The law first, because it changed two months ago
Article 4 of the EU AI Act is the one about AI literacy. Its text was replaced in July 2026.
| What it used to say (2 February 2025 – 26 July 2026) | What is in force today (since 27 July 2026) |
|---|---|
| providers and deployers shall take measures to ensure a sufficient level of AI literacy among their staff | providers and deployers shall take measures to support the development of AI literacy |
The change came with Regulation (EU) 2026/1744 (adopted 8 July 2026, published 24 July, in force 27 July). The difference in one sentence: you do not have to prove that your colleagues know something, you have to prove that you did something so they would learn it.
The dates that matter:
| Event | Date |
|---|---|
| Article 4 starts to apply | 2 February 2025 |
| The amended text enters into force | 27 July 2026 |
| Supervision and enforcement start | 2 August 2026 |
So the obligation has been alive for a year and a half, but supervising and sanctioning it has only been possible for about a month. That is why it is topical now.
Who it applies to, and what the sanction is
It applies to you too if anyone in your company uses AI in their professional work. Whoever uses an AI system under their own responsibility is a "deployer" in the language of the Regulation — a small business that drafts its emails with a language model counts as a deployer. There is no company-size threshold, and the circle of people is wider than your employees: on the Commission's reading it covers subcontractors and outside parties acting on your behalf.
The sanction, on the other hand, is where a lot of people get it wrong. Article 4 appears in none of the AI Act's fining bands. The Regulation knows three of them (35 million euros or 7% of global turnover for prohibited practices; 15 million euros or 3% for another group of obligations; 7.5 million euros or 1% for misleading the authorities) — AI literacy sits in none. Enforcement rests with the national market surveillance authority, with proportionate sanctions weighed case by case.
And one important concession: for SMEs the lower of the two figures applies, not the higher one.
Which authority that is differs from member state to member state, and it is rarely the data protection authority, which is what most people assume. The EU deadline for the designation was 2 August 2025, and not everyone met it. Open the one that applies to you:
Hungary — market surveillance sits with the minister for enterprise development
The designation is in Section 1 of Government Decree 344/2025 (X. 31.): the market surveillance authority and the single point of contact went to the minister responsible for enterprise development, and the notifying authority role to the National Accreditation Authority.
Under the Hungarian implementing act — Act LXXV of 2025 — that authority may impose an administrative fine, establish that a use was unlawful, and order the restoration of a lawful state; summary proceedings are excluded. So if anyone does ask about AI literacy, there is an authority, and it has instruments.
Spain — an AI agency and guidance exist, a national sanction does not
Spain was the first in the EU to set up a dedicated AI supervisory agency, AESIA, back in 2023. The formal designation under the Regulation, however, is still not finished: it sits in the bill that was still in the amendment stage in Congress in September 2026.
The Spanish bill does not sanction AI literacy either — Article 4 does not appear anywhere in its list of infringements. Which confirms the conclusion above: there is no standalone fine for skipping the training, at EU level or nationally.
What does exist, and is useful: in December 2025 AESIA published 16 non-binding guides to compliance with the AI Act — among them transparency, risk management and human oversight — with checklists. None of them is specifically about AI literacy, but their table of contents makes a good skeleton for an in-house course.
What to take away from this: anyone talking about million-euro EU fines for skipping AI training is wrong.
What it does NOT require — this part is reassuring
According to the Commission's own guidance:
- there is no single recipe, and no mandatory courses;
- you do not have to measure your colleagues' level of knowledge;
- you do not have to guarantee any particular level of literacy for anyone;
- there is no compulsory curriculum, certificate or exam.
There is one exception: for high-risk AI systems, the training tied to human oversight remains mandatory. This does not touch most small businesses — and the rules for high-risk systems have in any case slipped to 2 December 2027 and 2 August 2028.
What the Commission does ask for is four steps: a general understanding of what AI is and how it works; clarity about your role (provider or deployer); an assessment of the risks of the systems you actually use; and literacy initiatives tailored to your colleagues' level and to your sector.
There is an EU collection of practices too (the "living repository" launched in February 2025 and refreshed in August 2026, with more than 40 organisational initiatives) — but its legal status is stated openly: copying a practice from it does not in itself create a presumption of compliance.
Who should you train? The surprising answer
Not the enthusiastic technical one. This is the most counter-intuitive and the best-evidenced point in the whole plan.
In a measurement covering 5,179 customer support agents, an AI assistant raised the number of issues solved per hour by 14% on average — but it brought +34% for novice and low-skilled staff, and practically zero for the experienced ones. The authors' explanation: the model spreads the proven practice of the more skilled colleagues.
Whoever tried it on their own will use it on their own anyway. The gain sits with the worst-performing third — and that is also the best professional argument for holding any training at all.
There is a business confirmation as well: according to a US company survey from August 2026, 61% of service firms use AI and only 4% of them let anyone go because of it — while more than a third of the AI-using companies retrained their existing staff. The authors write literally that existing workers are far more likely to be retrained than replaced.
The four sessions
Each is 90 minutes, each has homework, and each produces one output you then actually use. The material is built from your own company's documents, not from generic prompt tips — that is exactly the mechanism behind the measurement above.
| Session | What it covers | Homework | What is left afterwards |
|---|---|---|---|
| 1. What it can and cannot do | How the machine guesses the next word, why it makes things up, what it is reliable at | Everyone brings one case from their own work where the machine got it wrong, and one where it was good | A shared list of what we use it for and what we do not |
| 2. Your own process | One real task worked through together, on the team's own material | Everyone does the same task alone and saves what they got | The company's first prompt collection, built from the best answers |
| 3. What not to hand over | Where the company is liable for the output, what must never be typed in, when human approval is needed | Everyone reviews three of their own outputs and marks what they would release without approval | The filled-in version of the one-page policy |
| 4. Measurement and routine | What we look at weekly, what is a signal and what is evidence | Two weeks of logged use in the chosen process | The table of weekly numbers and the name of the person responsible |
The output of the second session is the most valuable one: the prompt collection gets good because the best person on the team puts their solution into it. Building one is covered in the team prompt library, step by step. For the output of the third session there is a ready skeleton too — the one-page AI policy gives you the wording to fill in.
The third session is the most important one legally, because it is the one that actually delivers the third of Article 4's four steps (risk assessment), and because the question of liability is settled there.
What counts as a "measurable outcome"
Here you have to be honest: I could not find a randomised or quasi-experimental study measuring whether a four-session corporate AI course improves business results. There is evidence for the target group of the training and for the mechanism; for the format itself there is none. The four sessions are a proposal, then, not a proven recipe.
What you can measure — and what satisfies both the "measurable outcome" expectation and the law's documentation need:
- An attendance sheet for the four sessions — this is the evidence that you did something.
- Use: in the two weeks after a session, how many people used the company tool at least once in the chosen process. This comes from logs, not from a questionnaire.
- The three numbers of the process: volume, turnaround time, error rate — the same three you took the baseline with.
What not to measure: knowledge level, with a test. The Commission does not ask for it either, and the result says nothing about whether it gets used in the work.
And what you should really never do: ask the team how much time they saved. According to one rigorous experiment, people overestimate AI's effect on their own time by 40 percentage points on average — the question gives a wrong answer without any bad intent.
Three common misunderstandings
"An email saying you may use it from now on is enough." That is not a measure supporting the development of AI literacy, and it cannot be documented as one either.
"We have to buy an external course." There is no compulsory curriculum and no compulsory certificate. Four sessions built from your own material are better anyway, by the measurements, because they spread your company's own proven practice.
"Training is for the young ones." The measurement shows the opposite: the gain arises with the inexperienced, regardless of age — with an experienced professional it is roughly zero.
What to take away
The legal obligation is milder than most people believe, but it has been enforceable for a month now, and the authority behind it is a national one that differs from country to country. What it asks for, you deliver with a four-session in-house course and an attendance sheet.
The business case for the training is stronger than the legal one: the measurable gain arises with the worst-performing third, and the mechanism is that the best colleagues' answers spread. If the material for the four sessions is built from your own company's best answers, that is exactly what you are doing.
Where all this fits into the rollout schedule is shown by the three-month AI plan, month by month; and for the rest of the law — what applies and what slipped — the current state of the EU AI Act gives you a dated overview.