Insurance claim photo triage

Business value

• Reduces straight-through-processing leakage on claims with manipulated evidence. Even a small percentage of synthetic-evidence claims, multiplied by claim values, is meaningful loss avoidance.
• Front-loads SIU referrals on claims with visible manipulation evidence, so investigators focus on the claims most likely to warrant referral.
• Adds an audit trail to the photo-evidence step — evidence categories become part of the claim file for adjuster reference and regulatory documentation.

Agent job to be done

Be a claims triage layer. Allow claim photos that read as authentic damage documentation. Hold photos with manipulation cues for adjuster review. Refer high-confidence manipulation cases to SIU, but never claim the score is fraud evidence on its own.

format: otherintended_use: citedomain: insurance claim photo triage

When to call VeracityAPI

On photo intake at first notice of loss (FNOL), and again on any photo replacement during the claim lifecycle.

What image URL to submit

Public HTTPS image URL for each claim photo. Score each photo independently; for a multi-photo claim, aggregate the evidence at the claim level.

Decision policy

• allow: low synthetic_image_risk on all photos, no composite cues, EXIF metadata where expected.
• revise: medium risk on one or more photos — request additional photos from different angles, or photos with a held-up timestamp/insured-info card.
• human_review: high risk on any photo, OR claim value above $10K with any medium-risk evidence, OR pattern-match against known fraud rings in the carrier's history.
• SIU referral: high risk combined with claim-handling red flags (rapid total-loss claim on a new policy, prior-claim history, ownership-document inconsistencies). The score is one signal in the SIU intake form, not a conclusion.

Automation recipe

• Claims intake worker receives the photo bundle from the claimant's FNOL submission.
• Score each photo with intended_use=cite (the highest threshold — photos used as evidence should be held to a strict standard).
• Allow-tagged photos flow to the assigned adjuster's queue.
• Held photos route to specialist review with evidence categories pre-populated in the case file.
• Pattern aggregation: track high-risk-photo patterns across claimants, IPs, and shop networks for fraud-ring detection.

Evidence spans agents should inspect

• 'synthetic_damage' — damage that appears generated rather than photographed (impossible patterns, inconsistent debris physics)
• 'composite_seam' — damaged components pasted into the broader vehicle/property context
• 'metadata_inconsistency' — photo metadata that doesn't match the claimed event timeline or location
• 'recycled_imagery' — photos that appear in other claim files or in public salvage databases

Policy pseudocode

if (result.recommended_action === "allow") continueWorkflow();
if (result.recommended_action === "revise") rewriteWith(result.evidence, result.recommended_fixes);
if (result.recommended_action === "human_review") queueForHumanReview(result);
if (result.recommended_action === "reject") discardOrRebuild();

KPIs to track

• % of claims with a held photo at FNOL
• SIU referral conversion rate (referrals that result in investigation findings)
• claim-payout leakage attributable to manipulated evidence (the metric the gate exists to reduce)
• adjuster average handle time on clean claims (gate should not slow legitimate claims)
• regulatory-audit score on the photo-evidence step

What can go wrong

• The score is not fraud evidence. Never communicate a 'high risk' result as a fraud accusation to the claimant — it's an internal triage signal that routes the claim to a human.
• Phone-camera EXIF is often missing or scrambled. The absence of expected metadata is a signal to verify by other means, not proof of manipulation.
• State-specific insurance regulations govern claim-denial reasoning. The gate's output should never be the sole stated reason for a denial; it's one input the adjuster considers.

Cost and latency notes

Image analysis is a flat $0.02 per image. The endpoint accepts HTTPS image URLs, stores no image bytes, and logs only a URL hash plus hostname. Current v0.1 latency is vision-model-bound, so preflight balance and retry carefully.