Multi-Agent LLM Approach for Moderating E-Commerce Customer Service Responses

Language model (LLM)-based solutions have been widely adopted in automated customer service systems, particularly on e-commerce platforms. However, such solutions still face challenges related to the accuracy, contextualization, and reliability of the generated responses. This study proposes an LLM-based multi-agent architecture for the automatic moderation of textual responses. The architecture is composed of specialized agents operating in an iterative review workflow, which includes semantic and contextual evaluation, improvement recommendations, textual rewriting, and final decision-making. The agents share a common context and operate in a coordinated manner to identify deficiencies, propose corrections, and validate the quality of the responses. The proposed approach was evaluated using real-world data from a multilingual e-commerce platform, leveraging two models from the Qwen3 family (32B and 30B-A3B). The results indicate that the approach is effective in enhancing response quality, achieving average gains of more than two points on an evaluation scale and enabling the correction of over 60% of initially inadequate responses. Additionally, the solution offers advantages in terms of auditability, modularity, and potential adaptability to different domains.

Proceedings of the 31º Brazilian Symposium on Multimedia and Web