AI literacy under the EU AI Act: what Article 4 asks of employers since July 2026

If your company gives staff an AI tool, the EU AI Act expects you to help them understand it. That duty sits in Article 4. It has applied since 2 February 2025, and in July 2026 the EU rewrote it. Some teams read the new text as the end of the obligation. It is not. The duty is lighter, and it is still there.

What the text said before

The original Article 4 asked providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and the other people who operate or use those systems on their behalf. A deployer is any organisation that uses an AI system under its own authority. A company that rolls out Microsoft 365 Copilot or a chatbot to its employees is a deployer.

The word that worried legal teams was "ensure". It sounded like an obligation of result: you had to reach a level, and you could be asked to prove it.

What changed in July 2026

The Digital Omnibus on AI, published as Regulation (EU) 2026/1744 on 24 July 2026 and in force since 27 July, replaced Article 4 in full. Three points matter for an employer.

1. The duty is now to support, not to ensure. Providers and deployers must take measures to support the development of AI literacy among their staff and the people using the systems on their behalf. The measures still have to fit people’s technical knowledge, experience, education and training, the context of use, and the people the systems are used on.
2. No set level is required. The new text says the obligation does not require anyone to guarantee a specific level of AI literacy for any individual. You do not have to certify each employee.
3. The public side has a role too. The Commission and the member states must support companies' efforts, SMEs in particular, and publish practical examples of compliance. The European AI Board will adopt recommendations on common objectives.

The same regulation pushed most high-risk obligations back to December 2027 and August 2028. Article 4 was not delayed. It applies today, to every organisation that deploys AI, whatever the risk level of the system.

What did not change

You still have to do something, and you still have to be able to show what you did. "Take measures" is an obligation of means. If a regulator, a works council or a court asks how your staff were prepared to use the tool, "we bought licences" is not an answer.

The duty also keeps its scope. It covers staff, and also contractors, temps and anyone else who uses the system on your behalf. It covers every AI system you deploy, not only the high-risk ones.

One more point, often missed: Article 99 of the AI Act, which sets fines, does not list Article 4 among the obligations with a dedicated fine. That is not a reason to skip it. A gap in literacy shows up elsewhere, for example in a data incident under the GDPR or in a dispute about how an AI output was used in a decision about a person.

A plan that holds up

The text asks for measures adapted to people and context. That rules out one generic e-learning module for everyone. Here is what I would put in place for a Copilot or ChatGPT rollout.

1. Map who uses what. List the AI systems in use, and the roles that use them. A buyer who drafts contracts with Copilot and an HR officer who screens CVs with an AI tool do not need the same training.
2. Write the rules in plain language. One page per tool: what data can go in, what must be checked before an output is used, who to ask. If a non-specialist cannot apply it on a Tuesday afternoon, rewrite it.
3. Train on real tasks. Short sessions built on the work people do, with the limits shown live: a wrong answer, an invented source, a summary that drops the key point.
4. Keep help close after launch. Questions come in weeks two to six, when real cases appear. A network of colleagues who answer them is worth more than a second course. I wrote about why pilots stall without it.
5. Keep a record. Who was trained, on what, when, and which rules they received. Dates and attendance are enough. This is your evidence.
6. Review twice a year. Tools change every quarter. So do the ways people use them.

Literacy is not adoption

Article 4 is about understanding: knowing what the tool does, where it fails, and what you are responsible for. It says nothing about whether people use the tool, or whether it helps. Those two questions are your business case, and the law will not answer them for you. A training register proves you complied. It does not prove the licences pay off. For that, you need to measure usage, not licences.

Done well, the two go together. The sessions that make people literate are the same ones that show them where the tool saves time. That is the part I work on with teams.

Sources: Regulation (EU) 2024/1689 (AI Act), Articles 3, 4 and 99; Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1. This article is general information, not legal advice.