Multi-Agent Collaboration for Hallucination Correction in Complaint Management

Large language models (LLMs) can mimic people with different areas of expertise and, in this way, offer a common solution process for problems with agents having various expert perspectives. This study presents an approach that creates automatic responses to customer complaints using multi-agent collaboration in LLMs. In the proposed structure, agents with different expertise were allowed to develop solutions to the problem through brainstorming. An end-to-end complaint management process was imitated. The results produced by Claude-3.5-Haiku, Gemini-1.5-Flash, and GPT-4o were evaluated using a self-evaluation technique according to ISO 9001 standards. Claude showed higher scores in the results than the other two language models. The developed hybrid method scored higher by $\sim 2 \%$ on average.