AI decision-support systems often focus on predicting the “right” answer, yet this approach can conflict with how experts actually make complex, high-stakes decisions. Through a case study of sepsis diagnosis, we investigated how clinicians interact with existing AI support and found that final-stage predictions were often perceived as late, opaque, and difficult to act on. We designed SepsisLab, a human-centered AI system that instead supports intermediate stages of clinical reasoning through future risk prediction, uncertainty visualization, actionable lab-test recommendations, and counterfactual exploration. The project proposes a shift from AI as a competing decision-maker toward AI as a collaborator that helps experts explore hypotheses, gather evidence, and reduce uncertainty while preserving human agency over final decisions.