The human ceiling, measured first
Two of your engineers independently code the same subsample. That agreement rate is the ceiling, and every figure we quote afterwards is expressed against it.
06 How we're measured
We publish the ceiling before we quote the number.
No system can be more accurate than the coding it is graded against. So the first thing we measure is not our model — it is how often two of your own warranty engineers, shown the same claim, assign the same failure cause.
You own the measuring instrument
The held-out claim set and scoring harness are handed over as your property, with a written right to re-run them against any vendor — including us.
The failure envelope is published
We name the claim types we will not adjudicate — photo-only evidence, code-mixed narratives, contradictory DTC and text — and how often each occurs in your data.
A named switch-off threshold
Written into the contract: the metric at which we tell you to turn the agent off, and the reversion runbook for doing it.
Escrow from signature
Weights, adapters, scorer, container and runbook in escrow, released on insolvency, two consecutive quarters below floor, or exit for convenience.
The clock starts on access
Timelines begin at claim data access, environment sign-off and a scoped service account — not at signature. DMS integration is our scope, not your homework.
04 Return
What changes when every record is read.
Three outcomes our customers report. Each is stated against the before-state it replaced — and where we do not yet have a number, we say so rather than round one up.
Up to 3% additional fraud detected
In claims already processed — surfaced by cross-claim and historical intelligence that rule-based screening cannot reach.
Baseline: fraud already caught by existing controls. Counted only where the customer’s own investigators confirmed the finding.
Root cause on failed parts, far sooner
Safety and reliability teams reach RCA on a failed part much faster, because the observations, inspection reports and service history behind it are already read, classified and connected.
Baseline: investigation assembled by hand, one case at a time. We do not yet publish a figure for this — it is measured per customer during the Full Read.
Adjudication report in 1 hour
The report that governs whether claims are paid is produced within the hour, on the full population rather than a sample.
Measured before-state: one week.
07 Plain speech
Things you will not read on this website.
Every one of these is said by nearly every vendor in this category, which is precisely why none of them tells you anything.
That phrase belongs to national programmes and cloud regions, not to an application vendor. We will show you an egress diagram and the steps your security team runs to verify it.
Deployment was never the bottleneck. Claim data access, DMS integration and security review are. We put a start gate in the contract rather than a range in a headline.
A savings figure with no stated baseline is arithmetic, not evidence. We measure your current inspected share and cost per claim first.
On whose claims, on which field, at what sample size, against what agreement rate between your own engineers? Without those five things the percentage is decoration.
This is the opening line of almost every deck in the category. If our work is real, the Coverage Check shows it and the sentence is unnecessary.
Someone else usually can. We would rather be measured than be the only one making a claim that is checkable in a single search.
— Start here
Your closed claims are an answer key. Let us be graded against them.
Send us nothing. We send you one file, your own IT runs it offline against a single export of closed claims, and you keep what it finds either way.