AI isn't a yes-or-no question. It's a "for which task" question. For drafting and translating modules, AI cuts hours of work. For grading people and for compliance content, it adds risk you have to cover by hand.
Where AI clearly helps
Three tasks where AI shrinks the work from days to minutes:
- Drafting quiz questions. From a 10-page document, AI proposes 15 questions in 2 minutes. HR reviews them and drops half.
- Translation and localisation. One module from EN to DE/FR/HR without a translation agency. The cost drops from a few hundred EUR to a few minutes.
- Compressing content into microlearning. AI turns a 40-minute webinar into five 5-minute units with the key points.
These tasks share one trait: a human reviews the output before it reaches employees. AI speeds up the first draft here. It doesn't make decisions.
The "cost of verification" rule
Apply a simple test before every AI task:
- If verifying the output is faster than writing from scratch → use AI.
- If you must check the output sentence by sentence against a source → AI saves you nothing.
A quiz draft passes the test. A legal paragraph on GDPR often fails it.
When not to use AI
Three areas where careless AI use creates more work or legal exposure:
- Compliance and legal content. AI "hallucinates" articles and dates. One wrong reference in a safety module is the employer's liability, not the AI vendor's.
- Employee personal data. Feeding names, scores, or health data into an external AI model is a data transfer. Under GDPR you need a legal basis and a processing agreement.
- Assessment that decides about people. Here the law applies, not just good practice.
The line the EU AI Act draws
The EU AI Act (Regulation (EU) 2024/1689) classifies some AI uses as "high-risk". Obligations for that group apply from 2 August 2026.
Annex III lists as high-risk, among others:
- AI that evaluates learning outcomes and steers the learning process
- AI that decides admission to education or training
- AI for monitoring during tests, and AI in employment decisions (hiring, evaluation, termination)
A proposed "Digital Omnibus" would push the deadline for some employment systems to December 2027. Until it is formally adopted, 2 August 2026 stands. So AI that automatically grades and gates promotion based on a quiz falls under the heaviest duties: human oversight, logging, and discrimination monitoring.
Transparency applies separately (Article 50, also from 2 August 2026): if an employee chats with an AI tutor, tell them. Label synthetic video and audio.
Worked example: GDPR onboarding for 50 employees
A company of 50 employees is building a data-protection onboarding module.
- Use AI to draft 12 quiz questions from the internal policy. HR drops 5 and edits 3.
- Use AI to translate the module from EN to DE for international colleagues.
- Don't use AI for the text on data-subject rights. The DPO writes it, because a wrong reference is the company's liability.
- Don't use AI to auto-decide who "fails" onboarding. HR sets the threshold and a human reviews the result.
Result: module prep drops from three days to half a day. The legal part stays verified and hand-written.
Worked example: safety training in manufacturing (200 people)
A manufacturer with 200 employees repeats mandatory safety training each year.
- Use AI to localise the module into the three languages of seasonal workers.
- Use AI to summarise a 60-page regulation into five 4-minute units.
- Don't use AI to state specific exposure limit values. A safety engineer checks those against the current regulation.
- Don't use AI as the sole grader of the practical part. A supervisor rates the live performance; AI only logs completion.
Here AI saves translation and editing time. The engineer keeps responsibility for the limit values, because an error there is a safety risk.
Which tools to pick
The market of AI training tools falls into three groups. Judge each by the same cost-of-verification test:
- Generic chat models. Universal for drafts and translations. Downside: you can't share data without a processing agreement.
- Authoring tools with built-in AI. They generate quizzes and scenarios inside the LMS. Upside: content stays in the system and the log is automatic.
- AI graders and tutors. The most capable, but most likely to land in the high-risk category once they decide about people.
Rule: a generic model is enough for first drafts. For personal data or scoring, pick a tool that documents the processing and leaves the final call to a human.
How to organise this in practice
To keep AI an assistant and not a risk, set three rules:
- Every AI draft passes through a human before assignment to employees.
- No personal data goes into an external model without a legal basis; check how your LMS vendor processes data (ai-data-processing).
- AI does not decide on admission, promotion, or termination without human oversight.
It also helps to log which part of a module AI produced. That record is the basis for compliance after 2026.
In short
- Use AI for drafts, translations, and compression — where a human verifies the output fast.
- Don't use AI for legal references, personal data, and decisions about people.
- From 2 August 2026, the EU AI Act classifies AI grading of learning outcomes and employment decisions as high-risk.
- For compliance content, an error is the employer's liability — verification isn't optional.
AI works best as an accelerator for the first draft, not as the final authority. An LMS like Mentor logs who approved the content and separates completion from a score with legal consequences. That way human oversight stays built in, and you can read the data details on the security page.