An Automated System for Generating Chargeback Notices Based on Multi-Agent Collaboration and Its Practical Evaluation

Le résumé fourni par la source

The advancement of large language models (LLMs) has propelled the application of agent systems in complex tasks. As a critical component for recovering funds in corporate payment risk control, chargeback defense has long relied on manual expertise, resulting in low efficiency and high costs. This paper proposes an automated defense letter generation system integrating a non-agentic data cleansing process with a multi-agent collaboration mechanism: at the base layer, structured cleansing unifies multi-source interaction data; at the upper layer, four agent types—Router, Policy, Defense, and Judge—collaborate to perform template selection, policy parsing, text generation, and automated quality inspection, thereby simulating real-world task division and enhancing generation stability. In real-world experiments on an online service platform, the system achieved a 29.2 % dispute success rate on a high-value subset covering 16.5% of all declined orders. This performance aligns closely with the historical baseline (29.68%) set by senior operations staff and significantly outperforms single-agent prototypes on cases traditionally abandoned by manual reviewers. Furthermore, through multi-threading and batch scheduling optimizations, end-to-end processing efficiency improved by approximately 5.3 times. The results demonstrate that the multi-agent architecture possesses both feasibility and practical business value in financial risk control tasks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.