Local time

01AI at work · IA · AI Act · Formation

AI Act Article 4: training your teams, what is actually required

Published on · 6 min read

An SME of twelve people has been using an AI assistant for eight months, for quotes and meeting minutes. Nobody has ever been trained, and the question arrives one morning, from a client: “are you compliant with the AI Act?”

The answer comes in two parts. Yes, there is a European obligation that bears on AI training, and it has applied for more than a year. No, it is not quite the one you were told about: it was rewritten in July 2026, and what it asks is both softer and more concrete than what you read about it.

What is true in 2026

The text is Regulation (EU) 2024/1689, the AI Act. Its Article 4 covers “AI literacy”. The European Commission states, in its questions and answers devoted to AI literacy, that this article entered into application on 2 February 2025, at the same time as the general provisions and the prohibitions.

Article 3, point 56 of the regulation defines AI literacy as a set of skills, knowledge and understanding that allows the actors concerned to make an informed deployment of AI systems, as well as to gain awareness about the opportunities and risks of AI and the possible harm it can cause. Three words carry the meaning: skills, risks, harm.

Second part, the one many people missed. Article 4 was rewritten by Regulation (EU) 2026/1744 of 8 July 2026, the “digital omnibus” on AI, published in the Official Journal of 24 July 2026 and entering into force on the third day following its publication. EUR-Lex publishes a consolidated text applicable as of 27 July 2026. In that wording, the obligation on providers and deployers is to take measures to support the development of AI literacy, and the text specifies that it does not require guaranteeing any specific level for any given individual. Check the version in force on eur-lex.europa.eu before quoting the article in an internal procedure.

In plain terms: an obligation of means, not of result. The Commission says much the same in its questions and answers: there is no one-size-fits-all approach, and no mandatory training is imposed. It also recalls that, under Article 4(2), the Commission and the Member States support providers and deployers — small and medium-sized enterprises in particular — and that examples of compliance are published on the single information platform. A living repository of AI literacy practices is online; the Commission specifies that replicating a practice from that repository does not automatically grant a presumption of compliance with Article 4.

On penalties, let us be precise, because this is where the most damaged figures circulate. Article 99 sets ceilings: up to 35 million euros or 7% of total worldwide annual turnover for the prohibited practices of Article 5; up to 15 million or 3% for the provisions listed in paragraph 4, which cover among others the obligations of deployers of high-risk systems (Article 26) and the transparency obligations (Article 50); up to 7.5 million or 1% for incorrect information supplied to the authorities. For SMEs and start-ups, the lower of the two amounts applies. Article 4 does not appear in the list in paragraph 4. That does not mean “no risk”: Member States set their own penalty regime, and this is a point of interpretation to confirm with your legal counsel.

One last point on the timeline, and it changes how urgent the subject is. Since 2 August 2026, the Commission and the national authorities have been enforcing the regulation, and the transparency obligations of Article 50 apply: telling a person they are dealing with an AI system, marking generated or altered content. Systems already on the market at that date have until 2 December 2026 for the marking and detection obligations of Article 50(2). The rules on high-risk systems, for their part, have been postponed: 2 December 2027 for the Annex III use cases, 2 August 2028 for AI embedded in regulated products.

In Belgium, the FPS Economy ran an information campaign aimed at SMEs and start-ups, built around a guide that answers four questions: what the AI Act is, whether the company is concerned, which risk category its AI system falls into, and which obligations apply. Check its current availability on economie.fgov.be.

Why training remains your subject, even softened

The softening of Article 4 does you no favour if you read it as an exemption. It simply moves the question: what exposes you is no longer a level of literacy that was not reached, it is what your teams do with the tool without knowing what they are doing.

Three examples I keep seeing. A salesperson pastes a complete client contract into a consumer assistant to get a summary. Someone in administration sends a client a generated text citing a legal reference that does not exist. A customer service team puts a conversational agent online without saying it is an AI — and there, we are no longer in Article 4, we are in Article 50, which does appear in the list of provisions carrying penalties.

Training is not a box to tick for a regulator. It is what separates those three situations from their harmless version.

The method: three levels, one record

  1. An inventory of uses, on one page. Two columns: which tool, for which task. Count the tools nobody approved, they exist all the same. Without that inventory, you are training blind.
  2. Level 1, everyone, 90 minutes. What a large language model is and why it sometimes invents. What you never paste into a chat window: health data, personnel files, contracts, credentials. The obligation to tell the person on the other side that they are talking to an AI. Who to call when in doubt.
  3. Level 2, intensive users, half a day. Checking an output before sending it: figures, names, references, dates. Recognising the typical failures — invented fact, misplaced number, ignored instruction, plausible and wrong answer. Building a small set of control questions for one’s own use.
  4. Level 3, whoever decides. The roles in the regulation: provider and deployer are not the same thing. The timeline above. The clauses to read in a supplier contract. The threshold beyond which you ask for legal advice.
  5. The record. Two pages are enough: who attended what, when, with the session outline and the list of uses covered. Since the obligation is one of means, it is the evidence of means that counts.
  6. The review, once a year. Uses move faster than training. Take the inventory again, add what has appeared, remove what has been dropped.

What gets in the way in practice

Generic training. A one-hour talk on “AI and its challenges” produces nothing verifiable. What works is built on your own cases: your quotes, your shared inbox, your documents. Same core content, in-house examples.

Shadow AI. In every organisation where I have looked, the official inventory is shorter than reality. Personal subscriptions used at work are the most common source of data going to the wrong place. One declared and framed use beats three invisible ones.

Confusing training with an exam. Since the text does not require guaranteeing a level for any given individual, building a graded assessment is misplaced effort. Train, record, and put the remaining energy into checking outputs — that is where mistakes cost money.

Where to start, in two weeks

Week 1: the one-page inventory, and the list of what is never sent to an external tool, approved by management. Week 2: the level 1 session for the whole team, on your own examples, and the tracking table. You will then have the hardest thing to produce after the fact: a date, an attendance list and a content outline.

The rest — levels 2 and 3, the annual review — is built on that base. And if a client asks you the question again, you will have something other than an intention to show.


Benjamin de Bruijne is an independent digital and AI consultant in Belgium. He automates repetitive work for SMEs and puts AI where it removes a specific chore. Astérisme: AI inside your processes.