A wholesale distributor should let buyers override its inventory AI only on new items and put what they know about every other item into the forecast the AI reads.
The first half of the rule rests on a randomized field experiment that an automobile replacement parts retailer conducted with the corporate buyers who stock its affiliated stores across the United States, published in the November 2020 issue of Management Science by professors at the University of North Carolina and George Mason University. The retailer's tool flagged underperforming products for removal, and its buyers rejected those flags more than half the time, George Mason University reported in November 2021. The trial ran for 12 months on more than 30,000 products, with the buyers' overrides followed in some stores and ignored in a random selection of others; neither the abstract nor George Mason's account gives either group's store count. The abstract reports that the buyers' overrides reduced profitability by 5.77%, and that they increased profitability on growth-stage products while decreasing it on mature and decline-stage products. George Mason's account describes growth-stage products as new to market and gives the professors' theory that new products have no past performance record for the tool to base decisions on.
The second half rests on a large-scale field experiment at an automotive spare parts retail chain, published in the January 2026 issue of Management Science by two of the 2020 authors and a third at the University of Minnesota. There the buyers' adjustments to an inventory algorithm's forecast inputs increased profitability on average compared with relying on automation without human intervention. The abstract adds that the paper examined when adjustments improve profits by stock-keeping unit margin, life cycle and supplier size, and gives no sample size, period or result by stage.
The purchasing manager should have the ERP tag each item new when added and, on the first business day of each month, retag it established once 12 months have passed since its first sale at any branch, or since added if no branch has sold it. A replacement part number takes the old number's tag and dates. The tag measures new to the distributor, a stand-in for the 2020 study's growth stage. A buyer sets a new item's opening forecast and may keep stocking it after the AI flags it for removal. On an established item, a buyer enters what the sales history does not show, such as a contractor's job won but not yet ordered, as a forecast adjustment with the reason before the AI's next order run. The purchasing manager keeps each supplier's order and freight minimums in the AI's settings and approves each buy to reach a rebate tier or ahead of a supplier's announced price increase or discontinuation, and any stock a customer contract requires. In the first week of each month the purchasing manager reviews the previous month's 20 largest overrides and adjustments by dollar value. Each quarter the controller reports gross profit and average inventory on new and established items against the same quarter of the year before the rule began, and shows how much of the change on established items came from changes in purchase costs and selling prices.
The distributor pays for the AI and carries the inventory its orders create, so the report should come from the controller, outside the purchasing team whose calls it measures. George Mason's account cautions that what works for the auto parts industry in the United States today may not for another time, place or sector, so each distributor should read that report as its own check on the rule. An operating partner can ask each distributor in a portfolio that runs an inventory AI for the same report. George Mason's account gives the reason the tool came out ahead overall: at the retailer in the 2020 experiment, the vast majority of parts were mature or decline-stage.
