CurbAI: Making Policy-Safe LLM Agents Work in Real-World Curbside Delivery

Curbside delivery is one of the most operationally constrained and policy-sensitive layers in last-mile logistics. Although recent large language model (LLM) agents have demonstrated promise in decision-making tasks, their unconstrained use in deployment-oriented delivery operations raises concerns about safety, compliance, and reliability. This study presents CurbAI, a policy-safe LLM-orchestrated control plane that integrates probabilistic reasoning with deterministic policy validation and fallback mechanisms for curbside delivery decisions. Unlike end-to-end LLM planners, CurbAI constrains LLM reasoning to bounded candidate sets and enforces municipal, operational, and safety constraints through explicit policy gates. We evaluated CurbAI using a simulation-based study across 30 delivery routes spanning three regions (the US, EU, and JP), with heterogeneous stop densities, curb conditions, and operational profiles. The results demonstrate consistent route completion, controlled fallback behavior under uncertainty, and stable dwell-time distributions across regions, validating the robustness of policy-constrained LLM orchestration for real-world curbside delivery. These findings suggest that policy-constrained LLM control planes can augment last-mile logistics systems while preserving deterministic safety, compliance, and operational guarantees.