AI in Supply Chain Optimization

Abstract

Global supply chains operate under compounding uncertainty -- volatile demand, supplier disruptions, transportation delays, geopolitical risk, and climate-related shocks -- that traditional operations research methods address through deterministic optimisation and scenario planning but cannot dynamically adapt to in real time. AI-powered supply chain optimisation offers three capabilities traditional methods lack: demand forecasting that captures non-linear and causal relationships in high-dimensional signals, real-time replanning under disruption, and autonomous decision support for multi-echelon inventory and routing optimisation. This study evaluates eight AI approaches for supply chain optimisation across five problem domains: demand forecasting, inventory optimisation, supplier risk assessment, vehicle routing, and disruption response. Methods include gradient-boosted trees (XGBoost, LightGBM), Temporal Fusion Transformer (TFT), deep reinforcement learning for inventory control (DRL-Inv), Graph Neural Networks for supplier risk (GNN-Risk), large language model-powered planning agents (LLM-SCM), combinatorial optimisation with neural heuristics (OR-Neural), and a hybrid ensemble. Data from 12 European manufacturing companies across six sectors (automotive, pharmaceutical, electronics, food, energy, and logistics) over 2019-2023 -- including the COVID-19 disruption period - was used for evaluation. TFT achieves the lowest demand forecasting MAPE (8.4%) across product categories. DRL-Inv reduces inventory holding costs by 18.6% vs. EOQ baseline while maintaining service levels. GNN-Risk predicts supplier disruptions 14.2 days earlier than human supply chain managers. LLM-SCM reduces disruption response time by 42.4% in simulation but introduces hallucination risks requiring human validation. A sector-specific AI adoption roadmap is proposed