// Claims professional
Adjusters hate these tools. The data says they're right.
The most negative occupation in the analysis about AI is the one that has been handed the most of it. So this page starts where the interviews start — the named failure modes, in adjusters' own words — and every build below is answering one of them.
// The situation
The highest negative rate of any occupation analyzed.
98% of Glassdoor reviews mentioning AI, written by claims adjusters, were negative — the highest rate of any occupation in the analysis.
Glassdoor review analysis · reviews from June 2025 to May 2026.
A build that argues with that number will fail the same way the last one did.
So the first thing we build on a claims file is the check-path, not the summary: every extracted fact carries the page it came from, and a named reviewer with a written runbook exists before go-live rather than after it.
// The stack you actually run
Three hundred pages, a smudged fax, and a deadline.
58% of frontline claims professionals spend more than 20% of their time on manual data entry and compliance.
Verisk · May 2026 · vendor research.
And the work is getting harder. Bodily injury now takes a larger share of the loss dollar than physical damage does, which puts medical records and demand packages at the center of the file.
The named failure modes, from the interviews. Misclassifying claims at FNOL. Faulty claim summaries. Hallucinating from smudged or low-quality scans. Omitting details from medical reports, leading to inaccurate payouts.
“You often have to spend 15 to 20 minutes correcting AI-generated mistakes because supervisors don’t verify anything.”
Claims adjuster · Glassdoor review · 2026.
“There’s an AI fatigue. We’re pretty much exhausted as far as the amount of AI being shoved down our throats. AI is just a tool. It should never be given the keys.”
Geoffrey Conrad, adjuster · TechSpot · September 2026.
Read those two together and you have the design brief. The correction tax is real, it lands on the adjuster, and nobody upstream owns it.
// The build, differently
Four rules, each answering a named failure mode.
Every extracted fact cites its page
A summary you cannot verify is a summary you have to redo. Diagnoses, dates, amounts, treatment gaps and liability indicators each link to the source page in the record, so checking costs one click instead of re-reading the file.
Surface: the document store and the claim file in ClaimCenter, with page-level citations written back.
The summary is a starting point, never a decision
No reserve set automatically. No coverage position generated. The system assembles and surfaces; the adjuster decides, and the record shows who decided and on what basis.
Surface: claim notes and the decision record, with the human decision-maker named on every write.
Low-confidence output is labeled, not smoothed
Smudged scans and poor-quality faxes are where hallucination happens. Those pages are flagged as low confidence and routed for a human read rather than silently guessed at. The confidence thresholds are yours to own.
Surface: the extraction pipeline, with a routed exception queue your supervisors control.
Coverage-to-claim check at intake
Policy terms compared against the claim facts when the file opens, not after your team has worked it, so reserves get set on the right basis the first time. It also catches FNOL misclassification before it propagates.
Surface: the policy record at FNOL, read through the policy system's sanctioned API.
What we measure
Correction rate on AI output, tracked as a first-class metric and reported to the people doing the correcting. Minutes per file spent verifying. Reserve development on files where the coverage check fired against files where it did not. Rate of low-confidence pages routed and how many the reviewer overturned.
// We ran one of these
Adjusters do not adopt a tool they cannot check. The correction tax on a bad summary lands on the same desk, which is why the check-path is scoped first.
The operating record →// What's hard about this
Two things we tell you in week two.
The unstaffed seat is QA.
Insurance operating-role postings barely mention exception handling, escalation or QA, which are the exact skills a deployment creates demand for. If you deploy an extraction system and nobody owns reviewing its output, you have moved work rather than removed it. The posting data →
Automation buys capacity. It does not buy judgment.
Hiring is being cut at the junior end of claims while senior postings hold, so the pipeline thins at exactly the moment the surviving work gets harder. So the scope names which files never route to a thin bench and what the system must show a senior before a reserve moves.
// What ships with it
The file has to survive litigation.
Plaintiff firms are adopting generative AI of their own, which means the demand package arriving on your desk is itself AI-accelerated. The answer is not a faster summary. It is a file where every assembled fact points at its page, every decision names the human who made it, and every low-confidence read is on the record as low confidence.
What ships: a model inventory entry with risk tier, data lineage covering each document source, pre-deployment testing results, drift thresholds with remediation triggers, and a human-override specification naming the decision-maker.
NAIC and state bulletins · current at September 2026.
What the governance file contains →The check-path gets built first.
If the output cannot be verified in one click, it will be verified in fifteen minutes, and the fifteen minutes will land on you. That is the whole argument, and it is the first thing we build.