Every rotation relearned where the answers lived. Now they just ask.

The outcome.

Operators ask questions in plain language and get answers from their records and policy documents, with the workbook, sheet or document page behind each answer.

It has been live since February 2025, runs entirely on-premise, and passed its technical evaluation.

The problem.

A staff cell answers questions under time pressure by reading. Fleet state and personnel postings sat in spreadsheets: more than 10,000 equipment records across some 3,000 vehicles. Policy for every kind of decision sat in 150 documents of about seventy pages each.

Duties rotate between units. The person answering next month is rarely the one who learned where the answer lives, so that learning was paid for again with every posting.

The data was sensitive and compartmented. Nothing could leave the perimeter, and not every operator could see every file.

What we built.

First, the records themselves. The workbooks were built for the eye: merged title rows, headers stacked deep, the same vehicle in several files with different fields. We structured them, with checks that propose a fix for every problem and a person approving each one before anything loads.

Then the assistant. One question box answers questions about the records and about procedure. When a question is ambiguous, it asks instead of guessing, and follow-up questions stay in context.

Access rules apply before the AI sees anything, so a document an operator may not see never reaches it. We sized the hardware by measurement, with spare capacity so one server can carry the load if another fails.

What changed.

About ten operators use it at one site. They get answers to the questions the project scoped, each with its source.

It passed its technical evaluation, where the bar was no more than 5% incorrect answers. Retrieval invented nothing, and every answer cited the correct source file.

What we were honest about.

Simple counts came back right, but some other totals did not, so that class of question was not signed off. Where a list ran across a page break, an answer could come back partial. Page numbers were occasionally off, though the file was always right.

So the assistant asks to be checked: it tells the operator to open the cited document and mark the answer complete or incomplete, rather than presenting a partial list as a whole one.