What does a first pilot for freight invoice mismatches look like?

A model lab's data lead wants to test whether generated freight exception material improves an agent's handling of invoice disputes. They meet MirrorFirm with a one-page brief.

The brief names one domain, logistics, and one record pair: carrier invoice against proof of delivery and rate card. It lists four root causes: wrong rate applied, duplicate invoice, short delivery, and an accessorial charge not in the contract. It states the roles (a freight audit analyst and a carrier relations lead), an approval limit, and the expected outcomes: correct, dispute, wait or escalate.

The sample returned contains a small set of cases such as `FRT-2041`, where carrier invoice `CINV-7710` bills a fuel surcharge that rate card `RC-0093` excludes for the lane. The lab's reviewers solve cases by hand, run structural checks on the tables, and test whether a trivial strategy scores well. They find the duplicate-invoice cases are too easy because the amounts match exactly, and ask for near-duplicates with changed dates and references.

That feedback shapes the next iteration: a revised sample, a held-out template set for evaluation, and a decision on whether to widen to a second domain. Licence terms for a larger release are agreed afterwards.

**What it teaches:** a pilot works when it covers a narrow slice and the acceptance checks are written before the sample arrives.