AI literacy at work: what Article 4 of the AI Act asks of employers now
· By Bee Flow · 9 min read
Article 4 of the AI Act was rewritten in July 2026: no level, no certificate, but still a duty. What it asks of employers now, who supervises it, the dates that matter, and what a company of 20 to 200 people can do this quarter.

On 20 October 2026 the European Commission updated its questions and answers on AI literacy again. Put that page next to an article from 2025 and one line stands out: no specific or "sufficient" level is mandated any more. Article 4 of the AI Act was rewritten over the summer, and texts written before then still use the old wording.
This piece is about that one duty only: making sure the people in your company who work with AI know enough about it. For the wider picture of what the AI Act asks of a small business, see what the EU AI Act asks of a small company that uses AI chat. This is not legal advice. We report what the sources say, with links, so you can check them or hand them to your lawyer.
What Article 4 says now
The duty has applied since 2 February 2025, the date on which the general provisions and the prohibited practices took effect, according to the timeline on the Commission's AI Act Service Desk.
The original text asked providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff. Since 27 July 2026 the first paragraph reads, in the consolidated text on the Service Desk:
Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used. This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
Two paragraphs were added. The Commission and the Member States must support companies, SMEs in particular, in meeting the duty, and the Commission publishes practical examples for that. The European AI Board adopts recommendations with common objectives. The official text of the regulation is on EUR-Lex.
The change in one sentence: from "ensure a sufficient level" to "take measures that support learning". The duty was not deleted. In its note on the entry into force the Commission writes that the company-level requirement was simplified and that the Commission and the Member States take a stronger role.
Who it covers
Article 4 addresses providers (whoever develops an AI system or puts it on the market) and deployers (whoever uses one under their own authority). The second word covers almost every company whose staff use an AI tool at work.
The Commission's Q&A gives the example of a company whose staff use ChatGPT for advertising copy or translations. Those people, the Commission says, should be informed about the specific risks, "for example hallucination": a model that makes something up and writes it down convincingly.
It is not limited to employees. The Commission speaks of persons broadly under the organisation's remit, such as a contractor, a service provider or a client working with an AI system on your behalf.
What "enough" means now
The Commission's Q&A puts it like this:
- no specific or "sufficient" level is mandated
- there are no strict requirements and no mandatory trainings
- there is no need for a certificate
- you do not have to measure your employees' AI knowledge
- you can keep an internal record of trainings or other guidance you gave
It does name three things any approach should touch: a general understanding of AI within the organisation, the role your company plays (provider or deployer), and the risks of the AI systems you use. For high-risk AI, the separate duty in Article 26 to train the people who oversee the system stays in place.
The Dutch data protection authority (Autoriteit Persoonsgegevens, AP) published a second guide in October 2025, following Get started with AI literacy from January 2025. It was written before the amendment and still speaks of a sufficient level. The method it describes is still useful: four steps (identify, set goals, carry out, evaluate) that you keep repeating. The AP also writes that there is no obligation to measure the level per individual or to obtain AI literacy certifications, and that you should keep internal documentation of your approach and progress. We did not find a newer version written after the amendment.
Supervision and fines
According to the Commission's Q&A, Article 4 is supervised by national market surveillance authorities, not by the EU AI Office. They start supervising and enforcing on 2 August 2026. They could impose penalties and other enforcement measures, the Commission says, and any sanction must be proportionate and decided case by case.
Article 99 of the regulation sets maximum fines for, among other things, the prohibited practices, the obligations of providers and deployers of high-risk AI, and the transparency rules. Article 4 is not on that list. How a Member State sanctions a breach of Article 4 therefore depends on its national rules. The AP warns in its guide that a demonstrable lack of AI literacy can weigh in the assessment of an incident.
In the Netherlands the national implementing act is not in force yet. On 20 April 2026 the government put the draft AI Act implementation act out for public consultation until 1 June 2026. The proposal: existing regulators supervise AI in their own domain, the AP covers areas without a clear regulator, and the AP and the Dutch Authority for Digital Infrastructure (RDI) coordinate. The letter to parliament of the same day says the government accepts that the act will probably enter into force later than the AI Act requires. The same letter says the obligations apply in the meantime, because the regulation is directly applicable. It does not say which regulator will check Article 4. We would not guess.
The omnibus: what was adopted and what is still a proposal
On 19 November 2025 the Commission proposed two "omnibus" packages. They get mixed up all the time.
The AI omnibus is law. According to the European Parliament's Legislative Train, a deal was reached on 7 May 2026, Parliament approved it on 16 June, the Council adopted it on 29 June and Regulation (EU) 2026/1744 was published in the Official Journal on 24 July. It entered into force on 27 July 2026. This is the amendment that rewrote Article 4 and postponed the high-risk rules.
The omnibus on data and the GDPR (COM(2025) 837) is still a proposal. The Legislative Train records that the Council agreed its negotiating mandate on 21 October 2026. Parliament has no position yet. Until a text is adopted, nothing in the GDPR changes.
The dates in one table
| Date | What applies |
|---|---|
| 2 February 2025 | AI literacy (Article 4) and the first prohibited practices |
| 2 August 2025 | Rules for providers of general-purpose AI models (GPAI) |
| 2 August 2026 | Transparency rules (Article 50), and the start of supervision and enforcement, including of Article 4 |
| 2 December 2026 | New ban on AI that generates non-consensual sexual deepfakes or child sexual abuse material, and the end of a transition period under Article 50(2) for certain generative AI already on the market before 2 August 2026 |
| 2 December 2027 | Rules for high-risk AI in Annex III, such as AI for recruitment and staff evaluation |
| 2 August 2028 | Rules for high-risk AI in Annex I products, such as machinery and lifts |
Sources: the AI Act Service Desk timeline and the Commission's AI Act page.
When AI is used on your own staff
AI literacy is about the people who use AI. A second question is what applies when AI is about your employees. Three sets of rules meet there.
Annex III, point 4, of the AI Act lists as high-risk, among others, AI that analyses and filters job applications, AI that allocates tasks based on behaviour or personal traits, and AI that monitors and evaluates the performance and behaviour of workers (Service Desk). The heavy rules apply to those uses from 2 December 2027. An employer that deploys such a system must then, under Article 26(7), inform workers' representatives and the affected workers beforehand.
In the Netherlands the works council already has a say. Article 27(1) of the Works Councils Act (WOR) gives the works council a right of consent over arrangements for processing and protecting employees' personal data, and over facilities "aimed at or suitable for" observing or checking the attendance, behaviour or performance of staff. A works council is mandatory from 50 people. A smaller company with a staff representation body falls under Article 35c, which gives that body a shorter list of subjects, and these two are not on it in the text of the act.
And the GDPR applies, omnibus or not. If AI will process employees' personal data, a DPIA is often the first step. How to keep one small and useful is in a DPIA for AI chat in a small company.

What a company of 20 to 200 people can do this quarter
The law asks for measures, not an exam. What follows is our reading of what is reasonable for a company this size, based on the Commission's Q&A and the AP guide. Take a made-up example: Installatiebedrijf Voorbeeld, an installation firm with 45 staff.
Start with a list. Which AI tools are in use, by whom, for what? Ask the teams too, because the tools nobody officially introduced are often the interesting ones. After a week Installatiebedrijf Voorbeeld has six on its list, from the AI chat to a feature in the accounting software that reads receipts. Note for each tool whether personal data goes into it.
Then write a short policy, one or two pages. Which tools are allowed, what they are not for, what never goes in (customer data in a free account, for instance), and that anything going to a customer or the outside world is read by a person first. Name who people can ask. In Bee Flow, the Privacy Shield backs up the rule about what never goes in: by default it holds back a chat message with personal data before it reaches the model, in every licence including the free one (privacy and security).
Then explain it by role. Everyone gets the basics: what a language model does, why it can make things up, what the policy says. An hour is enough. People who use AI for work that affects others (the planner who assigns jobs, HR, whoever writes quotes) get a second conversation about their own tool and its risks. At Installatiebedrijf Voorbeeld that is an extra half hour for four people.
Record who had what. A spreadsheet with name, date, topic and the policy version will do. The Commission mentions an internal record as an option and sets no format. In the Bee Flow Learning Center an administrator sees per person which lessons and courses are done, which can be part of that record. Put the repeat in the calendar, for example once a year and whenever a new tool arrives. New staff get the basics in their first week.
That gives you something to show when someone asks. What we would skip: a paid certificate because a vendor says it is required. According to the Commission it is not.
How Bee Flow supports this
Bee Flow gives the steps above a place in the tool your team already works in. The duty itself stays with you as the employer, and no software fulfils it for you. The Learning Center has short lessons inside Bee Flow, for example on writing a good prompt and on what the Privacy Shield does to a message before it leaves. An administrator can require someone to finish a course before they create or change agents, knowledge bases or automations, while using them stays open to everyone. In the Compliance Center you confirm once a year that you took AI literacy measures, with a link to your own policy, and Bee Flow warns when that confirmation is more than a year old. Which licence includes these features is on the plans page, and if you want to see how they fit your team, get in touch.