Policy-Informed AI Assistant for Decision Support in Public Health System: Tool Development and UX Study With Health Service Managers
Abstract
Artificial intelligence in healthcare has concentrated on clinical applications, while support for the managerial work that sustains health services is yet underexplored. Health system managers decide against large, dispersed, and frequently revised bodies of official guidance, a task that differs from clinical question answering in the nature of its corpus. In this context, we developed and evaluated a policy-informed AI assistant for management decision support, using Brazil’s Unified Health System as a demanding instance of a decentralized system with a large policy corpus. Following a mixed-methods design, we (i) mapped and indexed 106 official Ministry of Health documents into a curated knowledge base, spanning primary care operational documentation, management portals, and national health planning instruments; (ii) developed a LLM-based assistant that grounds its answers in that base through retrieval, displays the consulted sources, constrains its scope regarding health management, and offers suggested questions; and (iii) evaluated exploratory user experience with 12 municipal health managers, who completed standardized use-case scenarios and then answered the User Experience Questionnaire (UEQ+) and open-ended questions. All seven UEQ+ scales were rated positively (means 1.08 to 1.69), with Efficiency, Dependability, and Perspicuity highest and Novelty lowest; every scale was above the neutral midpoint. Content analysis produced three categories: positive aspects (speed, ease of use, clarity, and practical usefulness), obstacles and constraints (conditional trust and a need to verify content), and security and transparency (data protection and source transparency). The quantitative and qualitative evidence converged, indicating adequate usability alongside a conditional, verification-oriented form of trust. A policy-informed AI assistant that cites its official sources proved feasible and was well-received as a decision-support tool for SUS management. Trust was conditional, and responsible adoption depends on source transparency, human verification, and constant refinement. The findings are exploratory and motivate larger, longitudinal, and comparative studies.