AI that cannot invent a number.
It reads your documents and it understands your question. It never does the arithmetic, because a governed engine does that every time.
The usual arrangement.
Letting a model answer questions directly against a database produces fast, confident, subtly different answers to the same question. Two people ask it slightly differently and get two numbers, both of which look authoritative, and neither of which anyone can check.
What is actually there.
The assistant writes the question, not the answer
Ask something in plain English and the model can only produce a report definition: the same subject, grouping, measures and filters a person would have chosen by hand. Then it stops.
A governed engine does the maths
Every figure is calculated by the product's own reporting engine, the same code path as a report built by hand. A mistake shows up as a visibly wrong question, never as a silently wrong number.
Money removed for those who shouldn't see it
Rate and margin fields are stripped for users without rate visibility, applied after the model has spoken so it holds even if the model ignores its instructions.
The product holds no AI vendor key
Not one, of any kind. All AI features call a separate gateway, so the product could be inspected end to end and contain no AI credential to steal or run up a bill on.
What it means on a Tuesday.
Someone asks what a lane made last quarter. They get a real report, built by the same engine that builds every other report, and they can see the definition it used. If the question was wrong, the definition looks wrong on screen, which is the failure you want rather than a plausible number that quietly informs a pricing decision.
Where this claim stops.
The AI features depend on an external gateway being configured and reachable. If it is not set up, document reading, document classification and plain-English questions are simply unavailable, and nothing else in the product depends on them. No core operational function stops working.