LLM-powered Triage Agent Boosts Fraud Case Accuracy by 30.6%
New AI agent conducts multi-turn conversations to route banking fraud reports accurately.
Banks receive millions of fraud, scam, and dispute reports annually, making it difficult to direct customers to the right support teams. The existing manual process is slow and stressful. To address this, researchers built an LLM-powered triage agent that engages customers in natural, multi-turn conversations, asks relevant probing questions, and classifies cases for policy-guided routing. The agent is embedded directly in the customer journey, aiming to reduce friction and improve accuracy.
To evaluate the agent, the team created synthetic digital twins of real customers—simulated personas with realistic dialogue histories based on historical data. This allowed scalable testing across thousands of scenarios without compromising privacy. The agent integrates safety guardrails, reasoning frameworks, and policy compliance checks. Results showed a 30.6% improvement in classification accuracy compared to existing methods, with high satisfaction from subject-matter experts. The approach demonstrates how targeted probing can lead to more effective triage in large-scale banking operations.
- 30.6% increase in classification accuracy for fraud/dispute triage.
- Uses synthetic digital twins of real customers for scalable, privacy-preserving evaluation.
- LLM conducts multi-turn conversations with probing questions to route cases per bank policy.
Why It Matters
Automated fraud triage with higher accuracy reduces customer stress and speeds up resolution for banks.