An AI-Enabled Predictive Procurement Model for Forecasting and Inventory Optimization

Abstract

This paper presents an AI-enabled predictive procurement model that integrates probabilistic demand forecasting, supplier lead-time prediction, and multi-echelon inventory optimization to secure target service levels while lowering working capital. The framework couples hierarchical time-series forecasting with gradient-boosted trees and lightweight sequence models, capturing seasonality, promotions, and regime shifts. It augments predictions with causal drivers including macroeconomic indicators, input price indices, marketing calendars, and policy shocks, producing calibrated distributions rather than point estimates. Operationally, the system ingests ERP and MRP transactions, purchase orders, invoices, catalog attributes, and supplier performance histories. Data quality pipelines execute de-duplication, unit harmonization, outlier handling, and late-arrival correction, while item-location-supplier embeddings transfer information across sparse series. Lead-time models use gradient boosting with monotonic constraints on distance and order size, and a compact LSTM captures intra-week receipt patterns. Forecast reconciliation ensures coherence from SKU to category to enterprise, and conformal prediction yields reliable uncertainty bounds. Optimization formulates a newsvendor-style objective that balances service penalties, holding, and ordering costs under budget and capacity constraints. A stochastic program solved on a rolling horizon outputs order quantities, reorder points, supplier splits, and dynamic safety-stock buffers aligned to service targets. A policy engine encodes guardrails shelflife limits, ABC/XYZ policies, MOQ/pack constraints, and ESG exclusions and writes approved plans back to ERP via governed APIs and event streams with full traceability. In simulation and a limited field pilot, the approach reduced backorders and emergency expedites, lowered average on-hand inventory, and improved forecast bias and MAPE relative to baseline heuristics, with the largest gains on intermittent-demand SKUs. Ablation studies show probabilistic lead-time modeling and cross-series transfer contribute most to service-level stability. Governance artifacts, change-management practices, and value tracking methods are provided to support scale-out across plants and regions. The contribution is a production-ready, interpretable design that turns uncertainty-aware predictions into auditable, financially grounded procurement decisions embedded in day-to-day workflows. We also outline model risk management, privacy safeguards, and human-in-the-loop review to maintain accountability, fairness, and resilience under data shifts and supplier disruptions globally