InventOR: A Human-Reviewed, Prompt-Configured Large Language Model Decision Support System for Material-Level Inventory Control

Material planners must set a reorder point, a safety stock and an order quantity for thousands of plant-material pairs, usually under ERP parameters that were last revised long ago. We define and evaluate InventOR, a decision support system (DSS) for this recurring decision. For each material, a large language model (LLM) configured only through a prompt, without training or fine-tuning, proposes a reorder-point/order-quantity (r,Q) policy with a written justification and an escalation flag. Deterministic software gates the proposal against minimum order quantities and storage limits and replays it on held-out demand before the planner, who remains the decision-maker, reviews it. We derive six design requirements and evaluate the artifact on 365 plant-material pairs from three plants of an industrial manufacturer, using three independent LLM runs, seven reference arms and an ablation that withholds the ERP planning parameters. In a hard-limit storage replay, the LLM policies cost about 11% less than ERP-derived parameters at a mean fill rate 0.1-0.4 points lower, but 6%, 8% and 18% more than history-only, Croston/SBA and empirical-quantile (r,Q) baselines. The saving comes from carrying 13-17% less inventory with 11-20% more stockout days; when shortages are priced at one unit cost, ERP-derived parameters are cheaper and a simulation-tuned policy is cheapest. Static storage checks are insufficient: simulated trajectories exceed storage limits for 53-66% of pairs. InventOR is a checkable, human-reviewed candidate generator, not an optimizer that beats data-driven baselines; benefits to planners are untested. We contribute design principles, review rules and a research agenda.