Data residency, security and compliance you can verify.

No, unless you configure it to. Jean reads your systems in place through your own permissions, and the knowledge graph, embeddings and ledger are all generated and stored inside the environment you nominate. If you enable an external frontier model for a specific data class, the Governance Gateway masks the entities your policy names before anything is sent, and records exactly what went and what came back.

Nothing is answered. The Gateway fails closed by design, so an unavailable control is treated as a failed control rather than skipped. The outage itself is a ledger entry.

It inherits them and never re-implements them. Jean queries each source as the person asking, at row and document level, so a person cannot reach anything through Jean that they could not open directly. When a relevant source is withheld for that reason, the answer says a source was withheld rather than silently working around it.

Providers under United States jurisdiction can be compelled to disclose data regardless of where it is physically stored. MetisJean is a UAE entity, and Jean, the knowledge graph and the ledger run inside your perimeter or a region you control rather than in our tenancy. There is no copy of your intelligence in our hands to compel.

The split matters. Articles 12, 13, 14 and 18 sit on the provider of a high-risk system: the logging capability, the instructions for use, the design that makes oversight possible, and the ten-year retention of technical documentation. Article 26 is the deployer article, and it is the one that describes you. It asks you to use the system according to those instructions, to assign oversight to people with the authority and competence to exercise it, to keep the automatically generated logs for at least six months, and to inform workers before an AI system is used on them.

That is why the evidence has to be produced where you are: it is your use of the system that is being recorded, not ours. The Governance Gateway writes every decision to the immutable ledger inside your environment, pauses high-consequence actions for a named approver, and records what that person intended at the checkpoint rather than only that they had permission. Article 4 adds an AI literacy duty from February 2025, and Article 50 adds disclosure where people are interacting with an AI system. Nothing here is legal advice; your counsel decides how the Act applies to you.

From your own environment. Data classification is applied at ingestion, restricted classes are blocked from leaving the tenancy, and access, rectification, deletion and portability requests are served against the systems that already hold the data. We are not a processor holding a second copy.

Your administrators do, by data classification. Anything confidential or restricted stays on an in-perimeter model: a larger one for hard questions, a small one for the constant stream of small ones. Every routing decision is recorded, including the candidates that were not chosen and the reason.

The knowledge graph, the fine-tuned weights, the workflows you built and the whole audit ledger. They are already in your storage, so there is nothing to hand back. What you lose is our support and future releases, not the asset.

How the licence, updates and services are priced

Because you can open it. Every answer arrives with the documents it was drawn from and the path it took across the knowledge graph, so explainable AI here means traceable answers rather than a confidence score. Hallucination reduction comes from the same place: the model is answering from retrieved company facts, not from memory, and Synthia, our synthetic company with verified answer keys, is used to measure how often it gets a known-correct question right.

When Jean cannot support a claim, the honest failure is the designed one: it says which source it needed and did not have.

It removes the reason for it, which is the only thing that actually works against shadow AI. People reach for a public model because it is faster than the internal route; Jean has to be the faster route, on the same phone, with the company’s own context already loaded.

For the traffic that remains, the Governance Gateway is the control. Prompts leaving for an external model are inspected first, and confidential material is masked or the request is routed to an in-perimeter model instead. Either way the attempt is in the ledger, so you can see where the pressure is coming from rather than guessing.