Artificial intelligence decision support system for small business supply chain resilience under uncertainty

Small-business supply-chain studies often report forecast improvements without showing how calibrated demand and logistics uncertainty should change feasible replenishment decisions. This study evaluates an uncertainty-aware artificial intelligence decision support system implemented as an auditable benchmark architecture that couples leakage-safe forecasting, conformal demand intervals, calibrated ex-ante logistics-delay probabilities, a conditional-value-at-risk stochastic linear program, and calibration-frozen ABC-XYZ policy assignment. The constituent methods are established; the contribution is their explicit and separately testable coupling from prediction to constrained prescription. The analysis used 350,000 FreshRetailNet-50 K records and 120 store-product series, DataCo for logistics-risk estimation, and UCI Online Retail for external robustness. XGBoost improved mean absolute error by 2.74% over a seven-day moving average. Under the baseline scenario weights, the risk-adjusted program achieved 3.93% stockout and 95.53% aggregate service and significantly outperformed fixed reorder, EOQ, forecast-only, and risk-neutral optimization after Holm correction, but it did not differ significantly from the adaptive policy. Reoptimization under equal, routine-heavy, and disruption-heavy probability profiles preserved policy non-dominance but changed the LP-adaptive relationship, demonstrating sensitivity to probability beliefs. AI-only retained the highest service and EOQ the lowest cost across all profiles. The results establish a context-dependent cost-resilience frontier rather than universal AI superiority.