If you think nobody in your company uses ChatGPT, Gemini or Claude for work - they do, they just don't tell you. Shadow AI, meaning bottom-up use of AI tools outside the organisation's knowledge and control, is the norm today, not the exception. The problem is not that employees want to work faster. The problem is what gets pasted into the prompt window along the way.
- An employee pasting personal data into a public chatbot exposes the company to a data breach - with no data processing agreement, legal basis or information for the data subjects.
- Trade secret protection requires "reasonable steps" to keep information confidential - doing nothing about widespread public chatbot use weakens that protection.
- The AI literacy obligation under Art. 4 of the AI Act has applied since 2 February 2025 and also covers companies that "only use" off-the-shelf AI tools.
- A hard ban on AI use does not eliminate Shadow AI - it pushes it deeper into the shadows and the company loses visibility.
- The 5-step plan to bring order: fact-finding, enterprise tools with a DPA, a short AI use policy, documented training, and a regular review.
Why is Shadow AI a legal problem, not just an IT one?
The GDPR: personal data in prompts
An employee who pastes a client's email, a candidate's CV or a list of training participants into a public chatbot causes the company - as the controller - to transfer personal data to an external provider. With no data processing agreement, no legal basis, no information for the data subjects, and often with a transfer outside the EEA. In the free versions of many tools, conversation content may be used by default to train models - meaning the data leaves the company and does not come back.
From a GDPR perspective, such an incident may be a personal data breach: disclosure of data to an unauthorised recipient. And that sets the familiar machinery in motion: risk assessment, possible notification to UODO (the Polish data protection authority) within 72 hours, the breach register. Except that the company usually never even finds out about a breach caused by Shadow AI - and that is the most dangerous scenario.
Trade secrets
The second front does not concern personal data, but it can be more expensive: source code, contract terms, financial data, product plans pasted into an external tool. Trade secret protection (under the Polish Act on Combating Unfair Competition, ustawa o zwalczaniu nieuczciwej konkurencji) requires the company to take reasonable steps to keep the information confidential. A company that knows about widespread use of public chatbots and does nothing about it weakens its own position: it is harder to argue that information was protected if people were allowed to send it outside freely. On top of that come NDA commitments to clients - pasting their materials into a public tool may simply be a breach of contract.
The AI Act: the AI literacy obligation (Art. 4)
Since 2 February 2025, Art. 4 of the AI Act has applied: providers and deployers of AI systems are obliged to support the development of AI literacy among the people who use those systems on their behalf - taking into account their knowledge and experience, the context of use and the persons affected by the AI. The Digital Omnibus (Regulation 2026/1744) softened the wording - it previously required an “adequate level” of literacy to be ensured - but did not repeal the obligation. This also applies to companies that “only use” off-the-shelf tools. Since employees actually use AI at work (and they do), a company without training and rules of use has an unaddressed legal obligation. We describe the full AI Act timeline in a separate article.
Why doesn't a ban on using AI work?
The “block it and ban it” reflex is understandable and almost always ineffective. AI tools are available on personal phones and built into search engines, browsers and office suites. A hard ban does not eliminate use - it pushes it deeper into the shadows: employees keep using the tools, they just hide it more carefully, and the company loses whatever visibility it had left. On top of that, a ban hands competitors a real productivity advantage. The more sensible strategy: channel the use instead of fighting it.
A 5-step plan to bring order
1. Take stock without a witch hunt
A survey or short conversations with teams: who uses what and for what. The key is the communication framing - the point is to choose good tools and rules, not to punish anyone. If your first move is disciplinary proceedings, you will never get a second round of fact-finding, because nobody will own up.
2. Choose and roll out enterprise tools
The most effective cure for Shadow AI is offering a legitimate alternative. Business versions of the main tools (including ChatGPT for business, Microsoft 365 Copilot, Gemini for Workspace, Claude for Work) usually offer: no model training on customer data, a data processing agreement (DPA), administrative controls and SSO. When choosing, check: the data processing terms, where processing takes place and the transfer mechanism, prompt retention, and the option to switch off history. This is the classic processor vetting under Article 28 GDPR - AI is no exception here.
3. Write an AI use policy - a short one
A document people will actually read, not 40 pages to tick off. The minimum:
- a list of permitted tools (company accounts) and a rule that anything else requires approval;
- what must not be pasted into any tool without a company agreement: personal data, trade secrets, client materials covered by an NDA, access credentials;
- a rule on verifying outputs - a human is responsible for the result of the work; AI can be confident and wrong at the same time;
- labelling and transparency where the context requires it (e.g. content published externally);
- a reporting path for new tools and incidents - without an atmosphere of informing on colleagues.
Introduce the policy formally (like other internal regulations) and link it to your GDPR breach procedure: pasting personal data into a non-permitted tool = a potential breach to be assessed.
4. Train people - and document it
Training closes out the AI literacy obligation under Art. 4 of the AI Act, but it also has a purely practical dimension: most Shadow AI incidents are not bad faith, just ignorance (“it's only a chat”). Scope: how language models work and what happens to the data you type in, the risks (hallucinations, confidentiality, copyright), the rules from the policy, good practice for specific roles. Keep attendance lists and materials - they are evidence that the obligation has been met, useful both under the AI Act and when demonstrating the “reasonable steps” protecting trade secrets.
5. Review regularly
The AI tools market changes every quarter, and the AI Act applies in stages - the bulk of its provisions applies from 2 August 2026, and the package for Annex III high-risk systems, as amended by the Digital Omnibus, from 2 December 2027. Once a year (at a minimum) review the list of tools, the policy and the training programme. If the company starts using AI in HR, customer assessment or other sensitive areas - check the risk classification before the supervisory authority does it for you. Quick start: our AI Act Check quiz.
Where to start this Friday
One board decision: “we are putting AI use in order, not punishing the past”. Then, in this order: fact-finding → company accounts → a one-page policy → training. The whole thing can be wrapped up within a quarter - and each of these elements immediately reduces legal risk and improves the company's position in case of an incident.
This article is for information purposes only and does not constitute legal advice.