In July the managing director of an engineering firm receives a quote: AI Act training, two hours per employee, certificate included for the file. In a reply underneath, the privacy officer asks whether this is now mandatory. The honest answer: the obligation is real, but the course in that form is not required.
Article 4 of the AI Act has applied since 2 February 2025. It requires providers and deployers (in plain terms, whoever builds an AI system and whoever uses one under their own responsibility) to take measures aimed at a sufficient level of AI literacy among their staff and others who use AI systems on their behalf. What counts as sufficient depends on knowledge, experience, education and the context of use. The article mentions no course, no certificate and no number of hours.
What does Article 4 actually say?
The text is short. Providers and deployers take measures to secure, to their best extent, a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of AI systems on their behalf. In doing so they take into account those people's technical knowledge, experience, education and training, the context in which the systems are to be used and the persons on whom they are to be used. That is the substance of the provision in Regulation (EU) 2024/1689.
The same regulation defines AI literacy as the skills, knowledge and understanding needed to make an informed deployment of AI systems and to be aware of the opportunities, the risks and the possible harm. So it is not about being able to code. It is about knowing what you are using, what it cannot do and when not to take an output at face value.
Two terms deserve attention. A "deployer" can be a hospital running a triage system, but equally an office where people use a chatbot to draft a proposal. And "on their behalf" reaches beyond the payroll: contractors and suppliers who work with your AI systems count too.
The Digital Omnibus of July 2026 recast the provision as a best-efforts obligation: take measures that support the development of AI literacy. What that does and does not change is set out in the AI Act timeline after the omnibus. For the practical question of what to do, it makes little difference: the obligation remains, and it still depends on context.
Why does the law not prescribe a course?
Because the same provision has to work for a medical software manufacturer and for an accountancy practice with eight people. A single standard would be too light for one and too heavy for the other. EU lawmakers chose an open-ended standard: the organisation decides what is sufficient and must be able to explain that choice.
The European Commission confirmed this in spring 2025 in its questions and answers on AI literacy. Neither a certificate nor a formal test of staff knowledge is required, and Brussels will not be issuing a set training package. An organisation can rely on its own record of the measures it has taken. We could not open that page ourselves while writing this piece and have relied on several legal summaries of it; read the text before you base policy on it.
The Commission also maintains a collection of practices, a "living repository" of what organisations have put in place themselves, from formal courses to team discussions about what does and does not belong in a prompt. It is a useful source of ideas, with one caveat the Commission adds itself: adopting an example from the list does not in itself mean you comply with Article 4.
What does "sufficient" mean for each role?
The question is not "how many hours of training?" but "what does someone in this role need to understand to use AI responsibly?" The answer differs by role, and those differences are exactly what you write down.
| Role | What the person needs to be able to do | What you record |
|---|---|---|
| Uses a chatbot for writing | Knows which data does not belong in it, which tools are allowed and that answers need checking | The briefing this group received and the decision on each tool |
| Uses AI in decisions about people (recruitment, customer onboarding, care) | Understands how the system reaches an output, what it misses and when to override it | The instructions for each system and who was trained on them |
| Buys or assesses AI tools (IT, procurement, privacy) | Can assess the vendor, its training policy, where it is based and the risk | The questionnaire and the register of assessed tools |
| Sets the policy (management, board) | Knows the status of each tool and where the line between allowed and not allowed lies, and understands its own responsibility | The policy, the decisions and who took them |
For most organisations the first row carries the most weight. That is where the largest group sits, and where the briefing is simplest: this is allowed and this is not, and if in doubt, this is who you ask. How the AI Act and the GDPR fit together here is covered in what the AI Act and the GDPR require of businesses.
Who supervises it?
Article 4 has applied since February 2025, but supervision came later. National authorities have had supervisory powers since 2 August 2026, when most of the regulation also began to apply. The regulation sets no penalty specifically for Article 4; it is up to member states to set sanctions in national law.
In the Netherlands this is done through the AI Act Implementation Act (Uitvoeringswet AI-verordening). The bill went out for public consultation in April 2026 and divides supervision among existing regulators, with a coordinating role for the Dutch Data Protection Authority (AP) and the Dutch Authority for Digital Infrastructure (RDI). Exactly which regulator will supervise Article 4 in the Netherlands, and with what powers, has therefore not yet been settled. Note that in your own policy too: it is no reason to wait, but it is a reason not to claim certainty that does not yet exist.
Incidentally, the question "can you show it?" rarely comes from a regulator first. It comes from a customer in a tender, from the auditor, from the works council or after a data breach in which an employee pasted a file into a chatbot. What matters then is not whether there was a course, but whether you can show what the organisation had decided and whether that decision reached the employee.
What does this look like in an ordinary organisation?
In spring an accountancy firm with 120 staff sends everyone on an AI course. Copilot is allowed; client data does not belong in free tools; if in doubt, ask the privacy officer. Attendance is 94 per cent and the list goes in the file.
An inventory in the summer shows that ChatGPT, Perplexity and a meeting transcription tool are in use alongside Copilot. No decision was ever taken on the transcription tool, and two teams have been using it in client meetings for months. That is not an incident, nor proof that the course failed. It shows that the course was a single moment and that work carried on, with tools that did not exist on the day or that nobody recognised as AI.
The firm then does two things. It decides on the transcription tool and records the decision in the register. And it makes sure that anyone who opens the tool sees that decision.
Why knowledge is no guarantee of behaviour
That second step is the part of AI literacy most often missing. A course changes what people know. What they do on a Thursday afternoon with a deadline looming is shaped more by habit than by knowledge. Why that is, and what a course can and cannot do, is explained in our article on why awareness training does not change behaviour.
For Article 4, passing on knowledge is only the first half. The second half is making sure the organisation's decision is there when someone opens an AI tool. Not as a block, but as a reminder: this is what was decided about this tool and this is the approved alternative, and any sensitive data in your text is highlighted before you send. Set that up, and you no longer have to hope that people remember a briefing from three months ago.
It also gives you something for the record. Every notice can be counted along with what the person did next: opened the alternative or carried on anyway. Every highlight can be counted with its outcome: masked, removed or sent regardless. As counts, without saying who it was. That is the difference between "we ran a course" and "we know what has happened since".
