Shadow AI: when staff use private chat services

Shadow AI means that employees use AI services that the company has neither provided nor checked, usually a private chat account, and enter project data there. According to a survey by the industry association Bitkom, this happens in one company in four, and a further 17 per cent suspect it. A ban alone helps little, because private use is hard to monitor and the need remains. An approved offer with clear rules is more sustainable. That is my assessment on the basis of the sources named below, and it is not legal advice.

How widespread is it?

On behalf of the association, Bitkom Research surveyed 604 companies in Germany with 20 or more employees by telephone in calendar weeks 27 to 32 of 2025. The results are in the press release of 21 October 2025. In 8 per cent of the companies, private AI use by employees is widespread (2024: 4 per cent), and in 17 per cent there are individual cases (2024: 13 per cent). A further 17 per cent are unsure but assume it happens. Bitkom sums up: four in ten companies at least assume that private AI tools are in use. Only 29 per cent are sure that it does not happen (2024: 37 per cent). As of November 2026, I found no more recent Bitkom survey on the subject.

The Bitkom figures reflect the view of the companies, not that of the employees, and they apply only to firms with 20 or more employees. In 2025, 26 per cent of the companies provided official access to generative AI, and 23 per cent had rules for its use (2024: 15 per cent).

The employees’ side is shown by a survey of almost 5,000 employees by the German Economic Institute (Institut der deutschen Wirtschaft, IW Köln) in March and April 2026. The short report appeared on 19 September 2026, so this paragraph reflects the position as of November 2026. Of the employees who use AI at work (2,793 respondents), 29.1 per cent say they use tools that the employer has not provided. In firms with fewer than 50 employees the figure is 33.5 per cent, in firms with 250 or more employees 25.5 per cent. The IW points out that “not provided” does not necessarily mean secret or prohibited, but it does mean outside the company’s control. The figures rest on self-reporting. A second finding stands out: employees who credit their employer with strong commitment to building AI skills use tools that were not provided more often, at 44.5 per cent, than those who see no commitment at all (21.3 per cent). The IW reads this as an eagerness to experiment that runs ahead of the official offer, and points out that the data do not show which way the effect runs. For small firms it sees a lack of official tools as a possible reason in many cases.

Why bans alone do not hold

A ban assumes that you notice violations. The Bitkom figures show that many companies do not know: 17 per cent are unsure and suspect private use, 24 per cent are unsure and suspect none, together 41 per cent. There is also the motive. Anyone who wants to summarise texts or draft emails in daily work and has no tool at the company is more likely to turn to a private one. Training or encouragement alone does not replace an approved tool, however, as the IW finding above shows.

The Data Protection Conference (DSK) sees it similarly: for professional use, employers should provide devices and accounts, and employees should not have to work with private accounts and devices. Its statements on missing rules are in the article AI policy for the office: what it should contain. Bitkom advises the same: “Companies should avoid uncontrolled proliferation of AI and prevent shadow AI from developing” (our translation). The association recommends setting out rules for the use of AI in the workplace in company policies.

What this looks like in the office

Three invented examples show how project data ends up in private services. A site manager has meeting minutes with the names of the subcontractors smoothed by a private chat account. An administrator uploads a list of the tenants of a residential estate in order to prepare letters. An estimator types in the prices of a live tender to have the justification drafted. None of the three means to do anything wrong. All three save time and do not know what happens to the data afterwards.

What a private service does with inputs depends on the account and the settings. For that reason the DSK requires checking whether inputs and outputs are used for training, whether people are informed about it and whether it can be ruled out. With private accounts the company cannot control this.

Which risks exist for project data

Confidentiality

Drawings, contracts, prices and calculations are often business secrets or subject to confidentiality agreements with clients. Under section 2 no. 1(b) of the German Trade Secrets Act (Geschäftsgeheimnisgesetz, GeschGehG), protection as a trade secret requires appropriate secrecy measures (more in the article AI policy for the office). Anyone who has neither rules nor control over where data goes may have a harder time in a dispute. How that turns out in a particular case is for a lawyer to assess.

Data protection

Project documents contain names of residents, owners, employees and tradespeople, sometimes photos. With a private account the company generally has no contract under Article 28 of the GDPR. The DSK describes open systems like this: the input data leaves the protected area and, depending on the design, may also be used to answer other users’ queries, often with transfer to third countries. Uncontrolled use could lead to data protection violations. With private accounts, profiles of individual employees can also arise.

Copyright, in general terms only

Under section 2(2) of the German Copyright Act (Urheberrechtsgesetz, UrhG), works are only personal intellectual creations. Whether and when a text that a service produces without any creative input of its own is protected is a legal question for the individual case. Conversely, drawings, texts and photos of third parties can touch on rights when they are fed into a service. The assessment is for a lawyer.

Accuracy

The DSK points out that providers of language models often state themselves that their texts make no claim to accuracy. Without checking, errors find their way into tenders and letters.

What a company can offer instead of just banning

1. An approved tool with a company account, a data processing agreement (Auftragsverarbeitung) and training switched off. How such an agreement can be checked is covered in the article data processing agreements for AI services.
2. A short policy with examples of what is allowed and what is not, and with data classes. The article on the AI policy shows the structure.
3. A point of contact where staff can ask whether a use is acceptable and where reports can be made without first asking who is to blame.
4. Training with examples from the company’s own daily work. The article on AI training by role describes the content for each role.

Technical blocks and logging of use can touch on co-determination and employee data protection. Section 87(1) of the German Works Constitution Act (Betriebsverfassungsgesetz, BetrVG) names questions of order in the establishment (no. 1) and technical devices for monitoring behaviour or performance (no. 6). Clarify this beforehand if you are considering anything of the kind.

What you can do now

1. In a short, friendly conversation, ask which AI services are used for work and for what.
2. Decide which tasks an approved tool should take over, and choose it with a contract in place.
3. Write one page of rules with three or four examples from your daily work and discuss it with the team.
4. Decide who answers questions, and clarify with your data protection officer how to deal with data that has already been entered.