What Article 4 asks for, and what it doesn't
One plain sentence first: this is not legal advice. It describes what a company of 20 to 250 people can sensibly do and where the obligation stops.
It has applied since February 2025. Providers and deployers must take measures on AI literacy for their staff and for anyone else operating or using AI systems on their behalf. The article says the measures should account for people's technical knowledge, experience, education and training, and the context the system is used in. It also points at the people the system is used on. A tool that summarises internal meeting notes needs less preparation than one that answers customer questions.
Since July 2026 the wording is softer. The EU rewrote Article 4 in the so-called Digital Omnibus, EU Regulation 2026/1744, which entered into force on 27 July 2026. The old text asked companies to ensure, to their best extent, a sufficient level of AI literacy. The new text asks them to take measures to support the development of AI literacy, and adds that this does not require guaranteeing any specific level of AI literacy of any individual. That turns an obligation of result into an obligation of effort. The obligation itself is still there.
No certificate, no AI officer. The European Commission's Q&A on Article 4 says there is no need for a certificate and that organisations can keep an internal record of trainings. It also says Article 4 mandates no specific governance structure, so you don't have to appoint an AI officer. An external auditor appears nowhere in the article. Anyone selling you an expensive certification programme "for Article 4" is selling something the law doesn't ask for.
No fine of its own, but a regulator. Article 4 is not among the provisions listed in the AI Act's penalty article, Article 99. In Germany, the national AI supervision law, the KI-Marktüberwachungsgesetz, has been in force since 29 July 2026 and makes the Bundesnetzagentur, the federal network agency, the market surveillance authority. Its fine provisions don't list Article 4 either. That's no reason to skip it. The Commission gives August 2026 as the start of supervision and enforcement, and when an AI tool causes real trouble, a wrong answer sent to a customer or personnel data pasted into the wrong account, the first question anyone asks is whether staff were trained.
Take stock of tools and roles
Training comes third. First you need to know what's running and who uses it.
Ask, don't hunt. Don't send IT through browser histories. Send a short survey with four questions. Which AI tools do you use for work? What for? On which account, private or company? What kind of data goes in? Say plainly that nobody gets in trouble for their answer. Otherwise you'll get the answers people think you want, which is a different list.
Count the hidden ones. Plenty of AI features aren't labelled AI. The meeting summary in the video call tool, the text suggestions in the CRM, the receipt recognition in the accounting software. Go through your software list and ask of each program whether a recent update taught it to write, summarise or read documents.
Sort roles by risk. Article 4 explicitly ties the measures to people's background and the context of use, which means you don't owe everyone the same course. A 50-person company usually ends up with four groups. Occasional users rephrase an email or ask for a formula now and then. Users who handle customer or personnel data sit in sales, customer service, HR and accounting. A small group sets systems up or lets them influence decisions about people. And management decides what gets bought.
Hiring is its own category. AI systems used to analyse and filter job applications or evaluate candidates are listed in Annex III of the AI Act, the list of high-risk areas. For those, Article 4 is nowhere near enough. If you use or plan something like that, treat it as a separate project, not a training question.
What you end up with is a table with six columns: tool, purpose, department, users, account type, kind of data. For 50 people it fits on one page.
One short training per role
A session that treats everyone the same bores half the room and loses the other half. An hour covers occasional users. People working with customer data need longer and need examples from their own week. Run it in-house, with your own tools and your own cases. In a German company with a works council, the training usually belongs in the works agreement for the tool anyway, and what else that agreement covers is in the article on AI and works councils in Germany.
What the tool is bad at. The Commission's Q&A names hallucination as a risk that people using generative AI should be told about. Show it live. Ask the tool to quote a standard, cite a section of a law or give a small customer's annual revenue. The answer sounds confident and is sometimes invented. Two minutes of that stick longer than a slide with the word hallucination on it.
What must not go in. Write the list with examples from your own company. Customer names with their terms, personnel files, health data, passwords, unpublished figures. Add the rule on accounts, meaning which account is allowed for which data. A free private account and a company account with a contract are different things, and everyone needs to know which one they're typing into. To turn the list into a rule for everyone, start from the AI usage policy template.
Who owns the output. The person who sends it. Obvious, and still worth saying out loud once, ideally with a customer service example where an invented delivery date turns into a real complaint.
Who to ask. A name, not a department. Someone unsure whether a spreadsheet can go into the tool should know who to call. Otherwise they decide alone, and they'll usually decide in favour of finishing faster. If nobody in-house has time to be that name, a fractional Chief AI Officer can take it on.
Write it down and repeat it
The record is the part most companies forget, and it's the easiest one.
A list of who got what. Name, role, date, topic, length, who ran it. A spreadsheet will do. File the material next to it: the slides, the list of forbidden data, the account rule. The Commission says organisations can keep an internal record like this. It doesn't say whether that will satisfy a supervisory authority in a given case. For what goes into each column, and a two-table template to copy, see AI literacy training records under the EU AI Act.
Repeat when something changes. A new tool, a new feature in an old one, a new role, a new hire. Each is a reason for a refresher. Put the basic session into onboarding and you no longer have to remember it.
Review the table once a year. Take the inventory from the start and ask each department what's been added. It takes half an hour, and it shows you where people already use AI well and where they're still working around you with private accounts.
Article 4 isn't a monster from Brussels. Read properly, it asks you to know which AI runs in your company and to make sure your people can tell when not to trust it. You needed that before the regulation existed. Now you have a reason to get it done this week.