Where is your data when you use AI?

You let an AI assistant read along in your inbox, or you paste a quote into ChatGPT to have it checked. At that moment, something of yours leaves the building. Where is your data when you use AI, and who can get to it? Almost nobody asks that question up front. And almost every business owner gets it six months later anyway, from a customer, from their bookkeeper, or from an employee wondering whether this is even allowed. Here is the honest answer: no scare stories, and no pretending you need an IT department for this.

What happens when AI reads one customer email

Say a customer emails a question about a delivery, and your AI assistant drafts the reply. That email travels past three places. It helps to pull them apart, because they are three separate choices that have nothing to do with each other.

• The place where your files and messages already sat, your mail program, your document folder, your customer system.
• The place where the thinking happens, the computers of the company running the AI model. That is where the text of the email goes and where the answer comes from.
• The place where the connecting piece runs, the bit that fetches the mail, passes it on and puts the answer back. That is usually the part someone builds for you.

Each of those three can sit in the Netherlands, elsewhere in Europe, or in the United States. And you can choose them separately. That is exactly where these conversations go wrong: someone says “our data is in Europe” and means only the first point, while the thinking happens in America. Or the other way around.

‘European servers’ and ‘European company’ are not the same thing

This is the part most business owners never hear. The American CLOUD Act of 2018 requires US companies to hand over data to US authorities when there is a valid order, even if that data sits on servers in Europe. The law looks at who the provider is, not where the box stands. A data centre in Amsterdam owned by an American company is therefore something different from a data centre in Amsterdam owned by a Dutch one.

On the other hand, there is a European arrangement that simply permits data traffic with American parties. That arrangement was upheld by the European court at first instance in 2025, but an appeal is pending at the Court of Justice. Translated to your situation: it is allowed today, and it is not certain the rules will be the same in two years.

What you can actually choose, place by place

The good news: you have more choice than you think, and it usually costs nothing extra. The bad news: you have to choose up front, because moving afterwards does cost money.

For the thinking, the model that writes the text, the big providers now offer European processing, but not in every subscription. OpenAI has done so since February 2025 for its business editions (ChatGPT Enterprise and Edu) and for projects on its developer platform, where the text is processed in Europe and nothing is retained. A regular Plus or Business subscription is not included. That difference is large, and it appears nowhere on the pricing page. There are also European model makers, such as France's Mistral, where the question simply does not arise.

For the connecting piece, the part we and other agencies build for you, the choice is simplest. It can just run on a server in a Dutch data centre. At a Dutch provider like TransIP, such a server starts at around five euros a month at the time of writing. So that is rarely the reason not to do it.

• Thinking: ask which model is used, which country it runs in, and whether anything is retained.
• Storage: ask where the files, conversations and logs are kept, and for how long.
• Connecting piece: ask which data centre it runs in and whose name the contract is in.
• Everything you agree here belongs in the data processing agreement, not in a text message.

What it costs to get it right from the start

In most smaller-business automations we come across, the answer is: nothing extra. The choice sits in the set-up, not in the price. You pick European processing and a Dutch server when you build it, and that is that. What does cost money is switching later, then you rebuild part of it and have to move existing data.

And sometimes it simply is not possible. For certain functions, speech, images, one very specific model, there is no European route yet. Then the honest conversation is: do not use that part for sensitive data, or write down once why you find it acceptable in this case. Pretending the problem is not there is the only wrong option.

From 15 August 2026, your customer will ask

Until now this was mainly a matter for you. That changes this month. The Dutch Cybersecurity Act, the national implementation of the European NIS2 rules, takes effect on 15 August 2026, after the Senate approved it on 7 July. It applies to organisations in eighteen sectors above roughly fifty employees or ten million euros in turnover. The number of organisations covered grows from about a thousand to an expected eight thousand.

Not covered yourself? It still reaches you, because those eight thousand organisations are obliged to have the risks in their supply chain under control. They do that by questioning their direct suppliers, including suppliers who do not fall under the law themselves. In practice that means a questionnaire. And such a list always includes: where is our data processed and stored? There is no transition period for the core obligations, so those conversations start immediately.

An answer of “I don't actually know” costs you points with a buyer who is under supervision themselves. An answer of “our customer system runs in the Netherlands, the text is processed in Europe and nothing is retained” closes the conversation in thirty seconds. You do not build that difference in a week, you build it the moment you have the automation made.

Three questions to ask any AI supplier

You do not need to be technical for this. These three questions sort out in five minutes who has their affairs in order. Ask us them too.

• Where is my text processed and where is it kept? A good answer names a country and says whether anything is retained. A vague answer only names a brand.
• Is my data used to improve the model? The answer should be no, and it should be in the data processing agreement, not just in an email.
• What happens if I stop, or if you stop? You want to hear that you take your data with you, that the contract can be in your name, and that you are not locked to one party.

If someone talks around this, you have your answer. Not because something is necessarily wrong, but because a supplier who does not know themselves cannot help you explain your customer's questionnaire either.

When it really matters, and when it does not

We do not sell panic. For a large share of what smaller businesses do with AI, this is not a heavy issue. Nobody gets fined because an advertising text was written in America. The distinction you need to make once:

• Does matter: data on customers and staff, case files, quotes with prices and drawings, anything covered by a confidentiality agreement with a customer, and anything you supply to a larger company.
• Matters less: public texts, marketing ideas, summaries of things already on your website, general questions without names or amounts.
• Always wrong: sensitive data pasted into a free tool without any agreement, regardless of where that tool runs.

That last point is the biggest risk in practice. Not the map, but the employee who quickly pastes a customer file into a free chatbot because it is convenient. No European server helps against that, only an agreement people understand, and a working, safe tool that is just as easy.

What you can do this week

• Write down which AI tools are circulating in your business. Just ask; you will be surprised.
• Per tool, check the settings for whether you can switch on European processing and switch off the retention of conversations.
• Put one line on paper per tool: what may go in, what may not. That is all you need to have the conversation with your team.
• Having something built? Put the three questions above in your request, before anything is built.

Then the question in the title is no longer a problem, but simply an answer you have ready. For yourself, for your customers, and for the questionnaire arriving this month.