PZU

Motor insurance claims processing is one of the most document-intensive workflows in the auto insurance industry. At PZU, Poland's largest insurer, every motor damage claim required a human adjuster to manually review photo documentation, evaluate damage severity across individual car parts, and decide whether each component qualified for repair or replacement. With high claim volumes, this sequential, labor-intensive approach created processing bottlenecks and constrained throughput — scaling output meant scaling headcount. The status quo imposed a hard ceiling on efficiency and introduced inconsistency in adjuster decisions, increasing both operational cost and settlement cycle times.

PZU partnered with a European AI startup to pilot a computer vision system purpose-built for motor damage assessment. Rather than building proprietary models from scratch, PZU leveraged the vendor's existing training infrastructure and supplemented it with over 20,000 historical motor damage cases from PZU's own claims portfolio. The resulting model can analyze submitted photo documentation, identify specific vehicle components, classify damage extent, and produce a repair-versus-replacement recommendation — all within 30 seconds per claim. The system was integrated into PZU's existing claims workflow, with a quality-gate mechanism routing cases that meet photo-quality thresholds through automated processing and flagging the remainder for human adjuster review.

PZU reported a tenfold increase in process efficiency for the automated segment of its motor claims workflow — the first deployment of this technology at scale by any insurer in Poland. Key outcomes include:

• 90% of submitted documentation meeting quality requirements is now processed automatically, with no human adjuster intervention required
• 30-second analysis time per claim, compared to a manual review process measured in minutes or hours
• The remaining 10% of cases — those with insufficient photo quality or atypical damage patterns — are escalated to experienced adjusters, preserving human judgment where it adds the most value

The shift reduced bottlenecks and allowed adjuster capacity to be redirected toward complex, high-value claims.

• 90% automation is achievable in routine damage assessment when the AI is trained on a sufficiently large, domain-specific dataset — 20,000+ real claims proved adequate for production-grade accuracy.
• Photo quality is the primary failure mode: building a quality-gate that routes low-quality submissions to human review is essential to maintaining system reliability.
• Startup partnerships accelerate time-to-production for incumbents — PZU avoided the multi-year build cycle of in-house AI development by leveraging an external vendor's existing model infrastructure.
• Human-in-the-loop design matters: reserving the hardest 10% of cases for adjusters ensures edge cases don't erode customer trust in the overall claims process.