Do you know where your data goes with AI?
When a company uses AI, its data goes somewhere. Where data is stored and which country's law applies are different things. Five questions to ask your provider.
Do you know where your company's data goes when you use AI? When an employee pastes a customer's message into an AI assistant, or a company builds an AI agent into its own process, the data goes to a provider. People often ask only whether the data is in the EU. That is a good question, but it is not enough on its own.
This post looks at what the location of data settles and what it does not. The example is this site's own AI architect. At the end there are five questions worth asking every AI provider.
Location and jurisdiction differ
The location of data tells you in which country the data is stored and processed. Jurisdiction tells you which country's authorities can order the provider to hand the data over.
The US CLOUD Act of 2018 requires a US provider to disclose data in its possession "regardless of whether such communication, record, or other information is located within or outside of the United States" (18 U.S.C. § 2713). So the law also covers data that is in Europe.
Under oath before the French Senate on 10 June 2025, the director of legal affairs of Microsoft France could not guarantee that French public sector data would not be handed to US authorities without France's consent (French Senate record). Digipooli, part of Finland's National Emergency Supply Organisation, puts it this way: technical sovereignty is not settled by the location of data alone (Digipooli and Technology Industries of Finland, 2/2026, in Finnish).
Example: this site's AI architect
This site's AI architect uses Google's Gemini model in the Vertex AI EU multi-region. According to Google, the model processing then stays in EU member states (AI processing) (Google Cloud: Vertex AI data residency). Google is still a US company, so the CLOUD Act applies to it. The EU multi-region settles the location, not the jurisdiction.
The EU multi-region also has its limits (data residency):
- Google's global endpoint can process data in any Google data centre. The location has to be chosen separately.
- The EU multi-region is supported only for some models.
Nor does the EU multi-region mean that nothing is kept (Google Cloud: Vertex AI data governance):
- Caching is on by default. The in-memory cache is deleted within 24 hours, and the persistent cache is also on by default. Both can be turned off.
- Google may log prompts to detect abuse.
- When answers are grounded in Google Search, Google keeps search logs for 3 days, and this logging cannot be turned off. That is why we do not use Google Search in the architect.
We explain these in the privacy policy, and the location and home country of every service are in the sub-processor list on the EU-first page.
Five questions for your AI provider
- Where is the data processed and stored? Is the location tied to a specific region, or can the service process data anywhere?
- Which country is the provider from, and who are its sub-processors? Ask for the home country, not only the location.
- How long are prompts, answers and logs kept? Can caching and logging be turned off?
- Is the data used to train models? Ask for the answer in writing, in the contract.
- How do we get out? How quickly can the service be replaced, in what format do we get the data out, and what does the switch cost?
The EU Data Act answers part of the last question. The right to switch cloud services has applied since 12 September 2025: a customer can start the switch with at most two months' notice, followed by a transition period of at most 30 days. From 12 January 2027 no separate switching charges may be made (Bundesnetzagentur: switching and switching charges). The law does not solve technical lock-in, though. In the German Bitkom survey, lock-in was the biggest obstacle to switching cloud providers, named by 59% of companies (Bitkom Cloud Report 2026).
Why this matters now
- The largest GDPR fine was for a data transfer outside the EU, and other large fines have followed for transfers. Meta was fined €1.2 billion in 2023 (EDPB) and TikTok €530 million on 2 May 2025 (Irish Data Protection Commission). Yango was fined €100 million on 8 May 2026, in a decision in which the Finnish Data Protection Ombudsman took part (tietosuoja.fi, in Finnish).
- The EU–US Data Privacy Framework is in force, but uncertain. On 31 July 2026 the European Data Protection Board asked the Commission to assess whether the framework holds after a US Supreme Court ruling (EDPB letter).
- Decisions have been reversed in Finland too. On 24 March 2026 the Ministry of Justice reversed its decision to build the new election information system on AWS, because "the international political situation has changed" (Finnish Government, in Finnish).
- The risk is recognised, but plans are missing. In the Finnish companies Digipooli interviewed, the risk is widely recognised, but concrete plans or measures have not been widely made. The study is based on interviews with 20 business decision-makers (Digipooli and Technology Industries of Finland, 2/2026, in Finnish).
Dependencies as deliberate choices
According to the Digipooli report, full technological independence is not a realistic goal. We agree, and that is why we do not promise "100% EU": it would not be true for anyone. The goal is that every dependency is a deliberate choice: you know where your data is, whose law applies to it and how you can get out if you need to.
In practice this means three things:
- For every service, we state its location and the company's home country.
- The architecture does not tie you to one language model. Gemini can be replaced with the European Mistral or an open model on your own or a European provider's servers without rewriting the whole system.
- Every critical dependency gets an exit plan.
Read more about EU-first. If you want to map your own company's dependencies, contact us or ask the AI architect.