A mid-sized manufacturer that builds finished goods to stock should aim its first AI project at the sales forecast those stock levels are planned from, and judge the project by days of inventory on the product lines that forecast plans.
The Census Bureau's Quarterly Financial Report, released September 8, 2026 and estimated from a sample of US manufacturing corporations with total assets of 5 million dollars and over, shows inventories above total receivables at the end of the second quarter of 2026 in every all-manufacturing asset band from 50 million dollars up. The report sizes corporations by total assets rather than shipments.
A make-to-stock plant, in the words of an April 2025 Census Bureau working paper, relies heavily on forecasting and inventory buffers to address uncertainty. A researcher at MIT's Global SCALE Network institute in Malaysia tested two machine learning methods at one multinational steel manufacturer in a journal paper published online August 4, 2020. Forecasting a year of monthly demand in its retail segment, they gave on average a 5 percent improvement in forecast accuracy over the company's methods, and a regression estimated the cash conversion cycle 21.6 days shorter using inventory the paper generated, not counted, from those forecasts and the parent's published accounts. The author calls the single dataset a major limitation.
A peer-reviewed paper by researchers at Lancaster, Castilla-La Mancha and Birmingham, partly funded by the European Regional Development Fund and the Spanish government, which the University of Birmingham repository dates 2019, used 229 items from a UK maker of household cleaning and personal hygiene products. Forecasts tuned in simulation to inventory results at the company's service targets lost as much as 9 percent of accuracy, added limited stock and cut stock-outs, a trade-off dominating the same method tuned for accuracy. Those forecasts used exponential smoothing, a statistical method.
The working paper's authors, at Toronto, Oklahoma, Stanford and Analysis Group, with partial funding from the Stanford Digital Economy Lab, drew on a 2021 Census survey of roughly 28,500 manufacturing plants that asked about AI in six business functions, sales forecasting among them, and in specific technologies, then combined the answers into one index. They found that industrial AI use increases work-in-progress inventory, investment in industrial robots, and labor shedding, while harming productivity and profitability in the short run. Census says its working papers have not had the review its own publications get.
In month one the demand planning manager lists the lines built to stock, and the plant controller records each line's month-end days of inventory, meaning inventory divided by a day's cost of goods sold, with work-in-progress beside it. During month two the planner runs a candidate model over the last 52 weeks through the plant's current reorder rule, comparing the stock on hand and stock-outs it would have produced with the current forecast's. In month three the COO lets the model's forecast set production where it held less stock at equal or fewer stock-outs, and the controller reports days of inventory, work-in-progress and stock-outs monthly. None of these sources prices the work. An operating partner can ask each make-to-stock portfolio company for those three numbers.
If the back-test cuts stock only by adding stock-outs, the model stays out of production planning. A line whose work-in-progress climbs after go-live returns to the old forecast. A plant that builds to order sits outside this rule: the working paper describes Lean make-to-order production as prioritizing responsiveness to realized demand over prediction.
In that survey, approximately 43 percent of plant managers considered cost prohibitive to AI use, and 28 percent named difficulty identifying business use cases.
