The short answer
Generative AI training in a company needs three things: content matched to each role, practice with your approved tools, and a simple list of who attended what and when. As a planning guide, we suggest one shared day for teams that work with AI regularly and two to three hours of foundations for occasional users of ChatGPT, Copilot or a similar assistant. IT, developers and managers need additional modules.
Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures that support the AI literacy of their staff and of other people who work with those systems on their behalf. Under Article 3, a deployer is anyone who uses an AI system under their own authority, except in the course of a personal, non-professional activity. In its questions and answers, the European Commission gives the example of a company whose employees use ChatGPT to write advertising copy or translate text. Article 4 applies there too: employees should be informed about the specific risks, such as hallucination (made-up facts).
Two topics have their own guides. “Plan AI team training around real work” covers use cases and learning outcomes. “EU AI Act Article 4: planning AI literacy at work” covers what Article 4 asks for and what to note for each measure.
Which role needs which depth
Article 4 specifies neither a format nor a number of hours. The Commission’s questions and answers say there is no one-size-fits-all approach and that no mandatory training is imposed. Simply asking staff to read the instructions for use may, in its view, be ineffective in many cases. For its own staff, the Commission distinguishes learning packages for generalists, managers and developers. The time figures below are our planning guide, not a legal requirement.
IT is on the list for a specific reason. According to Microsoft, Copilot in Microsoft 365 only surfaces organisational data for which the individual user has at least view permissions. If folders in SharePoint or Teams are shared too widely, Copilot can surface their content to everyone with access. Tidy up permissions before the training day, not after.
- Occasional users in every department: two to three hours on how the tools work, typical errors, approved tools and data rules.
- Frequent users in marketing, sales, HR or finance: a full day with exercises on task types from their own area.
- Managers: foundations plus about two hours on approvals, responsibility for outputs and where the team does not use AI.
- IT and administration: licences, permissions, logs and retention, plus an access-rights review for Copilot in Microsoft 365.
- Developers: coding assistants such as GitHub Copilot, reviewing suggestions, handling credentials and licence questions in code.
- Data protection, information security and HR: involve them early, because they help write the rules taught in the session.
Agenda for a one-day team session
One day is enough if the group practises with at most two tools and all exercises use fictional or approved material. Allow about six working hours excluding breaks and keep at least half for hands-on practice. The schedule below is a suggestion for a mixed group from business departments.
Two things do not belong in this day: deciding which tool to introduce, and debating the company policy. Settle both beforehand, or they eat into practice time. Two half-days are an equally valid option: between sessions, participants can try something at work and bring back real questions.
- 09:00 to 09:45: how a language model produces answers and why its mistakes sound convincing.
- 09:45 to 10:30: house rules on approved tools, permitted data and who decides in case of doubt.
- 10:45 to 12:15: describing tasks and comparing outputs, in pairs, using three prepared practice cases.
- 13:15 to 14:30: checking outputs by comparing figures, names and sources with the original material and marking errors.
- 14:45 to 16:00: work by department, for example Copilot in meetings and email or ChatGPT for draft proposals.
- 16:00 to 16:30: collect open questions, agree next steps and confirm attendance.
Data protection rules for prompts
As soon as a prompt contains information about an identified or identifiable person, personal data is being processed. The General Data Protection Regulation (GDPR) defines the term broadly: Article 4 lists names, identification numbers, location data and online identifiers as examples. Data must be limited to what is necessary for the purpose (Article 5). Processing specially protected information, such as data on health, religious beliefs or trade union membership, is prohibited in principle, with exceptions (Article 9). If a service provider processes data on your behalf, a contract must bind it to your documented instructions (Article 28).
One distinction matters most in daily work: personal account or company licence. Whether inputs are stored, used to train models or processed outside the EU depends on the contract and settings of the service. Clarify this before the training with whoever is responsible for data protection, and turn the result into a one-page traffic-light scheme. The scheme below is an example; your organisation’s own rules are binding.
If a planned type of processing is likely to result in a high risk to the rights and freedoms of the people concerned, Article 35 requires a data protection impact assessment beforehand. That question arises when a tool is introduced or put to a new use, not with each prompt, so it belongs in the manager module.
- Green: public information, your own drafts without personal data and fictional examples.
- Amber: internal documents and customer data only in the approved company tool and only for the intended purpose.
- Red: never passwords or access keys. Health data, job applications and personnel files only after explicit approval from the responsible function.
- Always: replace names with placeholders where the task allows, and check outputs about a person before passing them on.
Recording participation with Article 4 in mind
Article 4 mentions neither a certificate nor a particular form of documentation. The version that entered into force on 27 July 2026 also states that providers and deployers need not guarantee a specific level of AI literacy for any individual. The Commission’s questions and answers add that no certificate is needed, that organisations can keep an internal record of training and other initiatives, and that there is no obligation to measure employees’ AI knowledge.
A written record is still worth keeping. According to the Commission, national market surveillance authorities are responsible for supervision and enforcement, and those rules have applied since 2 August 2026. Measures nobody can evidence are hard to explain later.
Keep the record lean. Article 4 does not ask for test scores or assessments of individuals, and the less personal data the list holds, the easier it is to maintain. Decide who may see it and how long you keep it. An attendance confirmation from the training provider is useful evidence but does not prove legal compliance. The list below is our working aid, not an official form.
- Attendance list for each session with name, role and department, confirmed by the trainer.
- The agenda as delivered, with date, duration and tools practised.
- The version of the internal AI policy taught in the session.
- Absentees, with the date of their catch-up session.
- Contractors or freelancers who work with the tools on your behalf.
- A rule for new starters and a trigger for refreshers, such as a new tool or a changed policy.
How to brief a training provider
A good brief fits on one page and answers what the training provider cannot guess. Do not send real customer data or confidential documents, not even as practice material. Describe task types instead and have the provider build fictional cases for you to approve before the session.
Ask the training provider for a written answer to three questions. Which content is planned for which role? What is explicitly not covered, for example legal advice or configuring your systems? Who will deliver the session? NextNowa offers training on this topic on enquiry.
- Group: number of people, roles, language and prior knowledge in two sentences.
- Tools: which products and licences you have, and whether all participants have access on the day.
- Rules: your AI policy or the traffic-light scheme, so the session follows your everyday rules.
- Tasks: three to five typical tasks per department, without real data.
- Format: on site or live online, one day or two half-days, recording yes or no.
- Materials: timed agenda, reusable practice material, attendance list and attendance confirmations.
Sources and review date
Official information checked on 17 November 2026. General information, not legal advice.
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 3(4) and original Article 4 ↗
- EUR-Lex: Regulation (EU) 2026/1744, amended Article 4 ↗
- European Commission: AI Literacy Questions & Answers ↗
- European Commission: AI talent, skills and literacy ↗
- EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 4, 5, 9, 28 and 35 ↗
- Microsoft Learn: Data, Privacy, and Security for Microsoft Copilot ↗

