This page answers the question we are asked before any other: what happens to a document once it enters a system we build. The answer depends on what you are protecting, so it is given once per deployment option, followed by how we choose models, what is settled before anything goes live, and which European rules apply to whom.
Models run on servers you own, inside your network; nothing crosses the network boundary. For material that may not leave the premises under any policy. The trade-off is that only smaller, local models are available, and you run the hardware.
A dedicated tenant in the EU, hardened and operated with you — inference, storage and compute inside the EU, encrypted in transit and at rest, access-controlled and logged. For organisations whose customers or auditors check their security, and who need stronger models than local hardware allows.
Prompts and the documents you choose to send go to a European-hosted model under a contract that excludes training on your data and limits retention. For lower-sensitivity drafting and research, with the least setup.
For non-sensitive material where the strongest available model matters more than jurisdiction. Data leaves the EU; confidentiality rests on the provider's business terms. We say so plainly, and we help you decide whether that is acceptable for the material in question.
In practice most organisations combine them: sensitive material on the first two, everything else on the last two, with one verification layer and one set of rules across all of it. We design the split with you and document which material goes where.
One verification layer and one set of rules across all of it; the split documented, material by material.
We recommend European-hosted models by default, open- or closed-weight, and local or private inference where confidentiality requires it. In practice we help each client pick what works for their constraints: model quality for the task, data residency, contract terms on retention and training, and the cost of running it. The choice is documented and stays with the client; nothing is locked to a vendor. Our own development work runs in Europe, in the US or locally, depending on the client's situation, and never on client documents without that client's arrangement.
Where the documents go, under whose contract, and what is retained. Who can access the system and how that is enforced. What checks the answer — sources cited, assessments labelled. Who owns the system after we leave, and what the maintenance routine is. What is logged and for how long. What happens when the system is unsure or unavailable. These are the twelve items of the pilot-to-production checklist, and we do not skip them for a pilot.
The pilot-to-production checklist →How the verification layer works →
Three European regimes come up in every rail conversation about AI. They bind different parties, and being precise about that is usually worth more than a compliance promise.
Where the dates stand. The Cyber Resilience Act's reporting obligations for actively exploited vulnerabilities and severe incidents have applied since 11 September 2026; the remaining obligations apply from 11 December 2027. The AI Act's prohibitions and its AI-literacy duty have applied since February 2025, its transparency rules since August 2026, and — after the amendment that took effect in July 2026 — the high-risk obligations under Annex III from 2 December 2027 and those for AI in other regulated products from 2 August 2028, while for rail the requirements are to come through the rules on rail interoperability. The Data Act has applied since 12 September 2025. These regimes are still being amended; we check the dates in each engagement rather than relying on a page.
Our own side. The AI Act's literacy duty applies to us as well, both as a provider and as a user of these tools, and we keep a record of how we meet it. This website is a static site, so it is not a product with digital elements, and it has no chatbot. The one AI-generated image on the site, the portrait on /about, is labelled as such. If we ever add a chatbot, it will say that it is an AI.
