Addressing trust requirements in the design of an open-source multi-agent LLM-based domain-specific chatbot

What the paper says

Large Language Models (LLMs) have the potential to automate knowledge-intensive interactions in enterprise systems, yet their adoption is often limited. One reason is a lack of user trust. This study examines how trust can be systematically engineered into an LLM-driven, multi-agent chatbot that handles routine human-resources (HR) queries. We follow a two-cycle Design Science Research methodology. Cycle 1 triangulated a systematic literature review with a thematic analysis over semi-structured interviews of six employees at a global firm and a confirmatory workshop with five AI experts to elicit and validate trust requirements . Cycle II instantiated these requirements in a multi-agent LLM chatbot prototype artifact and evaluated whether the artifact satisfies them through controlled user sessions and expert walkthroughs, emphasizing perceived usefulness and trust captured in post-task interviews ( $$n = 11$$ ) and operationalizing trust via alignment-oriented measures (faithfulness, answer relevancy, and adversarial robustness). The study yields a refined taxonomy of external (transparency, organizational safeguards, third-party security) and internal (model provenance, bias risk, reliability) trust factors, identifying reliability as the primary determinant of adoption. The implemented design achieved $$\ge 0.86$$ on trust-aligned metrics and was endorsed by 9/11 participants as ready for field deployment. These findings demonstrate that trust can be proactively addressed through design and offer prescriptive guidelines for software engineers seeking to embed LLMs safely and responsibly in socio-technical contexts.