Three obligations of the EU AI Act (Regulation (EU) 2024/1689) already apply to almost any company that uses an AI tool: not using prohibited practices, training the people who use the tools, and saying when a piece of content or a conversation comes from AI. The heavy rules, those for high-risk systems, are postponed to 2 December 2027 and 2 August 2028.
In short
You are a "deployer" as soon as your company uses an AI system in its activity, even if it develops nothing.
Since 2 February 2025: certain practices are prohibited, and you must take measures to develop the AI literacy of the people who use these tools for you.
Since 2 August 2026: the transparency obligations apply, for example for a deep fake or a chatbot.
Fines go up to €15 million or 3% of worldwide turnover for transparency, and up to €35 million or 7% for a prohibited practice. For an SME, the lower of the two amounts applies.
Is your company concerned?
Yes, as soon as it uses an AI system in its activity. The regulation distinguishes two roles.
The provider develops an AI system, or has it developed. It then places it on the market or puts it into service under its own name (Article 3, point 3). The deployer uses an AI system under its own authority, except for personal, non-professional use (Article 3, point 4).
An SME that uses a writing assistant, an applicant-screening tool or a chatbot bought from a vendor is therefore a deployer. One that builds its own chatbot and puts it into service under its brand may become a provider. The difference matters: the obligations are not the same.
What already applies?
| Obligation | Since | Who | Text |
|---|---|---|---|
| Not using a prohibited practice | 2 February 2025 | Providers and deployers | Article 5 |
| Taking measures for the AI literacy of staff | 2 February 2025 | Providers and deployers | Article 4 |
| Informing a person that they are talking to an AI | 2 August 2026 | Providers | Article 50(1) |
| Marking generated content as such, in a machine-readable format | 2 August 2026, and 2 December 2026 for systems already on the market | Providers | Article 50(2) and Article 111(4) |
| Disclosing a deep fake, or a generated text published to inform the public | 2 August 2026 | Deployers | Article 50(4) |
The prohibited practices include manipulation that distorts a person's behaviour to their detriment. They also include exploiting vulnerability due to age, disability or social situation, and social scoring. Inferring a person's emotions in the workplace is prohibited, except for medical or safety reasons (Article 5(1)). A tool that infers your employees' emotions from their face or voice should be checked before any use.
AI literacy does not require a diploma. You must take measures so that the people who use these tools on your behalf understand their possibilities and risks, according to their training and the intended use. The text states that you do not have to guarantee a given level for each person (Article 4).
Transparency depends on what you publish. A deep fake is an image, audio or video generated or manipulated by AI. It resembles a real person, place or event and could pass as authentic (Article 3, point 60). If you publish one, you must say so. A generated text published to inform the public on a matter of public interest must also be disclosed. The disclosure is not required if a human has reviewed it and a person holds editorial responsibility for it (Article 50(4)). The information is given at the latest at the first interaction or the first exposure to the content (Article 50(5)).
What is postponed?
The obligations for high-risk systems. They will apply on 2 December 2027 for the systems in Annex III and on 2 August 2028 for those in Annex I (Article 113, point (c)).
If you use such a system, this is what awaits you as a deployer (Article 26):
- follow the provider's instructions for use;
- assign human oversight to people who are competent, trained and have the necessary authority;
- monitor the operation of the system and suspend its use if there is a risk;
- keep the automatically generated logs for at least six months;
- inform workers and their representatives before putting the system into service in the workplace;
- inform the people about whom the system helps to take a decision.
A high-risk system already in service before those dates falls under the regulation only if its design is then significantly changed (Article 111(2)). That is not a reason to wait: a tool contract signed today will still be running in 2027.
What is the risk?
Each Member State sets its own penalties, within the caps of the regulation (Article 99).
| Breach | Cap on the fine |
|---|---|
| Prohibited practice | €35 million or 7% of total worldwide annual turnover |
| Obligations of the deployer of a high-risk system, transparency obligations | €15 million or 3% |
| Incorrect information supplied to an authority | €7.5 million or 1% |
As a rule, the higher of the two amounts applies. For an SME or a start-up, it is the lower (Article 99(6)).
Where to start?
Three steps, in this order.
- List the AI tools actually in use, including those your teams adopted on their own. For each: who supplies it, what it is for, what data goes into it.
- Write an internal usage rule and train those who use the tools. This is the direct answer to the AI literacy obligation.
- Check what you publish and what talks to your customers: generated content, chatbot, synthetic voice. Add the notices where the regulation requires them.
If one of your tools sorts people or prepares a decision about them, have it checked now to see whether it is high-risk.
What I do in practice
I draft the transparency notices for a system and the internal AI usage policy (€890 to €1,450). For a tool that raises a question, an AI Act diagnostic says in writing which category it falls into and what you must do (€1,500 to €3,400). For several systems, a full gap analysis with its implementation plan (from €4,000). Amounts excl. VAT; the exact price is given in writing after first contact, which is free.
This work is covered on the Data and artificial intelligence page.
Frequently asked questions
Is a company that only uses a writing assistant covered?
Yes, as a deployer. The AI literacy obligation applies to you, and so do the prohibited practices. The transparency notices depend on what you publish with the tool.
Must a website chatbot say it is an AI?
The regulation requires the provider to design the system so that the person knows they are talking to an AI, unless this is obvious (Article 50(1)). If you built the chatbot and put it into service under your name, check whether you are not a provider yourself.
Must every AI-generated image be disclosed?
No. The deployer's obligation concerns deep fakes: content that resembles a real person, place or event and could mislead. An obviously fictional illustration is not one. The technical marking of generated content falls on the provider of the tool.
Do you want to know what the regulation changes for your tools? Write to me about your situation in three lines.
Sources
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence, consolidated version of 27 July 2026, Articles 3, 4, 5, 26, 50, 99, 111 and 113 (CELEX 02024R1689-20260727, EUR-Lex)
This article is general information, current as of 16 November 2026. It is not an opinion on your situation.
On this subject, see also the page GDPR and outsourced DPO: services, timing and fees.
Translated from the French original published on 16 November 2026. The analysis concerns Belgian and European Union law. Read the French original.
