Abstract for: The Role of Generative Artificial Intelligence in Adaptive Decision Making under Uncertainty: Evidence from Last-Mile Delivery

Last-mile delivery operations frequently experience disruptions such as extreme weather, which create substantial uncertainty for frontline couriers. In these environments, operators rarely solve formal optimization problems but instead rely on heuristics to generate feasible actions. This study examines whether generative AI (GenAI) can improve operational adaptation by expanding the set of routing and scheduling alternatives considered by frontline decision makers. To answer the research question, we adopt an empirically grounding analytics (EGA) approach by combing causal empirical analysis with an agent-based model of courier decisions. Using delivery data, a causal inference design estimates how extreme weather disruptions affect courier performance. The empirical findings calibrate the agent-based simulation of routing decisions. GenAI is then modelled as a decision-support tool that expands the set of candidate actions available to couriers under disruption. Preliminary empirical analysis shows that extreme weather significantly increases delivery delays and reduces task completion rates. Simulation experiments suggest that expanding the set of feasible routing and scheduling actions improves operational adaptation. The benefits of AI-supported decision generation appear strongest under severe disruptions and in dense urban environments where the number of feasible operational responses is larger. Our study introduces a decision-generation perspective on technology adaptation in the operations context. We argue that GenAI improves performance not primarily through optimization but by expanding the set of actions available to frontline operators. This expansion creates decision flexibility, a cognitive form of operational flexibility that enables more effective responses to disruptions. The framework illustrates how EGA can evaluate emerging technologies when large-scale field adoption data remain limited.