EU hosting and GDPR: what to actually check

“We host in the EU.” Every vendor says it by now, and it answers almost nothing.

Meanwhile AI stopped being an experiment at your company a while ago. Someone is summarising documents, drafting proposals, searching internal knowledge. Those prompts carry customer details, internal strategy, contract language, sometimes personnel files. Introducing a language model is therefore a decision about data flows, accountability and risk, not just about features. EU hosting is an important building block in that. It is simply not the seal of approval it gets sold as.

Why the place of processing matters

The GDPR does not prohibit cloud or AI services outside the EU. That is the first misconception I run into constantly. But the moment personal data leaves the European Economic Area, or a provider in a third country can access it, you pick up extra assessment and documentation duties: evaluate the legal basis, agree safeguards, account for possible access by public authorities.

A platform hosted in the EU takes a good part of that off your plate. Your data stays closer to your own legal structures, which makes impact assessments, internal approvals and conversations with customers or the works council noticeably easier.

The catch: what matters is not where the interface runs, but where the request is processed. Those are two different things, and marketing pages love to merge them into one.

The five questions that decide it

“EU hosting” does not mean the same thing everywhere. The only vendor review worth doing is one with concrete questions. If you take one thing from this piece, take this table.

Question	Why it matters	A good answer sounds like
Where do the application and storage run?	Processing and storage can sit in different regions.	Named EU regions for both
Are inputs used for training?	Your data must not drift into model improvement uncontrolled.	Training excluded by default
Which subprocessors are involved?	Downstream services shape the data flow too.	A current, open list with locations
How is data deleted?	Retention has to match the purpose.	Documented periods, a technical deletion process
Who has administrative access?	Support access creates a third-country link fast.	Restricted, logged, approved

An evasive answer is itself an answer. If someone cannot name the region for question one in a single sentence, they either do not know it or do not want to say it. Both tell you something.

Four layers that belong together

A responsible rollout needs more than a tick next to “EU”. In practice four things hang together:

1. Data and purpose. What may go in, for what, on which legal basis? Confidential material needs clearer rules than a marketing text.
2. Technology and access. Roles, single sign-on, logs, clean tenant separation. In short: who sees what.
3. Contracts. Data processing terms, subprocessors, retention, international transfers, matching the actual architecture rather than the intended one.
4. Organisation. Policies do not change behaviour. Your people need understandable guardrails, examples and an approved route.

Point four is the one nearly everyone underestimates. Block the public chatbots and the demand does not disappear. Your people still have the same deadlines, they just do it somewhere you cannot see. A central route does not win because it is safer. It wins because it is more convenient than the covert one. That operating model is what the term Corporate LLM describes: the category is defined not by the model, but by the layer of access, evidence and contract around it.

Several models, without new data silos

Different models produce the best results for different tasks. If every department sources its own service, you get new contracts, new accounts and data paths nobody oversees. Bundled access solves that without locking you to one model. Why that is more than convenience is in Why one AI model is never enough.

One more thing that appears in no vendor deck: far from every use case needs personal data. Plenty works just as well with anonymised or deliberately reduced input. Data minimisation belongs in the workflow, not in the review six months later.

How to start

A safe rollout does not take months. Take a few bounded cases where the value is measurable and the data controllable: structuring internal texts, drafting communication, searching approved knowledge.

Name accountable people from the business, IT and data protection together. Write down in plain language what may go in and what may not. Provide one central, approved route. Train your people on real situations from their own work, not on examples from a slide. And review usage, quality and open risks regularly.

Unclear rules slow you down more than strict ones, by the way. When every use is a case-by-case decision, everyone decides for themselves and nobody asks. A clean framework does the opposite: your people know what they are allowed to use, you know where the data sits, and you enable new models in a controlled way instead of banning them.

EU hosting is the foundation for that. It is just the foundation, not the house.