How this system works

A draft. Not a published communication, not a decision, and not advice. Someone in the organisation reads every draft, edits it, and decides whether it is used at all.

1. Someone picks one of 18 task types, such as a press release or an FOI response, and describes what they need.
2. The system searches that organisation’s own uploaded documents for relevant material: past examples, house style guides, standing text and policy evidence.
3. That material, the request, and a quality checklist for the task are sent to an AI model, which writes the draft.
4. The draft comes back with numbered citations pointing at the documents it drew on. Before it is saved, the system checks that every citation points at a source that was actually retrieved.
5. A person reviews it.

The system searches only the organisation’s own uploaded documents. It does not search the open web, and it cannot reach any other organisation’s material.

• Provider: Mistral AI, a company based in France. It is the only AI provider this system can reach.
• Writing model: mistral-small-latest
• Search model: mistral-embed, used to find relevant passages in the organisation’s documents
• Database: in Frankfurt, Germany

One part of the system, the application server, is still hosted outside the EU while a move to European infrastructure is completed. Until that finishes the deployment is restricted to sample material rather than real departmental content. This page will be updated when it changes.

Every generation is written to a log that no user can alter. Users can read their own entries and administrators can read their organisation’s, but the log can only be added to, never edited or deleted, and only the server can write to it.

For each draft generated, the log records:

• Who asked for it, and when
• The task type and the title
• The AI model and provider that produced it
• How much text went in and came out, and how long it took
• A fingerprint of the exact instruction sent to the model, so it can be shown later that the instruction has not been changed
• Every document passage that was retrieved and offered to the model, identified individually
• Whether the response was cut short

Edits and approvals are recorded too, including a marker that a human made the change.

Every draft moves through three states: Draft, then In review, then Approved. A person moves it, and each move is recorded.

The system has no way to publish anything. There are no integrations with websites, social media accounts, mailing lists or press distribution services. Text leaves this system only when a person copies it out. That is a deliberate design decision, not a feature that has not been built yet.

• Publish, send, or post anything
• Make a decision about any person
• Read another organisation’s documents or drafts
• Search the open web
• Use one organisation’s material to train a model

AI models can state things confidently that are not correct, and this one is no exception. In testing, one draft in eighteen cited a source that did not exist to support a figure it had invented. That is why citations are checked against the retrieved sources before a draft is saved, and it is the clearest illustration of why a person reviews every draft.

The checks catch citations pointing at sources that were never retrieved. They cannot tell whether a correctly cited source actually supports the sentence. Only a reader can.

If something this system helped produce has affected you, our complaints and appeal process is at /complaints.