Take a company with ten employees. Three of them use ChatGPT every week. The sales assistant rewords offers, the purchasing manager translates his exchanges with an Italian supplier, and the managing director prepares meeting notes. Since the spring, an n8n workflow has also been reading messages sent to the contact address and drafting replies with a language model. One employee reviews every draft before it goes out.
Nobody has had any training. The director has heard that the European AI regulation was amended this summer, and wonders whether he needs a certificate, an "AI officer" or an online course.
The company is covered by the AI Act AI literacy obligation in Article 4. It must take measures so that these people understand the AI they use. But the "mandatory AI training" that is often described does not match the text. No level is required and no certificate is expected. This article covers three things: what Article 4 has said since 27 July 2026, what the European Commission expects, and how to put together a short, documented plan role by role.
What Article 4 says since the Digital Omnibus
Regulation (EU) 2026/1744 of 8 July 2026, known as the "Digital Omnibus on AI", amends the AI Act. Published in the Official Journal of the EU on 24 July 2026, it entered into force on 27 July 2026, as set out in its final article and as noted by the law firm Cuatrecasas.
The regulation replaces Article 4 entirely. The new paragraph 1 requires providers and deployers of AI systems to take measures to support the development of AI literacy among their staff and other persons using those systems on their behalf. These measures take into account each person's technical knowledge, experience, education and training, as well as the context of use and the people the system affects. The last sentence states that the obligation does not require any particular level to be ensured for any individual.
According to the updated official FAQ, the obligation still applies, but a "sufficient" level is no longer required. Recital 8 of the regulation explains why. A strict obligation did not suit everyone and created an extra burden, particularly for small companies.
Paragraph 2 asks the Commission and the Member States to support companies, especially SMEs. To that end, the Commission is to publish practical examples of how to comply on the single information platform. Paragraph 3 requires the European AI Board to adopt recommendations that take European competence frameworks into account and set common objectives.
On timing, Article 4 has applied since 2 February 2025, like the rest of Chapter I of the AI Act (Article 113 as amended by Regulation 2026/1744). The FAQ states that the supervision and enforcement rules have applied since 2 August 2026.
Who is covered: deployers, providers and contractors
Article 4 targets providers, which develop AI systems, and deployers, which use systems built by others. A company using ChatGPT is a deployer. The obligation applies to both roles.
The FAQ covers exactly the situation from the introduction. A company whose employees use ChatGPT to write advertising copy or to translate is covered. Those employees need to be informed of the specific risks, such as hallucinations.
According to the same FAQ, the "other persons" are not employees. They are people acting within the company's remit, such as a subcontractor, a service provider or even a client. Suppose a freelancer maintains your n8n workflows, or an agency writes your content with AI. The Commission says these people need the AI skills appropriate to the task, like your own staff. Depending on the system and its level of risk, a contractual clause may be justified.
What the obligation does not require
The Commission's FAQ, consulted on 18 November 2026, rules out several common assumptions. These are the Commission's answers, not the text of the regulation:
- No certificate. No certificate is needed. The Commission says organisations can document their actions in an internal record of training sessions or other guidance initiatives. The record tracks real measures, it does not replace them.
- No knowledge testing. Article 4 does not require companies to test their employees' level.
- No AI officer or committee. No specific governance structure is mandated for Article 4.
- No set format. The Commission writes that there is no one-size-fits-all approach and that no strict requirements or mandatory training are imposed.
That last point is not settled for good. According to the FAQ, training requirements depend on the context and potentially on the future recommendations of the European AI Board (paragraph 3), which may refine what is expected.
There are two limits. The Commission warns that in many cases relying on instructions for use, or asking staff to read them, may be ineffective. And copying a practice from its repository of examples does not, for now, create a presumption of compliance.
Meeting the AI literacy obligation: a plan by role
The Commission organises its internal training paths by audience: generalists, managers and developers. In a company of 5 to 50 people, I recommend splitting the plan by what each person does with AI, not by seniority. The same person can belong to two groups.
| Role | Who, in the example | What they need to know | Suggested format |
|---|---|---|---|
| Everyday user | The three ChatGPT users | Hallucinations, data not to paste, approved tools, checking outputs | One one-hour session and a one-page policy |
| Reviewer of an agent's output | The employee who reviews the drafts | What the workflow reads, typical errors, when to reject, who to report to | A session on the real workflow, using failed cases |
| Automation builder | The person or contractor who runs n8n | Data sent to the model provider, prompt injection, logs, tests, limits | A written technical review, updated with each new workflow |
Everyday users. The essentials fit in four points:
- A model produces plausible text, not necessarily accurate text.
- Figures, names and references must be checked.
- Some data (customer files, contracts, health data) must not be pasted into a tool the company has not approved.
- Everyone knows the list of approved tools.
People who review what an agent produces. The FAQ treats this case separately. The person in the loop needs skills targeted at the system they work with. To get there, I recommend that they know which sources the agent consults and which mistakes it makes most often, and that they be able to reject or correct an output without having to justify themselves. The main risk is approving out of habit, without reading. For where this review fits in a workflow, see where to put human approval and the example of AI email triage with n8n.
People who build the automations. I recommend that they know which data goes to which model provider (see open vs proprietary models), that an email read by an agent can contain malicious instructions, and how to log, test and limit what the agent can do. If a contractor does this work, ask them for a short note covering these points for each workflow they deliver. To judge what can stay in-house, see what you can do yourself with n8n.
The managing director decides which tools are approved and keeps the log.
The minimum content, according to the Commission
The FAQ lists four steps that any programme should cover:
- General understanding. Everyone knows what AI is, how it works, which tools the company uses, and the opportunities and dangers involved.
- The company's role. For each tool, the company knows whether it develops the AI system or uses one built by another vendor.
- Risks of the systems in use. For each tool: what staff need to know, which risks, how to reduce them.
- Tailoring. The content takes into account each person's level, the sector, the use case and the people affected by the output (customers, candidates, suppliers).
The Commission encourages companies to add legal and ethical aspects, including the basics of the AI Act. A company just starting out can prepare this checklist while choosing a first AI project.
Documenting it: the internal training log
A shared spreadsheet works as a medium. I recommend one row per action, with these columns:
- date of the action;
- who attended and their role (user, reviewer, builder, contractor);
- tools and workflows covered, with their version or model where relevant;
- risks addressed (hallucinations, confidential data, agent errors, prompt injection);
- format and material (session, policy, written note) and who led it;
- date of the next review.
A fictional example row: "12 October 2026, Claire (reviewer), n8n workflow replying to contact requests. Risks: invented delivery dates, wrong tone, another customer's data quoted. 45-minute session on ten real drafts. Review in April 2027."
Update the log whenever something changes: a new tool, workflow, model, team member or contractor, or an incident. Keep the list of approved tools next to it.
Enforcement and penalties: why an incident matters
According to the Commission's FAQ, consulted on 18 November 2026, Article 4 has been enforced since 2 August 2026 by national market surveillance authorities, not by the AI Office. Penalties rest on national laws that Member States were due to adopt by 2 August 2025. This article does not verify the French penalty regime or which authority France has designated for Article 4, so do not assume either.
The Commission stresses proportionality. Any penalty depends on the case, including the seriousness of the breach and whether it was intentional or negligent. It becomes more likely if an incident results from a lack of training and guidance.
This is where the log matters. Imagine the workflow from the example sends a customer an invented delivery date, and the reviewer had never been told the agent could make things up. Without a written record, the company will struggle to show it had taken any measures.
On civil liability, the same FAQ states that the AI Act does not create a right to compensation. A person who suffers harm has to act under national law.
The rest of the AI Act timeline after the Omnibus
The Omnibus also postponed the obligations for high-risk systems. According to the official text and Cuatrecasas, they will apply from 2 December 2027 for Annex III systems (biometrics, employment, education, access to essential services). For AI built into regulated products (Annex I), they will apply from 2 August 2028.
The FAQ points out that deployers of high-risk systems must still train the staff responsible for human oversight (Article 26), a Chapter III obligation that follows these new dates. If your company screens job applications with AI, check now whether that use falls under Annex III.
For an overview, see what applies to companies under the EU AI Act.
A short plan rather than a course catalogue
For the company in the introduction, Article 4 comes down to four things: a one-page policy, two short sessions on its real tools, a note on the n8n workflow and an up-to-date log. Kirako also offers training for teams. The text leaves companies free to choose the format, but it expects them to take real measures and to be able to show what they have done.
Sources
Consulted on 18 November 2026:
- European Commission: AI Literacy, Questions & Answers, English reference version, updated after the Omnibus. The French version is a machine translation (eTranslation) whose header still mentions a "sufficient level".
- Official Journal of the EU: Regulation (EU) 2026/1744 of 8 July 2026, recital 8, new Article 4, amendment of Article 113, final article on entry into force
- Cuatrecasas: Digital Omnibus on AI has been published, 24 July 2026
Also available: Read in French