Trend — AI in claims
AI in Property Claims: Where Carrier Automation Helps Public Adjusters and Where It Hurts
Carrier automation is accurate on measurement and shaky on damage. The gap is where the recovery lives.
A hailstorm crosses Denver on a Tuesday. By Thursday the carrier has pulled an aerial image of the roof, run it through a damage-detection model, and issued a scope: 14 squares, one slope, $9,400 replacement cost value. The public adjuster walks the same roof on Saturday, finds bruising on three slopes the model scored as "no anomaly," and writes the file at $31,000. Same roof, same date of loss, a $21,600 gap that turned on whether a machine or a person looked at the north elevation.
That gap is the argument over AI in property claims, compressed into one file. Carriers stopped quietly piloting this technology a while ago. By 2025, algorithmic triage, aerial damage scoring, and automated estimate assembly were running on a large share of first-party property claims before a human adjuster ever opened the file. For the public adjuster on the other side, the question stopped being whether to believe the marketing and became something practical: which parts of this actually move a file faster, and which parts cost your client money if you let them stand unchallenged.
What the carrier's model is actually doing
Strip the branding off carrier-side property AI and most of it does three jobs. It measures, it scores, and it assembles.
Measurement is the part that works. Aerial and drone-imagery vendors have spent two decades building roof geometry from overhead photos, and the numbers are good. EagleView reports its measurements validated to 98.77% accuracy against an independent benchmark, and its sketches drop straight into Xactimate. If a desk adjuster says your client's roof is 22 squares at a 6:12 pitch, arguing the geometry is usually a losing fight. The tape measure lost that one already.
Scoring is where it gets slippery. A measurement model tells you how big the roof is; a damage model claims to tell you whether it is damaged, and how badly. Those are different problems with different error rates. Vendors have pushed hard into the second one — EagleView's Assess drone product and its 2026 Horizon engine rank properties by "damage likelihood" so a claims manager can decide which roofs to send a human to first. That is reasonable as triage. It becomes a problem when the score quietly hardens into the scope.
Assembly is the third job: turning a measurement and a damage call into a line-item estimate with depreciation already applied. It is fast and consistent, and it is only ever as right as the two steps feeding it. An assembly engine cannot know the damage model skipped the north slope; it just prices what it was handed.
Where the automation genuinely helps you
The tools the carrier runs are, for the most part, the same tools available to you. That symmetry is the part worth holding onto.
Pulling federal weather data and an aerial measurement for the date and address of loss takes minutes and hands you a defensible starting point you did not have to climb for. First-pass document sorting — reading the policy, flagging endorsements, surfacing the prior-claims history — is real time back in your day. When a model tags every photograph to a slope or elevation, your evidence package is organized before you write a sentence. None of that is theater. It is the unglamorous infrastructure of a faster file, and a public adjuster who refuses to touch it is climbing roofs the hard way for no reason.
The honest read is that automation compresses the parts of claim work that were always mechanical. It does not compress judgment, and it has never once walked a roof.
Where it hurts, and how to push back
The damage concentrates in three places.
First, false negatives. A damage model tuned to cut false positives — which vendors optimize for openly, because carriers do not want to pay for non-damage — will systematically miss soft hail bruising, wind-lifted shingles that reseated, and anything sitting under a tarp. Overhead imagery also flattens the very evidence that wins these files: a granule-loss pattern that reads plainly to an adjuster standing on the slope can vanish into roof texture at altitude, and a model has no test square to pull. Every missed slope is an underpayment the homeowner never notices, because the estimate in front of them looks finished, line-itemed, and signed.
Second, opacity. When a desk adjuster works inside a dollar range the software handed them, they often cannot tell you why the number is the number. This is the same dynamic that made algorithmic claims engines like Colossus contested in bodily-injury work for years: the adjuster plays inside the sandbox the model draws and calls it a settlement. Ask for the basis of a property scope and you can get an answer no human is actually able to reconstruct.
Third, speed cuts both ways. An automated denial or lowball lands in days, formatted and confident, and an unrepresented homeowner reads "the inspection found no covered damage" as the end of the road instead of the start of one.
The leverage hiding in the regulation: By 2025, roughly two dozen states had adopted the NAIC Model Bulletin on the Use of AI Systems by Insurers, which requires insurers to maintain a documented AI program — with validation, testing, and named accountability — for decisions made in regulated processes like claims handling. Regulators are openly demanding explainable systems. That is a paper trail you are allowed to ask about.
The leverage a public adjuster still has
Start with the thing the model cannot do: you stood on the roof. A demand that reads "the carrier's aerial scope recorded no anomaly on the north and west slopes; attached are 14 dated photographs of hail bruising on those slopes, with chalk markings and a moisture reading" does not argue with the algorithm. It goes around it, by entering evidence the model never had access to. You are not litigating the score. You are making it irrelevant.
The regulatory posture helps here. When a scope reads like it fell out of a black box, you can demand a human reinspection and put the program question on the record: what model produced this estimate, what was it validated against, and who signed off on it. A federal court in 2025 allowed discovery into an insurer's use of AI to deny claims, which tells you those questions carry weight beyond a strongly worded letter.
And the clocks do not care who scoped the file. Statutory prompt-payment and acknowledgment deadlines run from the dates written into the statute, not from the moment a carrier's software emitted a number. The interest and the penalties attach the same way whether a person or a model wrote the first estimate. claimOS for public adjusters is built around that reality — every photo, weather pull, and carrier message writes back to one claim record, so when you challenge an AI scope your counter-evidence is already dated, sourced, and addressable instead of scattered across a phone and three folders. If you are weighing how the major systems handle that kind of file, how the platforms compare is a reasonable place to start.
The carrier automated the first pass. It has not automated the appeal, and that is the half of the file where the recovery actually lives.
| What the AI does | Where it earns its keep | Where it costs your client |
|---|---|---|
| Roof measurement | Fast, accurate geometry; ends pointless square-count disputes | Rarely — measurement is the mature part |
| Damage scoring | Triages which roofs to inspect first after a storm | Misses soft hail, reseated shingles, tarped areas; tuned to suppress false positives |
| Estimate assembly | Consistent line items and depreciation in minutes | Inherits every error from the damage score; looks complete when it is not |
| Automated denial | Clears genuinely non-covered files quickly | Lands fast and final-sounding; unrepresented owners stop reading there |
Can a carrier deny my client's claim using only AI?
It depends on the state. Most states that adopted the NAIC Model Bulletin require insurers to keep a documented program with human accountability for AI used in claims decisions, and several legislatures have floated bills requiring human review of AI-generated denials. In practice you can demand a human reinspection and ask, on the record, which model produced the scope and what it was validated against.
Is the carrier's aerial roof measurement worth fighting?
Usually not. Measurement models report accuracy near 99% against independent benchmarks, and the sketches feed Xactimate directly. Spend your energy on the damage call rather than the square count — that is where the disputed money actually sits.
Why does an AI scope miss hail damage my inspection found?
Damage models are tuned to reduce false positives because carriers do not want to pay for non-damage. That tuning makes them skip subtle losses: soft bruising, wind-lifted shingles that reseated, and anything under a tarp. A dated, photo-documented, slope-by-slope inspection is the counter that the model cannot dispute.
How do I challenge an estimate that came from an algorithm?
Do not argue with the model on its own terms. Enter evidence it never had: dated photographs tied to specific slopes, moisture readings, a contractor's repair scope, and the federal weather data for the date of loss. Then put the explainability question in writing — what produced this number, and who is accountable for it.
Does carrier automation change the statutory deadlines on my file?
No. Prompt-payment and acknowledgment deadlines run from the dates set in your state's statute, not from when the carrier's software issued a scope. The clock and any statutory interest run the same way whether a human or a model wrote the first estimate.
Sources cited
- Model Bulletin on the Use of AI Systems by Insurers— National Association of Insurance Commissioners
- Nearly Half of States Have Now Adopted NAIC Model Bulletin on Insurers' Use of AI— Quarles & Brady
- Court Allows Discovery Into Insurer's Use of AI to Deny Claims— Hunton Andrews Kurth
- When Algorithms Underwrite: Insurance Regulators Demanding Explainable AI Systems— Buchanan Ingersoll & Rooney
- EagleView One Rolls Out Full 3D Coverage for Insurance Adjusters and Roofers— Claims Journal
Run your water claims on a system built for the file, not the spreadsheet
claimOS is the operating system for property-claims professionals — intake, evidence, communications, copilot drafting, and an audit-defensible record on one screen.