Resolution Audit — Verify Your AI Support Bill

It's the same relationship you already have with any metered supplier — you meter the meter. Most engagements find the billing broadly right and a specific pattern wrong, like conversations closed by a deflection that the customer immediately re-opened under a new ticket. That's a conversation with your vendor about definitions, not an accusation, and having the transcripts to hand is what makes it a short conversation.

Because we publish ours, and right now what we'd be publishing to you is an uncalibrated judge. The software ships a Calibration screen showing our agreement against a human-labelled sample with the confidence interval on it, per verdict class, including the classes we're bad at — and the product refuses to report a number as a judge accuracy when no model produced it. Being straight about where that stands today: the judge panel is built and its two-pass-plus-tiebreak logic is tested, but it has not been scored against a model on our side, so the screen currently reads "does not clear our 90% bar". Calibrating it against your labelled tickets is the first thing week one does, and if it doesn't clear the bar you shouldn't buy the second month.

The parts that don't need a model. Given correct verdicts the pipeline recovers 86% of a known planted misbilling with under 2% false disputes; the re-contact tracker runs at a 0.3% false-dispute rate, having been rebuilt four times because the first version confidently linked 218 tickets of which 3 were real; re-running an audit costs nothing because every verdict is cached on the transcript and the rubric version. All of it on synthetic data — no real support estate has been audited and no dispute has ever been filed. The full ledger ships in the repo as docs/VALIDATION.md and we'll send it before you buy.

Read-only access to your helpdesk's closed conversations and a copy of the vendor invoice — nothing that can change a ticket, and nothing that touches your customers. A CSV export path exists for teams that would rather not connect an API at all, and it's the path we build against first on purpose: this product tells your vendor their invoice is wrong, and a design that only works through their API is a design they can switch off.

Yes, and here's exactly how, because it matters. Six research agents worked one domain each against a strict evidence rule: every demand claim had to come from an analyst report, a vendor's own pricing page, a dated funding round, public earnings, a named survey or a regulatory text — no forum posts, no vibes. That produced 41 candidates scored on demand evidence, contract value, how verifiable the gap was, buildability and defensibility; this one ranked third, and it's the third of the ten we've actually built. What that does NOT mean: that anyone has paid for it, or that we've audited a real invoice. The research says the problem is real. Only a pilot says the product works.