Study Shows AI Spots Suspicious Claims Earlier Than Traditional Approaches

Machine learning models detect suspicious claims two weeks after submission–much faster than traditional methods, according to a new CLARA Analytics study on fraud detection in property/casualty insurance claims.

The research, completed in November 2024, analyzed 2,867 claims from 2020 to 2024 using an unsupervised machine learning approach.

The study found that 9% of open claims were identified as high potential for special investigation unit referral. Michigan and Arizona showed the highest percentages of potential fraud indicators.

The model’s predictions closely matched actual SIU referrals made by adjusters but detected potential cases significantly earlier–as soon as two weeks after the first notice of loss.

Network analysis revealed important connections between attorneys and medical providers that traditional methods might miss.

The findings suggest that cohort modeling across claim development periods can effectively identify cost and treatment outliers while mapping connections between providers and attorneys that may indicate fraudulent activity.

“This research represents a significant advancement in how the insurance industry can approach fraud detection,” said Pragatee Dhakal, director of Claims Solutions at CLARA Analytics. “By leveraging advanced analytics, we’ve shown that insurers can identify potential fraud much earlier in the claims process, potentially saving billions in fraudulent payout.”

The study also highlighted the importance of the “Sentinel Effect,” where the awareness of being monitored leads to improved behavior. Insurers known for effective fraud detection are less likely to be targeted, offering a preventive advantage extending beyond direct cost savings.

“What’s particularly promising about this approach is that it doesn’t rely on pre-established fraud indicators,” Dhakal added. “By using unsupervised learning techniques, the system can potentially identify novel patterns of fraudulent activity that might not match historical cases.”

The researchers employed cohort modeling across claim development periods and mapped frequency connections to providers and attorneys. The method allows a more comprehensive view of potential fraud patterns than traditional indicator-based approaches alone.

The findings could transform how insurers approach fraud detection, combining human expertise with sophisticated analytics to create more effective prevention systems, the researchers said.

Topics InsurTech Claims Data Driven Artificial Intelligence