Where does your data gowhen you use AI?

A fair worry

When we talk to companies about AI, the first question is often not what it can do, but where the data goes. That is a good instinct. Mails, contracts and client files contain personal data and business secrets. Putting them into a tool without knowing what happens to them is a risk you should not take.

The good news is that using AI safely is possible, and not even very complicated, as long as you ask the right questions.

The journey of your data

When AI reads a document, the text is sent to a language model, processed, and an answer comes back. Three questions matter. Where does that processing happen, in Europe or elsewhere? Is the data stored, and for how long? And is it used for anything else, such as training the model?

With consumer chat tools the answers are not always clear, and staff sometimes paste confidential information into free accounts. With business use through proper agreements, you can choose providers and settings where data is processed in Europe, not stored longer than needed and not used for training.

Questions to ask any provider

Where is data processed and stored? Which subprocessors are involved, including the model provider? Is my data used to train or improve models? How long is it kept, and can I have it deleted? Who on your side can access it? Is there a data processing agreement under the GDPR?

Any serious provider can answer these in writing. If the answers are vague, that is your answer.

Access rights inside the company

Security is not only about outside parties. An AI assistant that can search everything may show an employee information they are not supposed to see. A well built system respects the same access rights as your existing software: the assistant only answers from what that person is allowed to see, and every agent only gets the rights it needs for its job.

What the EU AI Act asks

The EU AI Act sorts AI uses by risk. Most business uses, such as reading mails, drafting replies or following up deadlines, are not high risk. Still, a few things apply to every company that uses AI.

Since February 2025 companies that use AI must take measures to support the AI literacy of the people who work with it. Since the Digital Omnibus came into force in July 2026, the text makes clear that this is about taking measures, not about guaranteeing a certain level per person. Since 2 August 2026 most other rules apply, including transparency: people must know when they deal with an AI system. The obligations for high risk systems, such as AI used in hiring decisions, moved to 2 December 2027. We explain the details in a separate article on the AI Act for SMEs.

What the system learns is yours

One more point that is easy to overlook: when a system learns from your corrections and approvals, that knowledge has value. Make sure your agreement states that your data, and what the system learns from it, belongs to you. If you ever switch providers, you should be able to take it with you.

The risk of free tools

The biggest data risk with AI in many companies today is not a sophisticated attack. It is a colleague pasting a client contract into a free chatbot to get a quick summary. It is well meant, and it happens daily. Forbidding AI rarely helps: people use it anyway, out of sight. Giving them a safe alternative and clear guidelines works far better.

A short policy helps: which tools are approved, which data may never be shared outside them, and who to ask when in doubt.

Logging and traceability

When AI takes actions, you want to know afterwards what it did and why. A good system logs every suggestion and action with the sources it used and the person who approved it. That is useful for security, for quality and for the AI Act, and it builds trust: when something goes wrong, you can see exactly what happened and correct it.

Which model, and does it matter?

For data protection, the model itself matters less than how it is used: through which provider, in which region and under which terms. Many large models can be used with processing in Europe and without training on your data. Choosing the setup is part of building the system, and a good partner explains the choice and puts it in writing.

Some companies prefer models they can run in their own environment. That is possible too, at a higher cost and usually with somewhat less capable models. For most SMEs, a European setup with the right agreements is the sensible middle ground.

A practical start

Write down which tools people already use for AI, including free ones. Agree on simple guidelines for what may and may not be shared. Choose providers that answer the questions above clearly. And when you build AI into your processes, build it into your own environment, with your access rights, instead of copying data into separate tools.