Nobody in the company wants another AI training session? If all a training session leaves behind is an attendance sheet, it is hard to establish afterwards which rules apply to the tools people actually use. The good news: Article 4 of the AI Act (the EU regulation on artificial intelligence) does not require a course. It requires measures — and measures carried out around a specific system leave a more specific trace than a certificate. This article shows five such traces and a record template that collects them. Preparing this documentation is part of our technical AI Act readiness sprint.
A caveat up front: we write down what the provision says and what the European Commission says — with dates and links — and what to build and record in practice. Whether your company’s measures are enough in its particular situation is for a lawyer to assess, not an implementer.
And one sentence about us, so it is clear where this article comes from: at ploch.ai we do not sell AI training or workshops as a standalone service. We practise working with a system only as a stage of an implementation — the AI Act readiness sprint, too, includes an Article 4 workshop on the tools the company actually uses. The reasons are on the page about what we don’t do.
What Article 4 says after 27 July 2026
Since 27 July 2026, following the amendment made by Regulation (EU) 2026/1744 (the Digital Omnibus on AI), Article 4(1) reads:
“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.”
A “deployer”, in the regulation’s language, is a company that uses an AI system in its business — including a ready-made one bought from a supplier. The duty has applied since 2 February 2025; the July 2026 change replaced the earlier wording about ensuring “a sufficient level of AI literacy” with a duty to take measures. The full AI Act timeline, and what the Polish statute and KRiBSI change, are covered in the AI Act in Poland — here we deal only with how to carry out and record those measures.
What Article 4 does not require — and what “without a course” does not mean
The provision says nothing about hours, an exam, a certificate or a dedicated AI officer. The European Commission confirms this in its AI literacy Q&A (last updated 27 July 2026). Four of its answers matter most here:
- Certificate: “There is no need for a certificate. Organisations can keep an internal record of trainings and/or other guiding initiatives.”
- Measuring knowledge: Article 4 “does not entail an obligation to measure the knowledge of AI of employees”.
- Format: “There is no one size fit all when it comes to AI literacy and no strict requirements or mandatory trainings are imposed”; what is needed depends on the context.
- Everyday tools too: a company whose employees use ChatGPT to write advertising copy or translate text is covered by Article 4 — “they should be informed about the specific risks, for example hallucination”, meaning information the model has made up but that sounds convincing.
That is not, however, permission to do nothing. In the same Q&A the Commission writes that “in many cases, simply relying on the AI systems’ instructions for use or asking the staff to read them might be ineffective”. An email saying “please read the policy” may therefore not be enough. That is why we recommend walking through the instruction on a case from the company’s own work, with the depth matched to people’s experience and to the risk of the use. “Without a course” means this: instead of generic AI training, short, specific measures around the systems people actually use, and a record that they took place.
Who it covers
Article 4 covers staff and “other persons” operating AI systems on the company’s behalf. The Commission explains that these are persons broadly under the organisation’s remit — for example a contractor, a service provider or even a client. In practice it is worth including in the record the external accounting firm working in your system and the agency that runs the company account in a content-generation tool.
The second place that is easy to forget is AI built into tools the company already has: summaries in email, suggestions in the customer relationship management system (CRM), a description generator in the online shop. Before you document anything, you need a list of those tools and the people who use them. How to compile it, including when staff have picked some of the tools themselves without the company knowing (so-called shadow AI), is covered in our article on shadow AI at work.
Three roles, different depths of preparation
Article 4 requires you to take knowledge, experience and the context of use into account. The simplest way to translate that is by role. The split below is our practice, not a requirement of the provision — the Commission only confirms that different levels of detail for different groups are acceptable.
| Role | Example | What is worth passing on to this person |
|---|---|---|
| Uses AI for drafts | a salesperson writes a first version of a quote in a chat tool | that the model can invent a fact; what must not be pasted in; that they, not the tool, are responsible for the text |
| Operates a process that acts on customers | the person running automatic replies to enquiries | how the system makes its decision, where it most often goes wrong, how to stop it and whom to report an error to |
| Approves AI output | an accountant approves the system’s proposed booking of an invoice | which errors a tired eye lets through, what to check every time, when to reject even if the result “looks fine” |
The approving role needs particular attention, because the person approving is the last safeguard against an error. How to design that human approval of AI output so that it does not turn into reflex clicking is covered in our article on human in the loop.
Five pieces of evidence that working with the system produces
This is the core. The documents below are a set we recommend and that is worth adapting to your own company — not a checklist whose completion settles compliance with the provision. Each of them serves to pass on knowledge about a specific system and at the same time leaves a trace of that measure. On their own they document measures; they do not settle whether the company meets the Article 4 duty. All of them are prepared separately for each AI system, not for each department.
1. A one-page work instruction. For each system: what it does, what data it works on, where it goes wrong (with two or three real examples from your company), what must not be entered into it, when to stop it and whom to report a problem to. The same page serves as the plan for a short walk-through with the team: we suggest half an hour at the screen with a real case instead of a generic presentation. Half an hour is our recommendation, not a standard.
2. An item on the onboarding checklist. If the company has a list of things to do with a new employee, add: “has read the work instruction for system X and gone through one real case with a mentor”. That way the measures do not stop at the people who happened to be in the company on the day of the rollout.
3. A warning inside the tool. A short text visible where the person works: a banner above the input field, the description in the company assistant, a message before approval. For example: “Always check amounts and account numbers against the source. Do not paste customer data into tools that are not on this list.” The wording and the date it was introduced are also a trace of a measure — and the text itself reminds people of the rules every time they use the tool, not once a year.
4. A register entry: who was briefed, and when. If you keep a register of AI systems, add a pointer to the record from the next section. If you do not keep one yet, this record is a good first step towards it.
5. A periodic “what went wrong” review. Once a quarter (again, a recommendation) the team reviews the system’s errors from recent months, adds new examples to the work instruction and notes who took part in the review. A review like this helps people learn from their own cases rather than from someone else’s examples.
| Evidence | What it documents | Who keeps it | When to update |
|---|---|---|---|
| Work instruction | which rules, limitations and error examples the company has written down | the process owner | after every change to the system and after the error review |
| Onboarding checklist item | that a new person was shown the system before starting to work with it | HR or the line manager | for every new person |
| In-tool warning | which rules are visible where the work happens, and since when | the tool administrator | when the instruction changes |
| Register entry | to whom, for which system, when and which material was presented | the person who keeps the register | after every walk-through |
| Error review | that the company analyses its own cases, and who took part | the process owner | every quarter (recommendation) |
Record template: one table to copy
The Commission speaks of “an internal record of trainings and/or other guiding initiatives”. An ordinary spreadsheet is enough. We suggest six columns — this is our proposal, not a format the provision requires:
| Person or role | AI system | What was presented (with version) | When | Who ran it | Next review |
|---|---|---|---|---|---|
| Customer service team — participant list OBS-01 | Automatic replies to enquiries | Work instruction v2; three errors from October; how to stop sending | 2026-10-14 | the process owner | 2027-01-15 |
| New person in accounting (staff number on list KS-07) | Invoice booking proposals | Work instruction v1; one case at the screen with a mentor | 2026-10-20 | a mentor from the department | 2027-01-15 |
| Marketing agency (external account, participant list MKT-02) | Product description generator | Work instruction v1; content-labelling rules v1 | 2026-10-22 | the person responsible for marketing | 2027-04-22 |
The rows are examples, and the dates, list references and roles are made up. Three rules make a record like this usable later: enter a specific system, not “AI in general”; in the “What was presented” column give the document version, so that it is possible to check exactly what someone saw; and always fill in the next review — the date helps plan further measures, and an empty cell shows the gap straight away.
In the “Person or role” column, enter only as much as you actually need to evidence the measures. For a group, a pointer to a separately kept participant list is enough; names do not need to be copied into every sheet. These are participants’ personal data, so it is worth agreeing their scope and retention period with whoever is responsible for data protection in the company.
The implementation as a source of evidence
A well-run AI implementation produces material for this record as a by-product. In how to implement AI in a company we describe the step in which the system runs for a while alongside the existing process and people record every error it makes. Here we are only interested in what that leaves behind for Article 4. Three things:
- The error list from the parallel run becomes the “where it goes wrong” section of the work instruction. The record notes who helped compile it and when.
- The approval record — who approved the system’s output during the trial period and how much of it they rejected — shows which real cases the approvers worked on.
- The warning text written from those errors goes into the tool as evidence number 3.
That is why this kind of trace says more than an attendance sheet: it shows not only that someone was in the room, but also which system, which cases and which material they worked with. With us, the team walk-through is a stage of AI implementation or AI automation, on the tools that actually run in the company.
When a training offer quotes the old wording
A simple test for any “Article 4 training” offer: if it quotes the provision with words about ensuring “a sufficient level of AI literacy”, it is quoting the wording in force before 27 July 2026. That alone does not disqualify the training, but it is worth asking what will be left on the company’s side afterwards: will the participants practise on your systems and your cases, or on generic examples, and will you get material after the training that can be entered into the record from the previous section?
When a course really does make sense
“Without a course” is not an absolute rule. A suitable training course is worth considering when:
- the company deploys, or plans to deploy, a high-risk AI system. For such systems Article 26(2) provides that human oversight is assigned to natural persons “who have the necessary competence, training and authority, as well as the necessary support”. The Commission stresses that this obligation remained in place after the July 2026 change. As a general rule it applies from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems; the regulation also has exceptions and transitional rules. Whether and from when this applies to your system is for a lawyer to establish — the dates are covered in the AI Act in Poland;
- someone with no AI experience at all is to approve output that reaches customers — one page and half an hour are too little as a starting point;
- many people join within a short time and walking through each system separately is no longer feasible;
- the team works with many different AI tools and needs common foundations before work instructions start to make sense.
In those cases, simply add the training to the same record — as one of its rows, alongside the instructions, warnings and reviews.
What next
If the company already runs several AI tools and has no record at all, start with the one system that affects customers most: a work instruction, a walk-through, the first row in the table. The rest comes with each further system. If you want to do it together with the register of AI systems, content labelling and documentation, we describe that on the page about our technical AI Act readiness sprint.
Sources, as at 12 October 2026
- Regulation (EU) 2024/1689 (AI Act), consolidated text of 27 July 2026 — Article 3(4) and (56), Article 4, Article 26(2) and Article 113; read 12 October 2026.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — the new wording of Article 4; read 12 October 2026.
- European Commission, “AI Literacy — Questions & Answers” — last updated 27 July 2026; the page’s introduction still describes Article 4 in the earlier “sufficient level” wording, while the answers already reflect the change; read 12 October 2026.
- European Commission, living repository of AI literacy practices — examples from other organisations; the Commission notes that replicating them does not automatically grant a presumption of compliance with Article 4; read 12 October 2026.