Employee AI training is a documented briefing that explains to staff which AI tools they may use at work, which data they may enter into them, and how to handle the output. The training record is the written evidence of who was trained, when, and on what content.
The training itself is rarely the problem. Most companies I talk to have run an hour on ChatGPT or Copilot at some point. What is missing is almost always the second half. No attendance list, no agenda, no link to the company’s own acceptable use policy. And a training session you cannot prove happened does very little for you in front of a supervisory authority.
This post does not go back over whether the obligation exists. I covered that in the article on the AI literacy obligation under the EU AI Act. This one is about implementation: what changed in 2026, what a training session should cover, and what goes into the record.
Is AI training for employees mandatory?
Yes, from two directions. The EU AI Act requires measures on AI literacy, and the GDPR requires that employees who process personal data know what they are allowed to do.
EU AI Act, Article 4. Deployers of AI systems, meaning any company whose staff use AI tools for work, must take measures on the AI literacy of their personnel. Company size makes no difference.
GDPR. The word “AI” appears nowhere in it, but the obligation follows anyway. Art. 32(4) GDPR requires that people acting under the controller’s authority process personal data only on the controller’s instructions. An instruction nobody knows about is not an instruction. Art. 24 and Art. 5(2) require you to put appropriate organisational measures in place and to be able to demonstrate compliance with them. And Art. 39(1)(b) explicitly lists awareness-raising and training of staff among the tasks of the data protection officer.
For professionals bound by secrecy in Germany, Section 203 of the Criminal Code (StGB) adds another layer. If you involve external service providers or staff in work on client data, you must make sure they are bound to confidentiality. What that means for AI tools in particular is covered in the post on client data and Section 203 StGB.
What changed in Article 4 in 2026?
Amending Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, rewrote Article 4 at the end of July 2026. Before, providers and deployers had to “ensure” that their staff had a sufficient level of AI literacy. Now they must take measures to support the development of that literacy. The new text makes clear that no specific level of literacy has to be guaranteed for any individual.
In legal terms, this is a shift from an obligation of result to an obligation of effort. That is real relief. You are no longer on the hook for every clerk reaching a specific level of understanding.
I do not read it as making training optional, though. Three reasons:
- The obligation stayed, only its standard moved. You meet an obligation of effort by taking measures. And documentation is how you show that you did. The record becomes more important, not less.
- The GDPR did not change. Art. 32(4) and the accountability principle apply exactly as before. If your team uses AI tools with personal data, you still have to tell them what is allowed.
- The practical benefit is the same. The typical risk with AI tools is not the technology. It is a well-meant input into the wrong tool. No configuration fixes that. Only knowledge does.
What should AI training cover?
Neither the GDPR nor the AI Act prescribes a curriculum. The purpose of the rules still makes it fairly clear what belongs in one. I work with five building blocks.
1. Which tools are approved. Specifically, by name and licence tier. “Copilot is allowed” is not enough if the free web version and the M365 licence are both available in the company and only one of them is approved.
2. Which data may go into which tool. This is the most important part. It works best with examples from the team’s own daily work: this client enquiry may be summarised, this health detail never. If you have defined data classes in your AI acceptable use policy, the training translates those classes into concrete cases.
3. How to handle AI output. Hallucinations, outdated legal positions, plausible but wrong numbers. Who checks the result before it goes to a client? Anyone who passes on AI-generated text needs to be able to answer that.
4. What happens after a mistake. Anyone who accidentally pasted data into the wrong tool needs to know whom to tell. A reported mistake is manageable, a hidden one is not. People only report when they know reporting is expected and will not be punished.
5. Role-specific content. A managing partner who decides on new tools needs different content from a trainee using Copilot for emails. For decision-makers, data processing agreements, data location, and the data protection impact assessment belong in the session. For the wider team, building blocks 1 to 4 are enough.
One hour for the whole team plus an extra session for decision-makers is a realistic scope for a small company. Longer is not automatically better. Training built around the tools people actually use sticks better than a general talk about artificial intelligence.
What belongs in the training record?
A training record is a document that captures, for each session, who attended, what was covered, and which version of the internal rules it was based on. An auditor should be able to see from it that your measure fits how AI is actually used in your company.
These details should be in it:
- Date, duration, and format (in person, online, e-learning)
- Name of the trainer and their role (internal, data protection officer, external consultant)
- Attendance list with name, function, and confirmation of attendance
- Agenda or content summary, at least the building blocks covered
- Tools covered, by name and licence tier
- Version of the acceptable use policy the training refers to
- Acknowledgement of the policy, by signature in smaller teams, otherwise by digital confirmation
- Optionally a short knowledge check, for example five questions on concrete cases
- Where the training materials are stored, so the slides from back then can still be found later
- Next planned date or the trigger that will prompt a repeat
For a team of 15, that is one spreadsheet and one folder. Nothing more is needed. The policy version is the item most often forgotten and the most valuable. It links the training to your documentation. When the policy changes, you see immediately who was trained on the old version.
Store the record wherever your other data protection material lives. Your record of processing activities can refer to it directly in the description of technical and organisational measures.
How often do you need to train?
There is no statutory interval. What makes sense is training tied to triggers, not to the calendar alone:
- When new employees join, as a fixed part of onboarding, before access to AI tools is switched on
- When a new tool or a new licence tier is introduced
- When the acceptable use policy changes in a way that affects daily behaviour
- After an incident, as a short refresher for the teams involved
On top of that, I recommend an annual 30-minute refresher. That is not a legal requirement, it is experience. AI tools change faster than any other software in the office. What was true a year ago often no longer is.
The most common mistakes
Mistake 1: One generic video for everyone. A purchased e-learning module on “AI in the workplace” explains what a language model is. It does not explain which tools are approved at your company and which data may go into them. Useful as a supplement, not as the only measure.
Mistake 2: Training without a policy. Without written rules, the training passes on opinions instead of instructions. Policy first, then training.
Mistake 3: Ignoring shadow AI. Training that covers only officially approved tools misses reality if half the team uses ChatGPT privately. Address it openly. More on that in the post on employees using shadow AI.
Mistake 4: Building the record after the fact. An attendance list reconstructed from memory months later convinces no one. Capture the record on the day of the training.
The next step
If you do only one thing today: set up the spreadsheet for your training record, even if the next session is a month away. Enter whatever AI-related briefings have already happened and note openly where details are missing. In ten minutes you will see how big the gap really is.
If you need training for your team that is built around your tools and your policy and delivers the record along with it, get in touch.
Jose Lugo is a CISSP-certified AI compliance consultant. He advises law firms, tax advisors, and financial service providers in Germany on GDPR-compliant AI adoption, including staff training and documentation.