Many studies have shown the potential of machine learning for vehicle routing problems (VRPs), yet most focus on isolated cases such as the Traveling Salesman Problem or Capacitated VRP. We extend this line of research to last-mile delivery by integrating demand management, through pricing and delivery time-slot decisions, with routing optimization. This creates a two-stage stochastic problem where customer demand depends on operational choices, and routing adapts to the resulting orders. We develop machine learning–enhanced heuristics to efficiently approximate this two-level decision process, enabling scalable solutions under real-world uncertainty.