Interface, agents, models and infrastructure sit inside one managed layer, with the secure gateway as the only connection to the outside world. Information comes in without personal data leaking out, and the administrator decides whether external AI may be approached anonymously.
Localign runs in one of two execution modes, and your organisation chooses which. Mode A is the default and allows no external providers at all. Every model that touches your data runs only on European infrastructure operated by us or by a vetted EU-based subprocessor, so in this mode your data does not leave the EU and there is nothing to configure. Mode B is opt-in and is where you select which external providers to use. An organisation administrator turns it on and picks exactly which providers are available: you can allow one, several, or any combination you want, choosing whichever providers fit your needs. Even once Mode B is enabled, each prompt that would leave the EU triggers a consent dialog with the user and a sensitive-data scan before anything is forwarded. Providers that are not enabled for the organisation are blocked explicitly rather than silently rerouted. Switching back to Mode A takes one click and is effective immediately.
Conversations, uploaded documents and account data are stored in the EU. The default language and reasoning models are hosted on European GPU infrastructure. Conversation state, PII mapping and orchestrator state are stored in the backend, while document binaries live in object storage and are served back through expiring presigned URLs when needed. The full processor list and hosting locations live on the Sub-processors page. We do not store your data with providers outside that list.
By default, identifiable data is processed exclusively within the EU, and external AI providers only ever see Localign as the requester, never your organisation's identity. The platform is designed around Schrems II, the EU ruling that protects against US surveillance law reaching your data. Organisations with stricter requirements can also deploy Localign on-premise or fully air-gapped.
Your browser keeps a short-lived access JWT and opens `ws/chat` only after the client has refreshed the session if needed. The backend authenticates the user, applies the organisation's AI policy and quota rules, and passes trace, conversation, assistant and project context into the AI engine. The AI engine runs routing and sensitive-data checks on the local secure path first, decides whether the question needs internal knowledge (your documents, our help centre) or external lookup, and only then dispatches the work: to EU-hosted models by default, or to an external model under Mode B if the user has consented for that prompt and that provider.
Document retrieval stays on the internal path. When a question needs files, the AI engine calls the document MCP with explicit scope headers such as conversation id, project id, attached document ids and assistant-library document ids. The backend still enforces the caller's access rights; the MCP only narrows what the model may search or summarise. Raw documents are not sent to an external model.
External providers receive only the filtered question text plus the selected provider and model settings for that turn. Conversation history, raw documents and raw images do not travel down the external path. The external-query guard also deduplicates requests, caps them to a small number of provider calls per turn, and blocks unavailable providers instead of quietly substituting another one.
Localign does not use your prompts, your documents or your conversations to train, fine-tune or improve any AI model, neither our own nor a subprocessor's. The same prohibition is mirrored in every Mode B contract with an external model provider, and is restated in the AI Annex you sign with us. In the routing and tool-execution path, documents are treated as internal-only content rather than material that can be forwarded to an external provider.
If your tab disconnects mid-answer, processing continues server-side and the conversation state is restored when you come back. If a primary model is unavailable, the request falls back along a pre-declared chain of equivalent models, never silently to a different region or a different consent posture. Long-running streams emit heartbeats, and document requests that hit still-indexing files time out cleanly with a retry instruction instead of fabricated output. Operational events that affect customers are published on the Service status page; documentation changes are listed on Updates.
We usually answer questions from CISOs, DPOs and procurement teams within two working days.
