Friday, November 6th, 2026 — Insurance fraud investigators may soon have a new way to analyze the enormous volume of digital evidence collected during complex cases. Researchers have developed an artificial intelligence system that connects information from mobile devices, financial transactions and other records to help investigators identify suspicious activity and relationships among suspects.
The system, called SAGE, was developed by researchers Weilong Ding, Haoyang He, Han Tian and Jie Guo. According to findings published Oct. 29 in Mobile Networks and Applications, the technology was evaluated using two real-world health insurance fraud datasets and deployed in a practical investigative project. During that deployment, researchers reported processing more than 1.34 million forensic records and identifying 11,200 pieces of fraud evidence.
SAGE uses multiple large language model agents working together to analyze different types of digital information. Individual agents handle tasks such as classifying images, transcribing voice messages and extracting information from behavioral activity logs. A central coordinating agent manages the process, producing organized records that investigators can examine chronologically.
Another feature, called the Unique Identifier Network, helps investigators connect individuals and other entities appearing across different devices and databases. The system combines information from multiple sources, including financial transactions and logistics records, to establish relationships that might otherwise remain hidden in separate sets of evidence.
The research focuses particularly on health insurance fraud involving prescription drug trafficking, where investigators may need to examine communications, transaction histories and other electronic records to reconstruct suspected criminal activity.
For insurance carriers and special investigation units, the technology offers a potential method for reviewing digital evidence that extends beyond traditional text-based analysis. Connecting communications with financial activity could help investigators identify patterns, examine suspect relationships and assemble more complete timelines of potentially fraudulent activity.
The researchers reported successful testing on two real-world datasets and an operational deployment. However, the publicly available research abstract does not specify the false-positive rate for the 11,200 evidence items or establish whether those findings resulted in confirmed fraud determinations.
The SAGE source code and illustrative sample data are publicly available on GitHub, although the real investigative datasets remain restricted because of privacy and legal requirements.
