Security & compliance · EU AI Act
The assistant tells people AI is answering, shows where every answer came from, and is contractually kept away from the uses the Act calls high-risk. Your deployer duties arrive already discharged.
The Act asks three things of a system like this: that people know they are talking to an AI, that the organisation deploying it keeps its people AI-literate, and that high-risk uses are kept out. Each one is a mechanism here, not a memo.
Every conversational surface — web chat, sign-in, invitations, messaging channels, the CLI — discloses that answers are AI-generated and cited. Not a setting that can be switched off.
The in-product help centre serves as the deployer's literacy resource: how answers are produced, what citations mean, what the assistant refuses and why.
Deployments are contractually excluded from high-risk uses; the product is a knowledge engine over business documents, and the terms keep it that.
Answers cite their sources or decline; nothing executes actions. A human reads, verifies against the passage, and decides.
Disclose: the surface says AI is answering. Ground: only the asker's permitted documents feed the answer. Cite: numbered sources, or an honest “I don't have that.” Defer: no actions executed — a human verifies and decides. Oversight is the workflow, not an add-on.
See it in the demo ›
A knowledge engine over your own documents: it reads them, grounds every answer in them, cites the source, and answers. It does not profile people, score candidates, evaluate employees, or trigger actions in other systems. Where a corpus contains people, GDPR mechanisms — export, erasure, records — run underneath every answer. And the boundary is contractual as well as technical: the terms exclude Annex III uses, so the system your counsel reviewed is the system your teams get.
The terms ›
The position paper is readable in the trust library beside the architecture overview it depends on — request forms never gate the reading. Signed copies, the answered questionnaire and the architecture record land on your own thread from a standing account request, the day each exists.
Read the position ›
The position and the records it rests on, readable right here — the EU AI Act measures, the architecture record, the GDPR mapping and the data-flow note.
Transparency by construction: AI-interaction disclosure on every surface, an AI-literacy resource in the product, high-risk uses contractually excluded.
One client, one deployment, one boundary, what that means concretely: containers, database, credentials, the AI boundary and the audit chain.
Each client runs their own complete instance: application, worker, database and document store in isolated containers on a machine that serves no other client. There is no shared platform, no pooled database, no cross-tenant anything, isolation is the deployment model, not a configuration option.
Indexing (reading and preparing documents for search) runs entirely inside the deployment. Generation follows the deployment's configuration: the fullest posture is a self-hosted deployment that answers entirely on its own hardware, where nothing leaves the boundary at all. Deployments using the Naxis AI service send the question and the permission-filtered excerpts (zero-retention terms, see sub-processors); the model never sees documents the asking person cannot see, because permission filtering happens before the request leaves.
In every configuration: your data is not used to train models, and nothing is retained after the answer returns.
Access is group-based and settled before a question is even considered, everywhere the assistant answers, never in any one surface. Accounts activate by invitation link; administrators never see or set a password. Channel guests receive nothing until groups are explicitly opened to them.
Every consequential event, questions, answers with their citations, administrative actions, syncs, erasures, lands on an append-only, hash-chained log. Verification recomputes the whole chain on demand from the console; a broken link is impossible to hide. Rows cannot be edited or deleted; retention prunes whole aged spans and records that it did.
Releases are versioned images; deployments install them in a quiet window after a safety backup and roll back on their own if the health check fails. What changed in each release is shown in the console in plain language.
An account action, the request files under your Naxis account, and the answer arrives on its thread.
The rights articles as product mechanisms: export, erasure, records of processing, each one built in, none of it paperwork after the fact.
Every Naxis deployment is a single-tenant system processing only the client's own documents, on infrastructure the client chooses. GDPR compliance is implemented as product mechanisms, not policies:
The generated manifest for a specific deployment is available to its administrator at any time; a sample is in Technical documentation below.
An account action, the request files under your Naxis account, and the answer arrives on its thread.
What leaves the deployment, what never does, and where things physically live.
An account action, the request files under your Naxis account, and the answer arrives on its thread.
Signed copies and the answered questionnaire land on your own thread from a standing account request; the full trust library carries the rest.
Yes, as a deployer you carry transparency and AI-literacy duties. Naxis Assistant is built so those duties are already discharged in the product: every surface discloses the AI interaction (Article 50), and the in-product help centre serves as the literacy resource Article 4 expects.
Answering questions over business documents is not an Annex III high-risk use, and deployments are contractually excluded from being turned into one. The product answers questions; it does not score people, filter candidates, or decide anything about persons.
Every conversational surface says so: web chat, sign-in, invitations, messaging channels, even the CLI. Disclosure is not a setting that can be switched off.
The assistant cites its sources or declines, it never executes actions. A human reads the answer, opens the cited passage, verifies, and decides. Oversight is the workflow, not an add-on.
The position paper is readable in the trust library, and the architecture overview it depends on can be requested as a signed document from your account.
Every paste into a consumer chatbot is a question your own systems could not answer fast enough. Security vendors sell detection, which reads the app and never the question. What the 2026 breach data actually shows, how far Microsoft's new Shadow AI controls really reach, and why the only fix that scales is making the company answerable.
Glean is Google for your company: it finds where things are stored. Naxis is the private knowledge engine that knows what is actually true right now. Different machines, different questions, different prices. An honest way to choose, and the Glean alternative for companies of 5 to 200 people.
Lidl wrote off €500 million trying. Birmingham City Council is £144 million in and reinstalling from scratch. Consolidation megaprojects fail for a structural reason: real work is too fluid to hold one shape. The part you actually wanted, one place to ask, never needed the migration.
