AI-Driven Orchestration of AutoML-Based Demand Forecasting, Fleet Allocation, and Route Optimization in Last-Mile Delivery: A Single-Instance Case Comparison of Metaheuristic and Agentic AI Routing

Background: Last-mile delivery couples demand forecasting, heterogeneous fleet allocation, and routing under time windows, yet these decisions are usually studied in isolation. The object of this study is the forecast-to-route decision pipeline of an urban parcel operator; we examine the hypothesis that no single routing paradigm dominates once solvers are evaluated under one shared feasibility-aware objective. Methods: We implement an orchestration architecture in which an AutoML supervisor selects transfer-learned recurrent backbones per zone, forecasts become Vehicle Routing Problem with Time Windows instances over the Casablanca road network, and five solvers—three metaheuristics (GA, PSO, ACO) and two multi-agent reinforcement learning methods (MAPPO, MADQN)—minimize one common fitness function. The demand corpus is synthetic, generated from the LaDe Hangzhou partition to preserve its spatial and temporal structure within the Casablanca boundary. Results: Across five backbones, accuracy differences were small (all R2≥0.845; best bidirectional LSTM, R2=0.865, wMAPE =29.4%). On one 50-customer instance solved once per solver, GA returned the best feasible solution, whereas MAPPO reached the shortest distance only by violating time windows. Conclusions: Backbone choice is not the decisive forecasting decision, and constraint handling rather than paradigm separates the routing solvers. Multi-instance, budget-matched evaluation remains necessary and is specified.