Providing Task Execution Capabilities in LLM-Based Conversational Assistants

Context: Conversational assistants (CAs) powered by Large Language Models (LLMs) excel in generating coherent responses, but struggle with task execution requiring long-term memory or temporal awareness. This research addresses such limitations, particularly in healthcare applications where task continuity is crucial. Problem: Existing CAs often fail to meet user expectations in scenarios requiring long-term task management, such as scheduling reminders or monitoring chronic health conditions. This lack of temporal reasoning and memory continuity undermines user trust and engagement. Proposed Solution: We propose a multi-agent system integrating LLMs with structured data processing, that enables CAs to recognize, manage, and execute user requests, such as scheduling reminders or tracking tasks over time. The solution bridges the gap between conversational fluency and actionable outcomes. IT Theory: The study draws on theories of Human-Computer Interaction (HCI) and Temporal Awareness in Information Technology for allowing structured task execution. Method: The research combines proof-of-concept implementation with healthcare case studies. Two experimental phases evaluated the solution: a pilot test with 35 users and an expanded trial with 437 participants focusing on chronic disease management. Summarization of Results: Results demonstrated an improvement in task fulfillment and user engagement. The assistant successfully addressed 82% of task requests in simulations, with reduced user frustration and longer interactions during real-world trials. Contributions and Impact on IT: The study enhances the capabilities of conversational assistants by introducing temporal reasoning and structured task execution, making them more effective in managing real-world tasks. These advancements improve the applicability of CAs, particularly in domains like healthcare, where accurate and timely task management is critical.