Computer Vision Development in London

Insurance / Property and Motor Claims

A London-based insurance group (GBP 8.5B GWP, processing 2.1 million motor and property claims per year, 12 million policyholders across UK personal and commercial lines) needed computer vision AI to automate damage assessment, accelerate claims processing, and improve fraud detection for motor and property claims. (1) Motor claims volume: 1.4 million motor claims per year — each requiring damage assessment from photographs submitted by policyholders, garages, and loss adjusters. Assessment was performed by 380 motor claims handlers who reviewed photos, estimated repair costs, and determined whether to repair or write off the vehicle. Average assessment time: 35 minutes per claim. The 8-photo average per claim meant handlers reviewed 11.2 million photos per year. (2) Property claims assessment: 700,000 property claims per year — weather damage (storms, flooding), escape of water, fire, theft, and subsidence. Property damage assessment was more complex than motor — requiring understanding of UK building construction types (Victorian, Edwardian, post-war, modern), building materials, and repair methods. Loss adjusters visited 40% of property claims (280,000 site visits per year at GBP 350 average cost per visit). (3) Reserve accuracy: initial reserves were set based on claims handler assessment of damage photos and descriptions — but varied significantly between handlers (coefficient of variation: 28% for motor, 35% for property). Inaccurate initial reserves affected financial reporting and reinsurance recovery calculations. Development factors (how reserves changed over time as the true cost emerged) were higher than competitors, indicating systematic inaccuracy. (4) Fraud in visual claims: the group's fraud team estimated 6-8% of motor claims involved some level of fraud — staged accidents, exaggerated damage, pre-existing damage claimed as new, and claims for vehicles not matching the policy. Visual fraud detection was largely manual — experienced fraud investigators reviewing photos for inconsistencies. The fraud team could only review 5% of claims, meaning most fraudulent claims went undetected. (5) Customer experience: claimants submitted photos through the mobile app or email — but received no immediate feedback on their claim. The average time from photo submission to initial damage assessment was 3.2 days for motor and 5.8 days for property. Competitors offering faster assessment were attracting younger customers who expected immediate digital responses.

We developed and deployed a claims damage assessment platform over 14 weeks. (1) Motor damage assessment: deep learning model trained on 2.8 million annotated motor claim images (the insurer's 5-year claims history). Damage detection: identifying damaged components (bonnet, bumper, wing, door, wheel, glass, lights) with 94% component-level accuracy. Severity classification: categorising damage as minor (cosmetic), moderate (structural component damaged but repairable), and severe (potential write-off) with 91% accuracy. Repair cost estimation: predicting repair cost ranges based on damaged components, severity, and vehicle make/model — using pricing data from 2,400 approved repair network garages. The model produced estimates within 15% of actual repair cost for 82% of claims (comparable to experienced claims handler accuracy). Write-off assessment: automated total loss assessment for clearly written-off vehicles — identifying cases where repair cost would exceed the vehicle's market value. (2) Property damage assessment: models trained on 1.2 million property claim images covering UK building types and damage categories. Damage classification: identifying damage type (water, fire, storm, impact, subsidence) and affected building elements (roof, walls, windows, flooring, ceilings, electrics). UK building knowledge: models trained to recognise UK-specific construction — slate vs. tile roofing, cavity wall vs. solid wall construction, period features vs. modern materials. Severity estimation: property damage severity scoring (1-5 scale) correlating with historical repair costs for similar damage on similar properties. Loss adjuster triage: automated determination of whether a site visit was required — simple claims (clear damage, below threshold) processed from photos only, complex claims flagged for loss adjuster visit with pre-populated assessment to guide the inspection. (3) Reserve automation: ML-based reserve estimation combining visual damage assessment with claim metadata (policy type, location, vehicle age, property age). Motor reserves: automated initial reserve setting for standard claims — within 8% of ultimate claim cost (from 28% variation under manual assessment). Property reserves: automated initial reserve for assessed claims — within 12% of ultimate cost (from 35% variation). Confidence scoring: each reserve accompanied by a confidence score — high-confidence reserves applied automatically, low-confidence reserves reviewed by experienced handlers. (4) Visual fraud detection: image forensics — detecting manipulated photos (clone stamping, splicing, metadata inconsistencies indicating photos from different times/locations than claimed). Damage pattern analysis: identifying pre-existing damage (rust around damaged areas suggesting old damage, weathering inconsistent with claimed date), staged damage patterns (damage patterns inconsistent with claimed cause — e.g., "parking damage" on areas inconsistent with another vehicle's contact), and repeat patterns (linking photos across claims to detect organised fraud — the same vehicle, location, or damage pattern appearing in multiple claims). Fraud scoring: every claim scored for fraud probability — high-risk claims automatically escalated to the fraud investigation team with specific visual evidence flagged. (5) Customer instant assessment: real-time damage assessment through the mobile app — claimants received preliminary damage assessment within 60 seconds of uploading photos. Guided photo capture: the app guided claimants to capture the right photos (4 corners, close-ups of damage, odometer, VIN plate) — improving photo quality and reducing the need for re-requests. Instant next steps: based on the AI assessment, claimants immediately received next steps — repair network referral (with nearest approved garage), write-off indication (with valuation range), or loss adjuster scheduling (for claims requiring site visit).

Motor assessment: automated damage assessment for 72% of motor claims (the remaining 28% — complex damage, disputed liability, commercial vehicles — routed to specialist handlers with AI pre-assessment). Assessment time: from 35 minutes to 3 minutes for AI-processed claims. Accuracy: repair cost estimates within 15% of actual for 82% of claims. Motor claims team: from 380 to 180 handlers — 200 redeployed to complex claims, fraud investigation, and customer service roles. Property assessment: 45% of property claims assessed from photos without loss adjuster visit (from 60% requiring visits). Loss adjuster visits reduced by 35% — GBP 34M annual savings. Property assessment time from 5.8 days to 1.2 days for photo-assessed claims. Reserve accuracy: motor reserve development factor improved from 1.18 to 1.06 (closer to 1.0 = more accurate initial reserves). Property reserve development from 1.25 to 1.10. Improved reserve accuracy released GBP 120M in over-reserved capital — generating GBP 6M in annual investment income. Fraud detection: visual fraud detection rate increased from 2.1% of claims to 5.8% — identifying GBP 45M in additional fraudulent claims annually. Organised fraud: visual pattern linking identified 8 previously undetected fraud rings in the first year. Customer experience: instant assessment adoption reached 65% within 12 months. Customer satisfaction with claims: NPS from +8 to +38. Mobile app claim submissions increased from 35% to 72% — claimants preferring the instant-feedback digital journey. Annual value: GBP 45M (fraud detection — additional fraudulent claims identified through visual AI), GBP 34M (loss adjuster reduction — fewer property site visits required), GBP 28M (claims handler efficiency — 200 FTE redeployment at GBP 42K fully loaded cost plus overhead), GBP 15M (customer retention — improved NPS driving 1.8% retention improvement on GBP 8.5B GWP), GBP 6M (reserve accuracy — investment income on released over-reserved capital) = GBP 128M. Development cost: GBP 450,000. Payback: under 2 days.