Since 2 August 2026, the AI Act’s transparency rules have applied. In practical terms, your chatbot must announce that it is an artificial intelligence, your AI-generated visuals must remain identifiable and your advertising deepfakes must be labelled. In Luxembourg, the CNPD is the default supervisory authority.
What has come into force, and what has been pushed back
The European regulation on artificial intelligence came into force on 1 August 2024 and became generally applicable on 2 August 2026. On that date, the European Commission announced the start of enforcement, through its AI Office and national authorities.
For a Luxembourg SME owner, the essentials come down to one sentence: it is the transparency obligations that concern you today, not the heavy obligations on so-called high-risk systems. Those have been postponed by the AI omnibus regulation, which came into force on 27 July 2026.
| What applies | Since when | Who is targeted in practice |
|---|---|---|
| Prohibited practices and AI literacy within the company | 2 February 2025 | Any organisation that uses AI |
| Rules on general-purpose AI models | 2 August 2025 | Model providers, rarely an SME |
| Transparency obligations (Article 50) | 2 August 2026 | Websites, chatbots, AI-generated marketing content |
| Annex III high-risk systems (recruitment, credit, biometrics) | 2 December 2027 | Postponed by the AI omnibus |
| AI built into regulated products | 2 August 2028 | Postponed by the AI omnibus |
In other words, the wave of paperwork compliance will wait. What is enforceable now can be seen from the outside, on your public pages. That is what makes the subject urgent: nobody needs an internal audit to notice that a chatbot does not say it is a machine.
Provider or deployer: the question that decides everything
The regulation distinguishes between two roles, which are often mixed up. Yet the distinction changes your whole list of duties.
The two roles not to confuse
The provider develops an AI system, or places it on the market under its own name. A chatbot vendor, an image generation tool.
The deployer uses that system under its own authority, as part of its business. Your SME, when it installs an off-the-shelf chatbot on its website or publishes a product visual made with an image generator.
The vast majority of SMEs are deployers. That lightens the load without removing it: the disclosure obligations towards your customers fall on you, not on your supplier. Two situations tip you into the provider role, with much heavier duties. The first: you put your own brand on an AI system bought elsewhere and offer it as your product. The second: you substantially change its purpose.
The common case in Luxembourg: an SME renames an off-the-shelf conversational agent “the Maison X assistant” and sells it to its own customers. It has just become a provider. The question deserves two minutes with your legal adviser before launch.
The 5 things to check on your website
1. Your chatbot announces that it is an AI
This is the most visible obligation and the easiest to fix. A system designed to interact directly with people must inform them that they are talking to a machine, unless it is obvious. The information must arrive no later than the first exchange, not in the legal notice.
In practice, a clear statement in the welcome message is enough. “I am an automated assistant. To reach an adviser, type adviser.” Check how it displays on mobile, where many widgets truncate the first message.
2. Your AI-generated content is marked
Synthetic content, whether images, audio, video or text, must carry a machine-readable marking that allows it to be detected as artificial. This technical obligation rests on the provider of the generation tool. Your job is therefore to choose tools that implement it and not to break that marking.
The trap lies in the production chain: cropping or re-exporting an image in an old editor often erases the metadata. Test your export chain before rolling it out widely.
3. Your advertising deepfakes are labelled
An image, sound or video generated or manipulated by AI that resembles real people, places or events, and that could pass as authentic, must be flagged as artificial. The statement must be visible to the public, not hidden in a file.
This directly targets campaigns that use synthetic models, cloned voices or reconstructed settings. A line such as “visual generated by artificial intelligence” on the creative or in the caption settles the matter. If you entrust your campaigns to a provider, this clause must appear in the contract.
4. Your texts of public interest are declared
Texts published to inform the public on matters of public interest must state that they were generated by AI. Good news for most companies: this obligation does not apply when a human has reviewed and approved the content, and a natural or legal person takes editorial responsibility for it.
This is one more argument for a formal review process on your blog. An article written with the help of AI, reviewed and signed by an identified person, falls outside the scope of the obligation.
5. Your behavioural analysis tools are declared, or withdrawn
If you use an emotion recognition or biometric categorisation system, you must inform the people exposed to it. Be careful: emotion recognition in the workplace has been among the outright prohibited practices since February 2025. An in-store attention measurement tool or an automated candidate screening tool deserves serious examination before it stays in service.
Who enforces in Luxembourg and what you risk
Luxembourg designated its authorities early. As early as November 2024, the government approved the bill giving the Commission nationale pour la protection des données the role of single point of contact, national coordinator and default market surveillance authority. Sector authorities keep their remit: the CSSF for finance, the Commissariat aux Assurances for insurance, the ILNAS for accreditation and conformity assessment.
For an SME in retail, the trades, consulting or healthcare, the contact is therefore the CNPD, which is already your contact for personal data. A useful continuity: companies that have taken their GDPR compliance seriously start with a head start.
As for penalties, the regulation provides for breaches of the obligations of providers and deployers fines of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. That ceiling is calibrated for large groups. The text provides a specific and more favourable ceiling for SMEs and young companies, to be checked case by case with a legal adviser.
The real risk for a Luxembourg SME is not the maximum fine. It is a complaint from a customer or a competitor through the complaint tool opened by the Commission, and the time lost responding to it. Fixing a chatbot welcome message costs an hour. Responding to an authority costs far more.
The realistic action plan for an SME
Here is the method we use to start this work. On a standard website, it takes half a day.
- Take inventory. List every place where AI touches your public: chatbot, recommendation engine, visual generator, writing tools, synthetic voices in your videos.
- Determine your role for each line. Deployer in almost every case, provider if you put the product under your own name.
- Fix the display. Chatbot welcome message, labels on synthetic visuals, caption on the campaigns concerned.
- Write the clause. Add to your supplier contracts the obligation to provide marked outputs and to flag any synthetic content.
- Train those who publish. Compliance is won in the daily actions of your marketing teams, not in a filed procedure.
This work overlaps with two other subjects already on the table of Luxembourg business owners: the NIS 2 cybersecurity directive and digital accessibility obligations. All three can be handled efficiently in a single website review, rather than in three separate passes.
Frequently asked questions
My website has no chatbot. Does this concern me?
Probably, yes. As soon as you publish visuals, videos or voices produced with a generative AI tool, the marking and labelling rules concern you. The chatbot is the most visible case, not the only one.
Is using ChatGPT to write my blog articles a problem?
No, in the vast majority of cases. The disclosure obligation targets texts published to inform the public on matters of public interest, and it does not apply once a human has reviewed the text and taken editorial responsibility for it. A commercial article that has been reviewed and signed is not a problem.
Who must display the notice: me or the chatbot vendor?
Both, in different ways. The design of the system is the provider’s responsibility. The actual display to your visitors is yours, since you are the one running the tool on your website. Do not count on a default setting, check what a visitor actually sees.
Do the rules on automated recruitment already apply?
No. The Annex III high-risk systems, which cover candidate screening and credit assessment among others, will apply from 2 December 2027 following the postponement decided by the AI omnibus regulation. However, emotion recognition at work has been prohibited since February 2025.
Can SME Packages fund this compliance work?
The Luxembourg scheme targets digitalisation and artificial intelligence projects, not a standalone compliance service. However, compliance built into a wider project, such as a website redesign or the rollout of an assistant, often falls within the scope. The question is settled when the application is put together, on our page dedicated to SME Packages AI.
Do our teams need to be trained, within the meaning of the regulation?
Yes. The AI literacy obligation has applied since 2 February 2025: people who use AI systems on the company’s behalf must have a sufficient level of skill. AI training for companies covers both this requirement and the productivity gain you are looking for.
How long does it take to get compliant?
For a brochure website with a chatbot and generated visuals, allow half a day for the review and a few hours of technical fixes. Timelines get longer as soon as an in-house tool is involved, or the provider role is in question.
Sources
- European Commission, press release of 31 July 2026 on the start of application of the transparency rules
- European Commission, official “AI Act” page, updated 3 August 2026 (timeline and postponement of high-risk systems)
- Official Journal of the European Union, Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence
- European Commission, guidelines of 20 July 2026 on the transparency obligations of providers and deployers
- CNPD Luxembourg, communication of 4 December 2024 on the designation of the national competent authorities









