Sovereign AI on your data, without letting it leave
The question is not whether AI is useful. It is where the documents you feed it end up, and who can reach them.
In short
Omniscient installs in three ways: shared, dedicated and managed, or entirely on your own infrastructure, which is also called an on-premise or self-hosted installation. In that last mode no data leaves your network, the language model included. The product is identical in all three cases, only the operator and the location of the data change.
Why do internal AI projects stall?
An internal AI project almost always runs into the same wall. Legal asks where the data is processed. IT asks what leaves the network. Nobody has a satisfying answer, and the project stays an experiment.
The objection is legitimate. Handing your document estate to a service whose location and terms of use you do not control is a heavy decision, and it commits far more than the team that started the trial.
What are the three installation modes?
Omniscient installs in three ways, and the choice is yours.
| Shared | Dedicated and managed | On-premise | |
|---|---|---|---|
| Where the data lives | With us, in Europe | With us, in Europe | With you |
| Who operates it | Us | Us | You |
| Isolation | By marking, applied on every read | Dedicated installation | Total, network included |
| Language model | Hosted service | Hosted service, or yours | Yours, on your machines |
| Time to start | Fastest | In between | Longest |
| Who it suits | Getting going quickly | Isolation without operations | Data that cannot leave |
The product is the same in all three cases. What changes is who operates it, and where the data lives.
On your own infrastructure, the platform runs on your machines, inside your network. Nothing leaves your premises. This is the mode chosen by organisations whose data cannot cross the perimeter, for regulatory or contractual reasons. The European AI regulation adds obligations of its own, which bear less on where the data sits than on how the tool is used: see what the AI Act asks of those who deploy an AI.
What does isolation mean technically?
On a shared installation, isolation is not a commercial promise. Every piece of data carries the organisation it belongs to, and that marking is applied on every read, in the graph as in the search index. A query cannot cross that boundary, because it is never built without it.
Permissions, before anything else
Sovereignty does not stop at geography. Data hosted in Europe but readable by the whole company is still a problem.
Authentication is fully handled by the platform. Accounts are managed end to end, sessions are time-limited and renewed by token, passwords can be changed and their rotation enforced. None of this calls for a third-party component.
Filtering applies during the search. An answer cannot rest on a document its reader may not consult, because that document never enters the scope being queried.
Rules are fine-grained, covering entities, relations and documents, per role and per action. Sharing is decided object by object: private, named people and roles, or the whole organisation.
Credentials for connected systems follow the same logic. A connection is declared once, then referenced by the runs that need it. The secret is never readable back: no response from the platform returns a password, not even to the person who entered it.
What is traced?
A verifiable installation means being able to answer "who saw what, and when".
Every run leaves a usable trace: what ran, what failed and why. Failures are grouped by cause, logs are searchable. The audit log can be exported, which matters on the day it has to be produced to a third party, an auditor or a supervisory authority.
How do you connect to our directory?
Everything above covers the platform's own authentication and permissions, which is to say what works from the moment it is installed, with nothing added.
Connecting to your directory is a different matter. Single sign-on (SSO), LDAP and multi-factor authentication are not options you tick: they are built specifically for your information system, as part of the installation. The reason is plain, and it beats a generic promise. A corporate directory, its groups and its mapping rules never quite look like the one next door.
In practice, without that work a user signs in with an Omniscient account. With it, they sign in the way they already do with your other applications, and your directory's groups feed the platform's roles.
It is one of the things the scoping conversation is there to settle, and it is better raised early than late.
Can you choose the language model?
The platform calls a language model to understand questions, extract entities and write answers. That model is a setting, not a given.
On an installation under your control, it can run on your own infrastructure. That is the configuration which guarantees no text leaves your network, during processing included. It is also what separates a genuinely sovereign platform from a service that merely keeps its servers in Europe while calling a model somewhere else.
How do you judge the sovereignty of a solution?
The right way to assess the sovereignty of a platform is not to read a page, it is to state your constraints and see whether they hold.
Required location, closed network or not, an authorised third party, sector obligations: these determine the installation mode, and therefore the conversation.
Once that frame is set, the rest of the platform does not change: the knowledge graph, search across your documents and the automations work the same way in all three modes.
Tell us yours, and we will say plainly what is possible.