Leading US-based Insurer (unnamed)

In Property & Casualty insurance, fraud in high-volume claims segments like Connected Living represents a significant and growing financial exposure. This insurer processed over 350,000 lifestyle claims monthly, yet its legacy fraud detection system flagged only 6.2% of claims as potentially risky — leaving the vast majority of suspicious activity unexamined. More damaging still, 74% of those flagged claims were false positives: legitimate customers subjected to unnecessary scrutiny, triage teams buried in low-value reviews, and investigative capacity diverted from genuine fraud. The dual failure — missing real fraud while overwhelming reviewers with noise — was eroding both operational efficiency and customer satisfaction at scale.

The insurer partnered with Zensar to design and deploy a real-time ML-based fraud scoring system on Microsoft Azure, integrating Databricks for distributed model training, Cosmos DB for low-latency data access, MLFlow for experiment tracking and model registry, and a serverless architecture for cost-efficient scaling. The solution applies predictive machine learning across 60+ engineered features — including behavioral signals, device metadata, and IP address patterns — to generate a fraud risk score at the point of claim intake. Explainable AI surfaces the top contributing factors for each flagged claim, preserving reviewer autonomy and accelerating triage decisions. An MLOps pipeline with drift monitoring ensures model performance remains stable as claim patterns evolve. The initial deployment handles 15,000 claims per day with architecture designed to scale to 100,000 per day without re-engineering.

The production deployment delivered material improvements across every key metric:

• False positive rate reduced by up to 90% — dramatically cutting unnecessary reviews of legitimate claims
• Review load cut by 50% — freeing triage capacity to focus on higher-confidence fraud signals
• Precision improved from 12% to 57% — meaning flagged claims are now far more likely to be genuine fraud
• Recall improved from 55% to 61% — capturing a larger share of actual fraudulent claims

Beyond the numbers, customer satisfaction improved as fewer legitimate claimants experienced friction from unwarranted scrutiny. The triage team gained actionable context through explainability features, accelerating investigation workflows.

• False positive reduction deserves equal priority to detection rate — in high-volume P&C operations, alert fatigue and CSAT degradation from over-flagging are measurable business costs.
• Explainable AI is a prerequisite for reviewer adoption — fraud analysts need to understand why a claim was flagged to act on it confidently.
• Design for scale from day one — architecting for 100K claims/day while running at 15K prevents costly re-platforming as volume grows.
• Feature engineering across behavioral and device signals unlocks fraud patterns that transactional data alone cannot surface.
• MLOps with drift monitoring is non-negotiable — fraud patterns shift continuously, and a model without monitoring degrades silently in production.