AI Claim Triage From Photos: First Notice of Loss to a Reviewed Assessment

An adjuster opens claim 20-8841 on Tuesday. The policyholder has uploaded forty-one photos from their phone: a flooded living room shot from four angles, a hallway ceiling, a close-up of a baseboard, three that are almost certainly the same wall. In another window sits a 62-page policy PDF. The job is to turn that into a defensible damage assessment with coverage checked. It takes hours, it cannot start until the claim reaches the top of a queue, and by the time it does, the policyholder has called twice.

This is the transformation we would run for that operation. It is presented as a concept, built on the same method we use in client engagements, so you can see exactly what the before, the after, and the path between them look like.

What does the manual claims desk actually cost?

The cost is not one dramatic number, it is the queue. J.D. Power’s 2025 U.S. Property Claims Satisfaction Study, released in March 2025 and based on 5,178 homeowner customers who had filed a claim in the previous nine months, put the average property claim at 32.4 days from filing to finished repairs and more than 44 days from first notice of loss to final payment: the longest since the study began in 2008. The same study found claims completed within 10 days scored 762 on its 1,000-point satisfaction scale, dropping 167 points to 595 when repairs ran past 31 days.

The honest reading of those numbers matters more than the numbers. Most of that elapsed time is physical: contractors, materials, and scheduling, none of which software touches. What software can touch is the desk time, the stretch where a claim waits in a queue and then consumes an adjuster’s afternoon in sorting, measuring, and policy lookup. That is a real slice of the calendar, and it is the only slice this work claims.

What does the claims desk look like before and after?

Drag the handle. The before is the recreated manual desk; the after is the triage copilot designed in its place.

The assessment is still an adjuster’s, checked against the same policy. What disappears is the sorting, the scrolling back through forty-one photos for each line, and the hours a claim spends waiting for someone to start.

What we built

The concept above is not a mockup exercise. It is the output of the same engagement steps we run on real companies:

• We map one claim type first, usually the most common water or wind loss, because it repeats constantly and its desk time is measurable against a baseline the operation already tracks.
• We build the extraction against the operation’s own closed claims: photos, adjuster notes, and final assessments become the evaluation set, so accuracy is scored on this book of business rather than on a vendor demo.
• We require the model to cite its image: every drafted line names the photo it was read from, and the reviewer can open that crop. A line the copilot cannot tie to a specific image does not get drafted.
• We ground coverage in the policy document rather than in the model’s memory, quoting the clause and page so the adjuster is reading their own policy language, not a paraphrase of it.
• We deliberately left settlement and fraud referral un-automated. The copilot assembles evidence and quotes clauses; deciding coverage, valuing the loss, and referring suspected fraud stay with the adjuster, because those calls carry legal and regulatory weight that a confidence score does not discharge.
• We designed the flag before the automation: anything ambiguous, like a ceiling stain whose source is not visible in any photo, is surfaced as a question rather than filled in with a plausible guess.

What that costs the operation is mostly one person’s attention, not a budget line. Expect roughly a day of a senior adjuster’s time to pick the closed claims that become the evaluation set and settle what a correct assessment looks like on each, a few hours from whoever administers the claims system to arrange read access to photos and policies, and about two hours a week from the same adjuster during the pilot to review flagged lines and correct drafts. Those corrections are not overhead, they are the eval data that decides whether the copilot earns wider scope.

How does it work?

Photos and the policy arrive the way they already do. The copilot groups images by room and damage type, drafts damage observations with a confidence score on each field, and cites the specific photo behind every line. It checks the loss against the policy and quotes the governing clause with its page number. The adjuster opens a claim where the sorting is done, the obvious lines are ready to confirm, and their attention goes to the flagged ones: the ambiguous ceiling stain, the photo that does not match the reported loss.

The failure mode worth naming is the one vision has and text does not: a misread field looks exactly as confident as a correct one, and there is no sentence to quote back as proof. That is why the design pins every line to its source image and puts an adjuster in front of the assessment before it counts. When the copilot is wrong, the cost is a corrected line in a review queue, and that correction becomes an eval case that decides whether the copilot has earned wider scope.

The operation’s data stays in its environment, the claims management system remains the system of record, and nothing is used to train anyone else’s models.

Why does this pay back?

For a claims operation the return lands on the two things the business is judged on. Cycle time improves at the only point software can improve it, by removing the days a claim spends waiting for an adjuster to have an afternoon free. And the assessment that results is better evidenced than the one it replaces: every line traceable to an image, every coverage call quoting the clause it rests on, which is worth as much in a dispute as it is in a fast month. The first slice is also a wedge, because the same extraction engine extends from water losses to wind, fire, and contents claims, one measured claim type at a time.

The honest limit: this pays back where photo-heavy, high-volume, relatively standard losses dominate the book. An operation whose claims are mostly large, complex, or litigated will find the copilot organizing evidence and little more, because the expensive judgment in those files was never the sorting. If policy documents are inconsistent across a legacy book, the first weeks are document work before they are AI. And an operation unwilling to keep an adjuster in the approval loop should not run this design at all: the whole safety argument depends on that gate.