Shadow AI at work
Employees use AI the company has not introduced — what needs to happen now?
Set a framework instead of banning. According to the German BAuA survey DiWaBe 2.0, 62.1 % of employees in Germany use AI, often with tools the company has not introduced. What is needed: approved tools, clear data rules, training under Art. 4 AI Act and the works council involved from the start.
What the German BAuA survey DiWaBe 2.0 shows — and what follows for the company
| Finding | Figure (BAuA DiWaBe 2.0, Germany, survey 2024) | What follows |
|---|---|---|
| AI use at work | 62.1 % of employees use AI, 30.4 % increasingly | The question is no longer whether AI arrives in the company, but under which rules |
| Most important AI application not introduced by the company | for 51.4 % of those using it increasingly, for 75.7 % of those whose use has stayed the same | First find out what is being used — then write the rules |
| Wish for more AI | 36.4 % would like to use AI that is not yet used in their company | Take up the demand: an approved tool instead of a ban |
| Chatbots as the main application | 48 % of those using it increasingly (11.6 % where use has stayed the same) | Write data rules specifically for typing in and uploading |
| Further training | AI users do not take part in further training more often than employees without AI | Training does not happen by itself — Art. 4 AI Act requires AI literacy |
| Unequal distribution | without a formal qualification just under a third, with a university, master craftsman or technician qualification almost 80 % | Rules and training for everyone, not only for those who already use it a lot |
How widespread is AI use without the company introducing it?
Very widespread. The BAuA survey DiWaBe 2.0 questioned around 9,800 employees subject to social security contributions in Germany in 2024; it was published in 2025. Result: 62.1 % use AI at work, 30.4 % increasingly over the last five years. For 51.4 % of these increasing users, the company did not introduce the most important AI application; among employees whose use has stayed the same, the figure is 75.7 %.
The BAuA describes this as a development from below: companies lag behind the actual and desired use of their employees. As possible reasons it names data protection concerns and missing strategies for bringing the technology into work processes. A third finding fits this: 36.4 % of respondents would like to use AI that is not yet used in their company.
For dealing with it, this means: your employees have long been trying out whatever makes their work easier. Generative AI does not need the company to introduce it; many services are freely available online. A company that sets rules now is bringing an existing practice into the light — it is not inventing a new one.
Why is a ban usually the weaker route?
A ban does not remove the need, it makes it invisible. Anyone who gets to a draft faster with a chatbot will, if in doubt, keep using it — just on their private phone. The company no longer sees any of it and can neither advise nor set limits. With texts that employees copy into an input field, that is a real risk: customer data, personnel data or trade secrets end up with a service the company has no contract with.
The stronger route is an approved tool with clear rules. It covers the most common purposes, is contractually secured (for personal data, a data processing agreement under Art. 28 GDPR), and it is clear which data may go in. The demand is there — 36.4 % want more AI. Meeting it brings use to where it can be shaped.
Where a tool stays excluded for good reasons, the reasoning belongs with it. A ban that is explained and names an alternative is more likely to be followed than one that only sits in a policy.
Which rules does the company need?
One page is enough to start. It answers which tools may be used, which data are off-limits, who is accountable for results and whom to ask with questions. What matters is that the rules fit everyday work: chatbots are the main application for 48 % of those using AI increasingly — so above all, the rules have to say what may go into an input field and what may not.
Three legal guardrails apply regardless of company size. First, Art. 4 of the AI Act: since 2 February 2025, deployers of AI systems must ensure a sufficient level of AI literacy among their staff. Second, the GDPR: the company must use rules and safeguards to ensure that personal data do not end up in services it has not chosen (Art. 24 and 32 GDPR) — and it bears the risk if it happens anyway. Third, co-determination (German law): where a works council exists, it must be informed in good time under § 90 BetrVG about the use of AI in work processes; AI systems capable of recording performance or behaviour regularly trigger mandatory co-determination under § 87 (1) no. 6 BetrVG.
Training does not run by itself. According to DiWaBe 2.0, AI users do not take part in further training more often than employees without AI — not even more often during working hours. Anyone who wants competence has to organise it and plan time for it.
- Approved tools: which ones, for which purposes, with which account
- Data rules: no health or personnel data, no customer data, no trade secrets in services that are not approved
- Responsibility: a person checks every result before it leaves the company or decides about people
- Contact person: who answers questions and reviews new tools
- Training: basics, limits, data protection — during working hours and documented
What does AI use have to do with strain and health?
More than it first seems. The DiWaBe 2.0 data show a double finding: employees who use AI increasingly decide more often themselves about new tasks and their workload — and at the same time report more often deadline and performance pressure and an information overload they cannot cope with. Multitasking and interruptions showed no significant difference. The BAuA found no statistically significant correlation between AI use and health.
Important for putting this in context: the data do not distinguish between AI introduced by the company and AI employees chose themselves, and the study shows correlations, not causes. Whether AI increases pressure or whether AI is used more often where pressure is high anyway cannot be decided with these data — the BAuA says so itself.
For occupational safety, something practical follows: what the company does not know, it cannot take into account in the psychological risk assessment. As soon as AI tools are officially introduced and change tasks, pace or collaboration, that is a changed circumstance within the meaning of § 3 ArbSchG (German law) — the assessment is then extended by an AI block.
How do you proceed in practice?
In five steps that can be implemented within a few weeks. The first is a stocktake without blame: a short, anonymous survey asks which tools are used, for what, and what is missing. Results are evaluated only for groups of five people upwards, so that nobody can be identified. The survey serves design, not control — that belongs in the invitation.
The answers show which purposes an approved tool has to cover. Then come the rules on one page, the training and the involvement of the works council, where there is one. Six to twelve months after the roll-out, an additional AI block in the risk assessment shows whether workload, pressure or information overload have shifted.
- 1. Survey anonymously: which AI is used for what, what is missing? (evaluation from 5 people)
- 2. Approve a tool that covers the most common purposes — with a data processing agreement
- 3. Rules on one page: tools, data rules, responsibility, contact person
- 4. Training during working hours, documented (Art. 4 AI Act)
- 5. Involve the works council early; after 6–12 months an AI block in the risk assessment
Related measures & topics
Key takeaways
- BAuA DiWaBe 2.0 (Germany, 2024, around 9,800 employees): 62.1 % use AI at work, 30.4 % increasingly.
- For 51.4 % of those using AI increasingly, the company did not introduce the most important AI application; 36.4 % want AI that is not yet used in their company.
- An approved tool with clear data rules brings use to where it can be shaped — a ban only makes it invisible.
- Art. 4 AI Act (since 2 February 2025): ensure AI literacy among staff — according to the BAuA, AI users have so far not taken part in further training more often.
- Involve the works council early under § 90 and § 87 (1) no. 6 BetrVG (German law); after the roll-out, extend the risk assessment by an AI block.
Frequently asked questions
May employees use private AI accounts for work?+
The company decides — and should do so explicitly. Without a rule, use is neither clearly permitted nor clearly prohibited. The risk then lies mainly with the company: it must use rules and safeguards to prevent personal data from ending up in a service with which there is no contract (Art. 24 and 32 GDPR).
Isn't a survey on AI use itself a form of monitoring?+
Not if it runs anonymously, only groups of five people upwards are evaluated and the purpose is clear beforehand: finding out which tools are needed. It asks about purposes and wishes, not about the behaviour of individuals. Where a works council exists, it is involved beforehand.
Must the works council be involved?+
Where there is one, yes. Under § 90 BetrVG (German law) it must be informed in good time about the use of AI in work processes. AI systems suitable for recording performance or behaviour regularly trigger mandatory co-determination under § 87 (1) no. 6 BetrVG — regardless of whether monitoring is intended.
What does Art. 4 AI Act require specifically?+
A sufficient level of AI literacy among staff who work with AI systems — measured against their knowledge, experience and the purpose of use. The regulation does not prescribe a particular training format or a certificate. A sensible approach is basic training for all users, supplemented by guidance for particularly sensitive tasks, and a short record of who was trained when.
Does AI use make employees ill?+
The German data have not shown that so far. The BAuA survey DiWaBe 2.0 found no statistically significant correlation between AI use and health — but it did find one between increasing use and more performance pressure and information overload. The BAuA itself names the limits: people who use a lot of AI may be particularly tech-savvy or resilient, and health effects often only show up with a delay. What matters for strain is therefore how work with AI is designed.
Ask first, then set rules
With EasyWHP you ask anonymously which tools your employees use and what they are missing — evaluated only for groups of five people upwards. The results feed into the psychological risk assessment and the measures plan.
Sources
- BAuA report F2573 — Digital transformation and the changing world of work (DiWaBe 2.0), 2025 ↗
- Art. 4 AI Act — AI literacy (in force since 2 February 2025) ↗
- Art. 28 GDPR — Processor ↗
- § 90 BetrVG — Information and consultation rights (use of AI) ↗
- § 87 BetrVG — Co-determination rights of the works council ↗
- § 3 ArbSchG — Basic duties of the employer (adaptation to changed circumstances) ↗
Last updated: 2026-11-11. Not legal or tax advice — have your specific case reviewed by a professional.