One task per engagement. Point it at a folder, define what you're looking for, and it reads every file it understands, extracting text locally, OCR'ing scans, and judging each document against your criteria.
The AI has already made a first pass. You verify it. No mark means accepted, cheap to move through at volume. Rejecting asks for the right tag, because a correction is a far more valuable training label than a bare rejection.
An export is a self-contained package: every document, the Excel index, and an HTML viewer with working relative links. It opens off a laptop or a USB drive with no server and no network, built for the courtroom, where wifi isn't a given.
Nothing is hardcoded. A CPA diligence engagement and a legal matter are just different criteria lists, you define the tags and what each one means, per task.