Skip to content

Case Studies Insurance

Cutting claims-fraud investigation time by 60%

A mid-market insurer partnered with Karsient to build a real-time fraud-scoring pipeline on top of a modernised claims data platform.

60% faster fraud triage35% reduction in false positivesReal-time scoring at claim intake

Client challenge

What was wrong with the legacy environment

Fraud signals scattered across claims, policy, and third-party data sources

Batch-only scoring meant fraud was often flagged after payout

Investigation teams relied on manual cross-referencing across systems

No feedback loop between investigator outcomes and the scoring model

Modernization approach

How Karsient approached the transformation

Karsient consolidated claims, policy, and external risk-signal data onto a governed Lakehouse, then built a real-time scoring pipeline that evaluates every claim at intake rather than after the fact.

Architecture

The modern target architecture

1

Streaming ingestion from claims and policy systems via CDC

2

Feature pipeline on Delta Lake feeding a real-time scoring endpoint

3

Unity Catalog governing access across claims, underwriting, and fraud teams

4

Investigator feedback loop written back to retrain the model on a schedule

Migration

Migration & re-engineering strategy

The legacy batch ETL was re-platformed incrementally, workload by workload, running in parallel with the existing nightly process until scoring parity was proven on historical claims.

Engineering improvements

What changed under the hood

Incremental, CDC-based ingestion replacing nightly batch loads

Delta Lake feature store shared across scoring and reporting

Automated data quality checks on every incoming claim

CI/CD pipeline for model retraining and safe rollout

Business impact

Measurable outcomes

Fraud triage time cut by 60%, from days to hours

False-positive rate reduced by 35%, freeing investigator capacity

Scoring now happens at claim intake instead of post-payout

A reusable feature store now powers two additional risk models

Technology

Technology used

Databricks
Delta Lake
Unity Catalog
Apache Kafka
Python

Want results like these?

Let's scope your next data or AI initiative