Mid-Market P&C Carrier · Insurance · Chicago, IL · 14 weeks
Fraud detection model deployed to production
A mid-market property and casualty insurer was manually reviewing every claim for fraud, using 12 full-time employees but still missing about $6 million in fraud each year. We implemented a real-time fraud detection model that scored all claims instantly. This cut the claims needing manual review from 18 of every 100 to about 5, and uncovered $6.2 million in suspicious claims for further investigation.
Challenge
The carrier depended on adjusters’ gut feelings and an outdated set of 23 static rules last updated four years ago. These rules flagged 18 of every 100 claims, causing a backlog that slowed down valid payments and annoyed customers. At the same time, more complex fraud slipped through undetected because the rules didn’t catch new patterns.
Approach
We developed a fraud propensity model based on four years of claims data, including adjuster notes, claimant behavior, and network connections between involved parties. The model scores each claim at submission and places it into one of four review categories: auto-approve, standard review, enhanced review, or refer to SIU. This approach replaced the previous 23 static rules with a flexible, data-driven scoring system.
Outcome
We cut the claims needing manual review from 18 of every 100 to about 5 of every 100, freeing up team capacity. SIU referrals rose from 3 to 18 of every 1,000, with 74 of every 100 of those cases confirmed as suspicious. In the first year, we flagged $6.2 million in fraudulent claims. At the same time, we sped up legitimate claim payments, cutting the cycle from 8.4 days down to 5.1.
How We Dug Into Network Data and Caught Fraud Rings in Action
Alright, here’s the scoop: these fraud rings were pretty slick. They filed claims under a bunch of different names but kept using the same addresses, phone numbers, or providers. We built a network graph linking all the claimants and ran some graph analytics on it. That’s how we caught suspicious clusters that the usual rules completely missed.
We put together an easy, step-by-step review to spot mistakes and keep everyone in the loop.
We dropped the basic approve-or-reject step and moved to a four-level review using fraud risk scores. If a claim seemed low risk, it bypassed the adjusters and got paid immediately. That way, the adjusters could zero in on the tricky or suspicious claims that actually needed their attention.
Results
- $6.2M Suspicious claims identified annually
- 18 to 5 of 100 Claims requiring manual review
- 74 of 100 SIU referral confirmation rate
- 8.4 to 5.1 days Legitimate claim payment cycle
We used to manually check 18% of all claims and still missed $6 million in fraud. Thinklytics built a model that narrows that down to just 5%, but it catches more fraud than we ever did before. Now our adjusters spend their time on the cases that actually need it.