AI use statement

What the assistant does, what it can see, what it never does, and how to tell an AI answer from a document. Last updated 2026-09-18. 1 item marked for completion.

DocuDriver includes an assistant that answers questions about a workspace's own documentation. This statement says plainly what it does, what it sees, what it never does, and how you can tell its answers apart from the documents themselves. It is a draft awaiting legal review.

What the assistant does

It answers questions such as "which reports use the Enrollment fact table" or "what does the Retention Rate measure mean" by finding the passages in the workspace's documentation that bear on the question and composing an answer from them. Every answer names the documents it came from, and nothing is cited that the person asking cannot open themselves.

It also proposes dictionary definitions for the terms it finds in documentation. Those are proposals: a person in the workspace reviews each one and approves, edits, or rejects it, and only approved definitions are treated as the workspace's dictionary.

What it can see

Only what the person asking can see. Before any passage reaches the model, DocuDriver filters the workspace's documents by that person's role, the grants on each document, and the hierarchy the workspace has set up. A document the person may not open is not consulted, is not cited, and is not revealed to exist. Two people asking the same question in the same workspace can get different answers, because they can see different documents.

The assistant sees documentation: report definitions, data model files, SQL, packages, notebooks, and Office documents, and the metadata extracted from them. It has no connection to the databases that documentation describes and no access to the data in them.

What it never does

It never answers from outside the workspace's documents. When the documentation does not cover a question well enough, it says so and stops, rather than filling the gap from general knowledge. The threshold for "well enough" is fixed by DocuDriver, not by the person asking.

It never extracts or stores personal data from the systems the documents describe, because DocuDriver never reads those systems.

It never uses a workspace's documents, questions, or answers to train any model. Requests go to the model provider under commercial terms that forbid training on the content, and DocuDriver keeps nothing from a request beyond the answer shown, which stays in the workspace with its citations.

It never acts. It does not change documents, definitions, memberships, or settings; every change in DocuDriver is made by a person and appears in the audit log under that person's name.

How to tell an AI answer from a document

Answers appear in the assistant panel marked Assistant, with the model that wrote them named beside the answer and the documents it drew on listed under it. Documents and approved dictionary entries are shown as themselves, with the person who uploaded, wrote, or approved them and when. A proposed definition the assistant wrote is marked as proposed until a person approves it, and the approver's name replaces the mark.

Which model is used

To be confirmed at beta: the platform's default model provider, the model family, and the terms under which requests are made. The intended default is a commercial provider whose terms exclude training on API content; the provider in force is shown to workspace owners under Workspace settings, Assistant model.

A workspace owner may instead connect the organisation's own model account (currently Anthropic). Requests then go to that account under the organisation's own agreement with the provider, DocuDriver stores the key encrypted and never displays it again, and the owner can test the connection and remove the key at any time. Answers written by the organisation's own model are labelled with that model's name.

Limits worth knowing

The assistant is only as good as the documentation in the workspace. Missing, outdated, or contradictory documents produce cautious answers or declines, which is the intended behaviour, and the Review queue is the place to fix the documents. Answers can still be wrong: they should be read as a guided pointer into the documentation, with the citations checked, not as a substitute for it.

Questions

Questions about this statement go to the support address on the Help and support page.