AI document generation service

The situation

Some documents can only be produced by someone who knows the domain. Combining a week’s worth of individual work scopes into a single consolidated request is one of them: the rules are real, they are specific, and they live in the heads of a few experienced people. That makes the document slow to produce and impossible to delegate.

Traditional automation cannot help much here, because the difficulty is judgement rather than arithmetic.

What we built

A web service where anyone on the team uploads the week’s scope workbooks, adds plain-English instructions if there is anything unusual, and gets the finished document back.

Behind it, a language model reads the workbooks directly and applies a written specification of the domain rules, the same rules the senior planner would apply, now written down once and executed consistently.

Why it works

The critical design decision was to keep the rules as an explicit, editable document rather than burying them in code or in a prompt nobody can find. When the rules change, a planner edits the rules. That keeps the expertise with the team that owns it, and keeps the system honest about what it is doing.

The result is used by colleagues who are not the person who built it, which is the only real test of whether an internal tool works.