AI Adoption GuideInsuranceClaim
Computer vision damage assessment
Computer vision estimates repair costs and severity from policyholder-submitted photos or video for auto and property claims.
Insurance processQuoteUnderwriteBindIssueBillServiceRenewClaim
By Don, DoneThat’s AI coach · updated
A usable estimate cites the photo and the guide vintage
A computer vision damage assessment is a draft estimate, not a payment. Each line that proposes a part, a labor operation, or a severity grade should point at a specific photo or video frame and at the estimating guide vintage used for that line. If either cite is missing, the line is not ready for an adjuster to accept.
This is a quality outcome. An auto or property claims lead should open the file and see why a bumper cover, a quarter panel, or a roof square appeared, and which published labor and parts catalog the draft used that day. Empty stays empty when the image cannot support the line. An adjuster still owns the estimate. Nothing here auto-pays. Nothing here invents a part or a repair hour because the model usually sees that on this kind of impact.
Photo estimating sits after first notice, not instead of it. Intake and routing still decide whether the file is a photo-estimating candidate. See ai FNOL intake and claim triage for how photos get onto the claim.
CCC Intelligent Solutions, Tractable, Snapsheet, and Guidewire sit in this class of claims and estimating systems. Treat them as vendors that can run photo-to-estimate work under your rules. Do not assume any one of them cites the photo, stamps the guide vintage, or keeps an unpaid draft from becoming a check. Those controls are yours.
Load photos into the claim, not into a side channel
Load policyholder photos and video onto the claim record before the model runs. The file needs the same objects an adjuster would use: capture time if the device sent it, which side of the vehicle or which elevation of the building, and a stable photo ID the estimate can cite. A zip of images in an adjuster's inbox is not a load. A text-thread screenshot of a bumper is not a load.
For auto, collect the damaged area, a wide shot when the loss suggests related damage, and a VIN plate or door jamb if you need to confirm the vehicle. For property, collect the claimed area, a wider context shot, and a scale reference when the loss is a measurement problem, such as a crack, a hole, or a missing shingle. Video helps when stills miss sequence, a walk-around, or water travel. Still extract frames the estimate can cite. Do not let the model cite the video as a whole.
If the submission mixes usable and unusable images, keep all of them on the file. The unusable ones are evidence that the model should not have filled a line. They are not trash to delete so the estimate looks complete.
Fraud review is a parallel track, not a reason to skip loading. Duplicate photos, capture data that does not match the loss date, or a stock-looking interior can still sit on the claim while real-time fraud pattern detection scores the file. Vision estimating should not drop a photo because it looks inconvenient.
Estimate only what the image supports
Run the model only after photos are on the claim. For each proposed line, require two cites: the photo ID or frame that shows the damage, and the estimating guide vintage, including publisher, region, and effective date your shop or independent appraiser would use. Labor hours and part numbers come from that vintage, not from the model's memory of last week's similar files.
If the photo shows a creased rear bumper cover and a cracked lower grille, the draft can propose replace bumper cover and replace grille only if those operations are visible and the guide vintage lists them for that vehicle. It cannot add a radiator support because that impact usually pulls the core support. It cannot add R&I for a headlamp that is not in any frame. It cannot pick a part number for a trim level the photo does not establish.
Take a parked-car scrape with three photos. Photo A is a sharp, close shot of the right rear door with a long crease through the character line. Photo B is a wide rear three-quarter that is motion-blurred. Photo C is the interior cargo area, unrelated to the loss. The draft should cite Photo A for a door outer repair or replace decision, leave related quarter-panel and bumper lines blank because Photo B cannot support them, and ignore Photo C for estimating. An adjuster can still request a reshoot of the rear corner. The model does not invent the missing corner from Photo A.
The same rule holds for property. A roof photo taken from the driveway can support a note that granule loss is visible on the front slope. It cannot support a square count, a full-roof replace, or a specific underlayment line unless the image actually shows those quantities and materials. Hail files fail here often: dark, distant, or wet shingles look like impact to a model that is rewarded for filling the estimate.
When the file also has medical bills, attorney letters, or a recorded statement, do not fold those into the photo estimate. Damage from images is one work product. Narrative from PDFs is another. Keep medical and legal document synthesis on its own path so a synthesized injury summary never becomes a parts list.
Leave the line empty when the photo is unusable
Unusable means the image cannot support the line an estimator would write. Nighttime glare on a windshield, a thumb over the lens, a distant roof shot, a cropped bumper with no adjacent panels, a screenshot of a screenshot, or a photo of a different vehicle than the one on the policy: the correct output is blank, plus a reason the adjuster can act on, such as a reshoot request, a field inspection, or a later supplement.
Empty stays empty. Do not substitute a typical labor hour. Do not copy a line from a similar VIN in another claim. Do not round up severity so the file can close. A blank line that says Photo 4 is underexposed and the right front is not visible is a quality estimate. A complete-looking estimate with no photo cite is a quality failure, the first of the three that quietly corrupt a file.
If half the loss is visible and half is not, estimate the visible half and leave the rest blank. Partial completeness is the point. Completeness without cites is how silent overpay and silent underpay both enter the book.
The adjuster owns the estimate, not the payment queue
The adjuster, staff or independent, accepts, edits, or rejects every line. Computer vision does not own coverage, comparative negligence, betterment, or whether a bumper is repaired versus replaced under your guidelines. The draft is a structured suggestion with cites. The signed estimate is a human work product.
Do not wire this draft to straight-through claims payment. Straight-through payment is a different control problem: it assumes the amount is already fit to pay. A photo estimate is not fit to pay until an adjuster has owned it, and often not even then, given deductible, rental, total loss threshold, and subrogation. Treating the model's number as paid is the second failure mode.
If you later choose a narrow auto-pay path, design it as a separate control: eligible loss types, dollar caps, dual control, and an explicit payment event. It is not a toggle on the vision job.
The third failure is inventing a part or a repair hour. The model fills a radiator, a condenser, or a structural hour because the impact story is familiar. If it is not in the cited photo and not in the cited guide vintage for that operation, it does not exist on the draft. A short estimate and a reshoot beat a complete estimate that nobody can defend.
Keep the loop boring on purpose: load photos to the claim, estimate only with cites, leave blanks when the image cannot carry the line, and leave ownership with the adjuster. That is the quality bar for computer vision damage assessment.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first