.png)
A mid-market P&C carrier was losing its best examiners to routine files it never needed them for. Working with Tezo, it rebuilt claims intake and decisioning so straightforward losses now close on their own, and complex ones reach an adjuster with the file already assembled.
A carrier whose growth had outpaced its claims operation.
A mid-market property and casualty carrier writing personal auto, homeowners and small commercial across 11 US states, with around $600M in direct written premium and 240,000 claims a year. Its 180 adjusters and examiners worked across two service centres on a mainstream core platform, while loss notices arrived through a call centre, an agent portal, an app, scanned mail and a shared inbox. Nothing in claims had been automated before.
Loss notices came through five channels and left the same way: emailed ACORD forms, estimates in two vendor formats, medical bills, phone photos, police reports scanned at 200 dpi. An examiner spent the first half-hour of every claim reconstructing it, and nothing was queryable until someone had typed it in.
Two adjusters looking at comparable property-damage losses reached materially different reserve and settlement positions, depending on caseload, tenure and the day of the week. The file recorded what had been decided and never the reasoning behind it, which is a fair-claims-practice exposure as much as an economic one.
A $1,400 windshield claim and a $140,000 bodily-injury claim entered the same queue and competed for the same reviewer. The carrier’s most experienced examiners spent roughly 40% of the week on files requiring no judgement at all, while the files that genuinely needed them waited in line behind those.
Suspicious patterns were caught at reserve review or at the point of payment, weeks after both the money and the goodwill had already moved. SIU received a high volume of undifferentiated referrals with no way to rank one against another, so genuine cases sat in the same pile as the noise around them.
Market-conduct exam responses and DOI complaint replies meant reconstructing the decision rationale for a claim from free-text adjuster notes. A single exam response consumed weeks of manual file pulls, and the answer was only ever as good as whatever somebody had thought to write down at the time it was handled.
The first six weeks produced no software. One quarter of closed claims was instrumented and segmented by document mix, severity, channel and touch count, to establish which claims were straight-through eligible and which never would be. About a third qualified, lower than the executive expectation and higher than the adjusters’. Nothing was trained until claims leadership signed it off.
Email, portal uploads, agent submissions, mobile photos and scanned mail were routed into a single intake service that classified, deduplicated, scanned and attached every artefact to the correct claim number before a human saw it. The shared inbox stayed open for six months as a fallback, then closed on evidence that nothing was falling through, rather than on a date.
Extraction models were trained on the carrier’s actual document set rather than a generic corpus: ACORD loss notices, CCC and Mitchell estimates, medical bills, police reports and handwritten statements. Confidence thresholds were tuned field by field, so a low-confidence value escalated to a person instead of quietly entering the claim file and being trusted downstream.
Every structured claim is now scored on complexity, severity trajectory and fraud signal, with an ISO ClaimSearch check at first notice rather than at reserve review. Routing stopped depending on who was on shift. A claim of a given profile goes to the same queue at the same priority every time, and claims operations can see and change that behaviour without a release.
Settlement recommendations for straightforward claims come from a decision layer where claims operations author and edit the rules themselves, in a format they can read. The model contributes a scored view with its top contributing factors exposed alongside it. Payment authority limits are enforced as hard rules, never as model outputs, so nothing can recommend past its authority.
Adjusters approve, adjust or escalate inside the core claims system they already use, and no claim pays without a human authorisation event. Every override is captured as labelled signal and fed back, which is why accuracy improves rather than plateaus. Each automated decision writes an append-only record: inputs, rule and model version, score, factors, human action, timestamp.
Adjusters open the file that actually needed them. Roughly 19% of claims-handling hours moved from routine processing back to complex work, without a change in headcount.
Every automated decision carries its own evidence, so answering a market-conduct request or a DOI complaint starts from a record rather than from a reconstruction.
What this project changed about how we approach claims work.
Straight-through eligibility is smaller than executives assume and larger than adjusters do. Measuring it first turned a contested business case into an agreed one, and it is now the opening phase of every claims engagement we run.
Extraction accuracy is a document-set problem before it is a model problem. Generic models performed well on clean forms and failed on the two things that mattered most here: handwritten statements and low-resolution scans.
Rules the business can read get used. Rules only engineering can change get worked around. Handing rule authorship to claims operations did more for adoption than any accuracy improvement we shipped.
Override data is the compounding asset. The system improves only because every adjuster correction is captured as labelled signal, which means the review interface matters as much as the model.
The audit record was scoped as a compliance obligation and turned into the adoption argument. Adjusters trusted recommendations they could interrogate, and stopped trusting the ones they could not.
Tell us where the work is stuck. We will tell you what we would do, what it would realistically take, and whether we are the right partner for it.
Vasavi Sky City, Telecom Nagar, Gachibowli, Hyderabad, Telangana 500032
