In short
- Article 4 of the EU AI Act has applied since 2 February 2025. Rewritten in July 2026, it asks you to take measures “pour favoriser le développement de la maîtrise de l'IA” (to promote the development of AI literacy) among your teams, without guaranteeing a specific level.
- As a general rule, a company that uses an AI tool at work is covered, even just to write an advert or translate a text.
- The regulation sets no fine amount for this obligation: each country sets its own rules. Training is still worthwhile: lack of expertise is the reason most cited by European companies that considered AI without using it (Eurostat, 2025).
“Is AI training compulsory for our employees?” Short answer: you must act. But according to the European Commission, no standard training is imposed and no certificate is required.
What Article 4 said before July 2026, and what has changed
Regulation (EU) 2024/1689, often called the “AI Act”, governs artificial intelligence in the European Union. Its Article 4 covers “maîtrise de l'IA” (AI literacy). It has applied since 2 February 2025, according to the European Commission.
The original text asked companies to take measures to ensure, in the Commission's words, a “sufficient” level of AI literacy among their staff. One question stayed open: who decides that a level is sufficient?
Regulation (EU) 2026/1744 rewrote this article. It is an “omnibus” regulation: a text that amends several laws at once, here to simplify the application of the AI rules. It has been in force since 27 July 2026, according to the Commission. In its frequently asked questions, it states that no specific, or “sufficient”, level is required any more.
In its consolidated version (the original text with its amendments built in), Article 4 now provides that:
- providers and deployers of AI systems “prennent des mesures pour favoriser le développement de la maîtrise de l'IA” (take measures to promote the development of AI literacy) among their staff and the other people who use these systems on their behalf;
- these measures take account of “leurs connaissances techniques, leur expérience, leur éducation et leur formation” (their technical knowledge, experience, education and training), the context of use and the people affected by the system;
- “Cette obligation ne contraint pas les fournisseurs ou les déployeurs à garantir un niveau spécifique” (This obligation does not require providers or deployers to guarantee a specific level) of AI literacy for any given person.
In plain terms: the text asks you to take suitable measures, not to reach a specific level. To apply it to your case, talk to your adviser. The same article also asks the Commission and the Member States to support companies, “en particulier les PME” (in particular SMEs).
The regulation defines “maîtrise de l'IA” (AI literacy) (Article 3, point 56) as “les compétences, les connaissances et la compréhension” (the skills, knowledge and understanding) that allow people to use AI in an informed way and to weigh up its opportunities, risks and possible harm.
Who is covered: as a general rule, any company that uses an AI tool
The regulation distinguishes the provider, who develops an AI system and places it on the market under its own name or brand, from the deployer. The deployer is a person or organisation that uses an AI system “sous sa propre autorité” (under its own authority), except “dans le cadre d'une activité personnelle à caractère non professionnel” (in the course of a personal non-professional activity) (Article 3, point 4).
An SME that uses a chatbot, a machine translation tool or an AI agent (a program that chains tasks together from instructions) is therefore, as a general rule, a deployer. The Commission confirms this in its frequently asked questions: a company whose employees use ChatGPT to write adverts or translate texts is covered. Its teams must be informed of the risks specific to these tools, for example “hallucinations” (invented answers presented with confidence).
According to the Commission, in addition:
- the obligation also covers people who are not employees but act within your organisation, such as a contractor;
- no particular structure is imposed: no dedicated committee and no compulsory “AI officer”;
- if you use a system classed as “high-risk”, a separate obligation remains in place: training the people in charge of overseeing it.
Penalties: national regimes, no EU-wide amount
The AI Act sets maximum fines for certain infringements (Article 99, paragraphs 3 to 5). Article 4 is not among them.
For the rest, Article 99, paragraph 1, applies: “les États membres déterminent le régime des sanctions” (Member States determine the penalty regime). These penalties must be “effectives, proportionnées et dissuasives” (effective, proportionate and dissuasive). States also take into account “des intérêts des PME” (the interests of SMEs) and their economic viability.
Enforcement falls to national market surveillance authorities (the bodies in charge of checking that the regulation is applied). According to the Commission, they have been checking since August 2026, and any penalty must be proportionate and take the particular case into account.
Any amount therefore depends on the law of your country: check with your adviser. If someone tells you about a specific EU-wide fine for “lack of AI training”, ask them for the exact article.
Why train anyway: lack of expertise, the most cited obstacle
The best reason to train is a practical one.
Eurostat, the Union's statistical office, surveyed European companies with 10 or more people in 2025. Among those that had considered AI without using it, the three most cited reasons (several answers possible) are:
- lack of expertise: 70.89%, or nearly 71%;
- lack of clarity about the legal consequences: 52.52%;
- fear of breaching data protection and privacy: 48.83%.
The most cited obstacle is therefore skills, ahead of law and data. A team that knows what a tool does well, where it gets things wrong and what must never be entered into it uses it better, and with fewer risks.
This is even truer with an AI agent. At Kendu Intelligence, at the first autonomy levels, the agent prepares and the team approves every action. This approval is only worth something if the person can read the proposal, spot an error and correct it. Otherwise, approving becomes a click out of habit. The Commission puts it in its own way: in many cases, relying on the tool's user instructions, or asking staff to read them, may be ineffective.
A proportionate approach: inventory, charter, short session, written records
The Commission writes that there is no one-size-fits-all approach, and no compulsory training is imposed. A simple, tailored approach helps you show that you have taken measures. For example, an SME with 30 employees can proceed as follows:
- Take an inventory. List the AI tools in use, including those that employees adopted on their own.
- Write a one-page usage charter. Authorised tools, data never to be entered (clients' personal data, confidential information), compulsory review before anything is sent, who to alert in case of doubt.
- Run a short session for each group. Tailor it to the job: accounting does not use AI in the same way as the sales team. Work on real cases made anonymous and show typical errors.
- Keep a record. Note the date, the participants, the content and the materials. The Commission states that no certificate is needed: you can keep an internal register of training and other support actions.
- Keep it up to date. Review again whenever a new tool arrives.
The hands-on introduction included in every Kendu Intelligence set-up
A hands-on introduction for your team will be included in every AI agent set-up by Kendu Intelligence. It will be a short session in which your colleagues will learn to:
- read the agent's proposals;
- approve, correct or reject an action;
- report a problem or a doubt.
The aim: for your team to be able to approve the agent's work on its own.
This introduction is part of our method. Each agent's autonomy progresses by level, from 0 to 3. At level 0, the agent suggests without acting and the team does the work itself: we compare the two, and the team practises before the agent acts. The agent moves up a level only on the strength of the measurements (agreement rate, errors) and with your written agreement. It drops back a level after a serious alert. Certain actions, such as a payment, never go beyond level 1, where a person approves every action. In the stop cases written in its rulebook, the agent stops and hands over to you rather than guess.
Remember to note this session in your internal register: it may form part of the measures you take under Article 4.
To go further, other training courses are in preparation; they will be offered by quote.
Please note: Kendu Intelligence is not Qualiopi-certified, the quality certification for training providers in France. Since 1 January 2022, a training provider must hold it to benefit from public or pooled funds, paid by funders such as the OPCOs (skills operators), the State or the Regions. The training courses we are preparing therefore cannot, for now, be funded by these bodies.
Sources (accessed on 12 and 13 October 2026)
- Articles 3, 4 and 99 in the version consolidated on 27 July 2026: AI Act Service Desk, European Commission
- European Commission: questions and answers on AI literacy and entry into force of the omnibus
- Eurostat, use of AI in enterprises
- DREETS Normandie, Qualiopi
Official texts: Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744, on EUR-Lex.
Going further
- Our training: the hands-on introduction included and the courses in preparation.
- Our method: how an agent goes from suggesting to autonomy, level by level.
- Request an audit: a one-page report on your tools, your processes and the useful agents, by quote.