AI without reach: The last mile problem in preventive public health

Objectives Examine the "last mile problem" in artificial intelligence (AI) applications for preventive public health. Study design Conceptual short communication introducing a five-stage pathway to examine where AI-enabled prevention breaks down. Methods This draws on emerging literature in AI, health services research, and implementation science to distinguish between predictive capacity and real-world impact. A five-stage pathway model of prediction, translation, delivery, uptake, and sustainment is introduced to characterize where breakdowns occur between identifying risk and changing outcomes. Results Current AI applications in public health are heavily concentrated in prediction, with limited attention to delivery, engagement, and sustained behavior change. These gaps are particularly consequential in prevention, where action must occur before consequences are visible, and in contexts such as family caregiving, where individuals operate as de facto delivery systems without adequate support. Traditional measures of healthcare access further obscure the problem by suggesting that services are available when real-world barriers prevent their use. The last mile problem is not evenly distributed; it disproportionately affects populations facing structural disadvantage, meaning that AI systems that fail to account for delivery and context risk widening existing inequities. Conclusions Addressing the last mile requires reorienting AI from a primarily predictive tool to one embedded in systems of care and daily life. This includes integrating real-time data, designing for context, and evaluating AI on whether it enables timely, sustained action, not on predictive accuracy alone. Until AI reaches the last mile, its contribution to prevention will remain incomplete, accurate in prediction but limited in impact.