Shadow AI: when your team secretly uses ChatGPT

It's Thursday evening and the quote has to go out. A colleague copies the customer's enquiry into his private ChatGPT account, has the text polished and sends it off. No ill intent, quite the opposite: he wanted to deliver good work quickly. Even so, customer data has just flowed to a US service the company knows nothing about. That is exactly what shadow AI is, and it has long been part of everyday working life.

What is shadow AI?

Shadow AI refers to AI tools that employees use for work without the company having checked or approved them. That includes the private ChatGPT account, the AI feature in a free app, the browser extension, the free tool picked up online. The term borrows from shadow IT, but the phenomenon is bigger: AI tools are installed in seconds, available for free and built into more and more everyday software.

How widespread is the problem?

The figures are clear. According to the German digital industry association Bitkom, the use of private AI tools is widespread in 8% of German companies, happens in isolated cases in a further 17%, and another 17% suspect it. Taken together, shadow AI is taking place, or is at least suspected, in more than 40% of companies. At the same time, only 23% have any rules at all for using AI.

Things look no better internationally: in Microsoft and LinkedIn's Work Trend Index, 78% of AI users said they bring their own tools to work, and in small and medium-sized businesses the figure was as high as 80%. A global KPMG study found in 2025 that almost half of employees have already entered sensitive company data into public AI platforms. And more than half even keep their own AI use a secret.

What shadow AI costs

IBM's Cost of a Data Breach Report 2025 shows that this is not just theory: one in five companies hit by a data breach was compromised through shadow AI, and such incidents increased the cost of the damage by around US$670,000 on average. Customer data was affected strikingly often.

Four cases that show how it goes wrong

• The classic: At Samsung in 2023, engineers entered confidential information into ChatGPT three times within a few weeks, including source code and the minutes of a meeting. The group then banned generative AI completely.
• The sharing trap: In summer 2025, tens of thousands of shared ChatGPT chats turned up in Google's index, complete with names and business details. The share function had made the content publicly findable.
• The insecure free tool: In early 2025, a database belonging to the Chinese provider DeepSeek with over a million entries, including chat histories, was openly accessible on the internet. Berlin's data protection commissioner later reported the app to have it removed from the app stores.
• The malware trap: A Disney employee privately installed a free AI image tool that contained a Trojan. Through his accounts, 1.1 terabytes of internal data were siphoned off. He lost his job afterwards.

The legal position: why your company is liable

The tricky thing about shadow AI: legally, the company remains responsible, even if it knows nothing about the use.

• GDPR: If an employee enters customer data into a private AI account, a third party processes personal data without a legal basis and without a data processing agreement. This can be a reportable data breach, and the 72-hour deadline under Article 33 GDPR runs regardless of whether management suspects anything.
• Trade secrets: The German Trade Secrets Act (GeschGehG) only protects those who can prove they took appropriate confidentiality measures. A company with no AI rules at all will hardly be able to provide this proof and risks losing protection for its customer lists, price calculations and processes.
• Contracts: Confidentiality agreements with customers apply independently of any law. Customer information in a private AI chat can amount to a breach of contract with liability for damages. For small service providers, this is often the most tangible risk.
• EU AI Act: If you use AI at work, you are a deployer within the meaning of the AI Act, with obligations such as AI literacy under Article 4. Tools that nobody knows about can neither be classified nor covered by training.

And the employees themselves? If AI use has been explicitly prohibited, unauthorised use can justify a formal warning, and disclosing trade secrets can justify considerably more. So far there is no published German court ruling on a dismissal over shadow AI. Nobody should rely on that, though.

Why bans don't work

Many bosses' instinctive reaction is a ban. The data speaks against it: in a Software AG survey, 46% of employees said they would keep using their AI tools even if they were banned completely. A ban only shifts the use to private phones, where you can't see anything at all. On top of that, your team's motivation is actually a good one: people want to work faster and better. That is exactly the energy you should harness instead of fighting it.

Out of the shadows in five steps

1. Take stock without blame: Ask your team openly which tools are in use, and make it clear that this isn't about punishment but about visibility.
2. Provide an official alternative: A business account with a data processing agreement and model training switched off takes away the main reason for shadow use. People who are given good tools don't need secret ones.
3. Set up a short AI policy: Permitted tools, prohibited inputs, a duty to check results, one contact person. Two pages are enough. Our guide to writing an AI policy shows you how.
4. Train your team: People who understand why customer data doesn't belong in a free chat tool are more likely to stick to the rules. Along the way, you also meet the AI literacy obligation under Article 4 of the EU AI Act.
5. Keep at it: New tools appear every month. A simple approval process and a well-maintained AI inventory keep the system up to date.

How Klaus takes the five steps off your hands

In practice, the five steps rarely fail for lack of will, but because of the effort involved: who builds the inventory, who writes the policy, who organises the training? This is exactly the work Klaus takes on:

• Inventory by conversation instead of questionnaire: Klaus walks you through your AI use in a chat. For common tools like ChatGPT or Copilot, Klaus suggests the typical setup and you just confirm it. Each tool gets a profile with its purpose, data categories and a clear classification: permitted, permitted with conditions, or prohibited.
• The shadow rule is included automatically: For tools that are not on the list, you set a default rule, for example "prohibited until approved" or "quick request to the AI contact person". The finished policy then states word for word: "The covert use of non-approved AI (shadow AI) is prohibited." That closes the grey area shadow AI thrives on.
• Training brings the 'why' to your team: The Klaus training courses use each department's own tools to explain why customer data doesn't belong in a free chat tool. Every completed session is automatically logged in the training records, which also takes care of Article 4 along the way.
• Keep at it without project overhead: If a new tool turns up, you update the inventory in a short chat. Klaus only asks about the change, and the policy is automatically regenerated as a PDF.

Conclusion

Shadow AI is not a sign of a disloyal team, but of missing tools and rules. The risks are real: data breaches, lost trade secret protection, breached customer contracts. Fortunately, the solution is unspectacular: provide good tools officially, set clear rules, run a short training session. That way you bring AI use into the light and turn a risk into a productivity gain.