What Article 4 requires, and what it leaves open
This is practical guidance, not legal advice.
The duty has applied since 2 February 2025. The AI Act is Regulation (EU) 2024/1689, and its official text is published on EUR-Lex, the EU's legal database. In its original wording, Article 4 required providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy. That covers their staff and other persons dealing with the operation and use of AI systems on their behalf. The European Commission says this includes contractors and service providers. The measures should account for people's technical knowledge, experience, education and training, and the context the systems are used in. If your staff use ChatGPT, Copilot or the AI features in your CRM, your company is a deployer and the article applies to you.
Since July 2026 the wording is softer. The Digital Omnibus, Regulation (EU) 2026/1744, amended Article 4. Companies now have to take measures to support the development of AI literacy, and the text no longer sets a specific or "sufficient" level for any individual. The obligation itself stays. That helps with record-keeping, because what you document now is the company's measures. Whether each individual reaches a given level is no longer part of the text.
No certificate and no set format. Neither the regulation nor the Commission says what a record must look like. In its AI literacy questions and answers, the Commission states that no certificate is needed and that organisations can keep an internal record of trainings or other guiding initiatives. It also warns that relying on a system's instructions for use, or asking staff to read them, may not be effective.
Penalties come from national law. Article 99 of the AI Act sets fine ranges for a list of specific infringements, and Article 4 is not on that list. That doesn't mean a breach has no consequences. According to the Commission's Q&A, national market surveillance authorities can impose penalties and other enforcement measures for infringements of Article 4, based on national laws that member states were due to adopt by 2 August 2025. The Commission adds that any sanction has to be proportionate to the individual case, and that one becomes more likely when an incident can be traced back to a lack of training and guidance. A record that existed before the incident answers that question better than one put together afterwards.
What goes into a credible record
A record is credible when an outsider can see who was prepared for what, and when it matches the tools that are actually in use. "All employees received AI training" fails that test. It doesn't say who learned what, or which tool the training was about, and Article 4 explicitly asks you to take both into account.
Who. Name, role and department. Include outside people who work with your AI systems on your behalf, for example the agency writing LinkedIn posts from your company account, or the freelance bookkeeper with access to the receipt-scanning feature.
Which systems, in which role. Name the actual tool and the account type, "ChatGPT Team, company account" rather than "AI". Then the role the person uses it in. Someone who rephrases the odd email needs different preparation from someone who answers customer requests with it or configures it for everyone else.
Content. The topics in a few words, plus the version of the material. Without a version number you can't tell later who knows the old data rule and who knows the new one.
Date and duration. A date and a length in minutes. The length shows whether the training fits the role. An hour is enough for occasional users. People handling customer data need more.
Format and trainer. In person, video call, e-learning or a walkthrough at the desk, and who ran it. No one requires a signature. The attendee list from the video call or the completion report from the e-learning tool is enough to show that someone took part.
Materials. The slides, the list of data that must not go in and the account rule, filed in one fixed place with a version number and date. The record points to them and doesn't need to repeat them.
Refresh cycle. When the next session is due, as a date or as a trigger. "Annually" is a date. "When a new tool is approved" is a trigger. Write both down.
A template in two tables
If you've run health and safety briefings, you know the pattern. The trap is writing the training content out again for every person. After the second update nobody can follow it. Two tables work better. The first describes the modules, the second records who completed which module.
The modules, shown here for a company with company ChatGPT accounts and an AI feature in its CRM:
| Module | For | AI systems | Content | Length | Format | Materials | Refresh |
|---|
| M1 Basics | everyone using AI | ChatGPT Team, Copilot in Outlook | spotting invented facts, forbidden data, account rule, who to ask | 60 min | in person with live demo | slides v1.2, data list v1.1 | yearly and with each new tool |
| M2 Customer data | sales, customer service | ChatGPT Team, CRM text suggestions | checking answers before sending, customer data, own cases | 90 min | workshop | case collection v1.0 | when the CRM adds an AI feature |
| M3 Administration | IT | ChatGPT Team admin console, CRM | permissions, chat retention, approving new features | 2 h | walkthrough | checklist v1.0 | with each major update |
Attendance, with made-up example rows:
| Name | Role | Modules | Date | Format | Trained by | Evidence | Next refresh |
|---|
| A. Example | Sales back office | M1 v1.2, M2 v1.0 | 14 Oct 2026 | in person | Head of IT | attendee list | Oct 2027 |
| B. Example | Accounting | M1 v1.2 | 14 Oct 2026 | in person | Head of IT | attendee list | Oct 2027 |
| Agency, 2 people | LinkedIn posts | M1 v1.2 | 21 Oct 2026 | video call | Head of Marketing | email confirmation | Oct 2027 |
For 50 people that comes to a few dozen rows, and a spreadsheet handles it. If you already run a learning platform, export the attendance from there and keep the module table next to it.
The list holds names, which makes it personal data. Store it where your other training records live, with the same access rights.
Keeping it current when tools and roles change
A record from last year proves last year's state. AI tools change faster than most software in the building, and new features arrive through updates that nobody ordered.
Attach it to processes you already run. The basic module belongs on the onboarding checklist, next to the laptop and the email account. HR already handles role changes, so add one line there asking whether the new role needs another module. Then nothing depends on someone remembering.
Use tool approval as the trigger. Clause 3 of the AI usage policy template keeps every new tool away from company data until someone approves it. If your policy works the same way, that approval is the natural moment. No new tool goes live until someone has asked which module changes and who needs a refresher.
Track versions. When the list of forbidden data changes, it gets a new number. The attendance table then shows who still knows v1.1. A small change often needs no more than fifteen minutes in the next team meeting and a new entry in the table.
Reconcile every six months. Put your list of AI tools in use next to the module table. A tool that appears in the list but in no module is a gap. So is a department that uses AI and has nobody in the attendance table. The check takes about half an hour.
Give it an owner. One named person keeps the list. In a 50-person company that's a few hours a year, often for the head of IT or the managing director's assistant.
The record won't prove that everyone in your company uses AI safely, and since July the amended Article 4 no longer asks you to. It proves that you know which tools are running and what the people using them have learned. Keep the two tables up to date and the answer is ready before anyone asks for it.