Built to be trusted. Designed to be audited.
Principles of company AI with EU data
Four principles we hold to on every deployment, from the first pilot. Not slogans: architectural decisions you can verify.
Updated
Why principles come before features
A corporate AI environment touches the most delicate material a company holds: contracts, payroll, records, client data that is not even its own. Before talking about models or dashboards, it is worth being clear about what the system does with that information, where it sends it and what trace it leaves. If that is unresolved, the rest hardly matters.
What we do not promise
No detection is perfect. Some sensitive data follows no recognisable pattern, and in some contexts the right call depends on human judgement. We work to reduce risk measurably and to record every decision, not to sell a guarantee no system can give.
How to verify them
A pilot starts in audit mode: for about ten days the system watches without blocking, and at the end you get a report of what would have left your company with no controls at all. That report is the proof. Our pitch is not.
The four principles, one sentence each
EU data residency: sensitive data is routed to models hosted in the EU and everything else leaves sanitised. No training: your team's content does not train third-party models, and it is set once for the whole company. Evidence log: every decision the perimeter makes is recorded and exportable. Single sign-on: one login with Google or Microsoft, and joiners and leavers handled in one place.
Who it helps to have them written down
The security lead or DPO who has to defend AI use to management. The person who runs a tax firm, a law firm or a clinic and answers to clients. And whoever in IT wants to stop firefighting personal accounts. The four principles are the questions each of them ends up asking, ordered from most to least urgent.
Frequently asked questions
What is a data perimeter for AI?
A layer placed between the people in your company and AI models. It classifies each request before it leaves and applies your policy: let it through, mask the sensitive data, route it to a model hosted in the EU, or block it.
How can I check that these principles are met?
With a pilot in audit mode: for about ten days the system records what it would have done without blocking anything, and the final report can be verified against your own data. Afterwards, the evidence log stays exportable for audits.
What is recorded about each conversation?
Who, when, with which model, which data categories were detected, which action was applied and what it cost. The content of messages is not part of that record: it is stored encrypted on each person's device and Acotia does not log it (the gateway processes it in memory). More in trust.