This research investigates how predictive AI affects the learning and performance of medical novices during lung cancer diagnosis. In two preregistered field experiments, we find that providing AI input during training or practice significantly boosts individual accuracy and helps develop long-term expertise. Moreover, integrating AI input into both phases yields the best results, allowing novices to reach performance levels near those of medical experts. Our study also highlights that, depending on the deployment strategy, AI input can increase or reduce human error diversity, which is vital for the effectiveness of group decision-making. Our findings suggest that, with the right deployment strategy, predictive AI can be a powerful coach that strengthens human capability (individually and collectively), rather than being merely a temporary crutch for task completion.