What the AI actually does
One job: reading fields out of documents. When you upload a client email, a decision notice or a set of proposed plans, Incursus sends them to an AI model and asks for specific values — the client’s name, the site address, the planning reference, the condition number, the drawing references.
That is all it does. The report itself is assembled by ordinary software from a template and from the fields you have confirmed. Root protection areas, incursion percentages and tree categories are calculated arithmetically from your survey data — no AI is involved in any figure.
Who it goes to
Anthropic, in the United States. We use their Claude model through their commercial API.
Anthropic may not train their models on what we send. Their commercial terms say so directly, and they state that you keep all rights to what is sent and own what comes back. They hold it briefly to check for misuse and then delete it.
We do not use your documents to train anything either, and we have no plans to. If that ever changed it would be something you switched on deliberately, off by default, and we would ask first.
What is not sent
- Your tree schedule
- Your CAD drawing
- Your account details or password
- Anything from another practice
- Any report Incursus has generated
Who checks it
You do. Every value the model returns lands in a field you can see and edit before it reaches a report. Where Incursus is unsure, it says so rather than guessing — and where it could not verify something against the page it was reading, it leaves the field empty instead of inventing a plausible answer.
A model can misread a scanned page, and the report will look exactly as confident either way. Check every figure before a report leaves your hands.
This is not a disclaimer for its own sake. The realistic failure is not nonsense — it is a reference that is almost right, on a document that otherwise reads perfectly.