentic, nothing for clearly decorative illustrations.
Question your vendors: ask in writing what their tools mark, what they display, and from which date.
Document: keep a dated record of your decisions and of the human reviews, that is what makes the difference during an inspection.
Belgian companies have taken a lead on AI adoption. The lead on transparency is a shorter piece of work, and it shows: telling your visitors when they are talking to a machine costs a few lines and earns trust. Those who deal with the subject before the deadline will simply have one less conversation to have with a regulator, and one more to have with their customers.
Unsure what the AI Act changes in your tools? We review your automations and put the notices in the right place: let's talk about your project.
Automation
August 1, 2026
9 min read
