Designing a Generative AI Co-Pilot for Specific Health Guidance: A Viewpoint on Task Allocation and Formative Evaluation (Preprint)

UNSTRUCTURED Generative artificial intelligence (AI) could assist preventive counseling by drafting records, organizing information, and supporting communication between scheduled encounters. Its capacity to produce plausible text, however, does not establish that a service will be useful, manageable, or safe. This Viewpoint uses Japan’s Specific Health Guidance (SHG) program, with a focus on workplace delivery, to examine how tasks should be allocated before introducing a generative AI co-pilot. We advance three propositions. First, task allocation should depend on whether outputs can be verified in practice and whether errors can be contained before consequential use. Second, professional oversight should be defined at the point when an output is released, distinguishing approval of a support pathway from review of an individual generated response. Third, generative functions should demonstrate added value over conventional digital support when the complete workflow is considered. Drawing on sociotechnical work-system design and research on human interaction with automation, we apply these propositions to documentation, between-session support, and follow-up. We distinguish these functions from predictive prioritization and unrestricted autonomous counseling, which require separate justification and evaluation. We then outline a formative evaluation agenda addressing local needs, verification and correction, participant understanding, total workload, equitable access, and organizational capacity. Hypothetical SHG examples illustrate how evaluation could challenge, rather than merely confirm, the proposed design. The model has not been implemented or validated and does not establish the feasibility or effectiveness of AI-supported SHG. Its contribution is a task-level argument about the evidence and responsibilities needed before claims of scalability are made. Development should remain open to a narrower AI function, a conventional digital alternative, or a decision not to introduce AI.