Industries / Industrial equipment and services / Industrial distribution
Inventory
Deciding which products each branch and distribution center keeps on the shelf and how many, which to order only when a customer asks, and what to do with stock that stops selling.
4
Milestones achieved, from 1962 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1962
Order points by formula
IBM's 1962 manual for IMPACT (Inventory Management Program and Control Techniques), a library of computer programs for "distribution industries", noted that a wholesaler may stock from 10,000 to more than 100,000 items. It replaced stock clerks' rules such as "When available stock gets down to three weeks' average usage, order enough to bring it up to six weeks" with reorder points set from a demand forecast, a measure of the forecast's error and a chosen service level, the share of demand to be met from stock (95% in its worked example). Its forecasts used exponential smoothing, a running average that weights recent sales most, which Robert G. Brown developed for US Navy inventory systems in the 1950s and published in a 1959 book.
Credit
IBM; Robert G. Brown
1984
Zone warehouses feed the branches
In 1984 Graybar adopted a zone warehouse plan and opened its first regional zone warehouse, in Bethlehem, Pennsylvania. By 2025 it had 22 regional distribution centers that hold both standard and specialized products, refill its local branches and ship to customers directly, while each branch keeps stock for its own territory. Of the roughly two million products it sells, more than 100,000 are stocked; for specialized or very large orders the maker ships straight to the customer.
Credit
Graybar
1999
The system counted phantom stock
W. W. Grainger moved its branches and zone distribution centers from older systems to a new enterprise resource planning (ERP) system, SAP's R/3, in 1998 and 1999. Grainger's annual report estimated that the resulting service problems lost about $20 million of sales in 1999. Computerworld reported $19 million of lost sales and $23 million of lower earnings in the second and third quarters, and that a physical count at the end of the year found the software was recording more products than were in the warehouses, so the inventory figures had to be written down.
LessonAfter a system change, count the shelves before trusting the books.
Credit
W. W. Grainger; SAP
2023
Shortage stock could not return
During the chip shortage of 2021 and 2022 some suppliers increased the share of orders that could not be cancelled or returned, which Avnet, a large electronic parts distributor, said limited its ability to bring its inventory down. From the second half of 2023 sales fell, as customers held high stocks and underlying demand weakened. Avnet's inventory rose from $2.73 billion in June 2020 to $5.47 billion in July 2023, and was still $5.47 billion a year later.
LessonReturn rights let a distributor undo a stocking mistake, and makers cut them back when supply is short.
Credit
Avnet, in its filings with the US Securities and Exchange Commission
Watch for
AI forecasts across a distributor
Sonepar, a Paris-based electrical distributor, says predictive models on a global data and AI platform built with the French consulting firm Onepoint now support businesses that make 95% of its revenue, for demand forecasting, inventory levels and product availability. Sonepar reported €33.6 billion of sales in 2025 and runs 180 distribution centers in 40 countries; it has published no before-and-after figures for stock or availability.
Achieved whenSonepar or an outside party publishes measured changes in inventory, fill rate or stockouts for the businesses using the models.
Claimed by
Sonepar, September 2026
Open
01
Open sales data for branches
An anonymized record of two years of daily sales by item and branch from a real industrial or electrical distributor, about 20 branches and one distribution center, with lead times, stock on hand and which lines were special orders, plus a benchmark task: choose what each branch stocks and how many, scored in simulation by fill rate (the share of order lines filled from the shelf) against the money tied up in stock.
The briefCredit · unclaimed02
Find stock that can return
A public table of the return and stock rotation terms that makers publish for their distributors (stock rotation means sending unsold stock back to the maker for credit): how many rotation returns a year, the cap as a share of last year's purchases, the handling fee, the age limit on returned goods and the kind of credit, for 100 makers of electrical and industrial products. With it, a checker that reads a branch's list of slow-moving stock and marks which items can still go back, to which maker, and by when.
The briefCredit · unclaimed
Open 01 · the brief
Open sales data for branches
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An anonymized record of two years of daily sales by item and branch from a real industrial or electrical distributor, about 20 branches and one distribution center, with lead times, stock on hand and which lines were special orders, plus a benchmark task: choose what each branch stocks and how many, scored in simulation by fill rate (the share of order lines filled from the shelf) against the money tied up in stock.
Why it is still open
Public demand benchmarks come from retail; FreshRetailNet-50K, for example, holds hourly sales of 863 fresh-food items from 898 stores. AI agents for inventory, such as InvAgent, are tested in simulated supply chains whose customer demand comes from a set formula. The nearest industrial set, a synthetic spare-parts demand benchmark Joo Ern Chin posted in June 2026, is a masked stand-in for one confidential company's records, because, its description says, the industrial records where demand is patchy are confidential; it has no branches.
Built with what exists
A distributor's ERP export; Python with pandas; Nixtla's statsforecast library for baseline forecasts (exponential smoothing and Croston's method); a simple inventory simulator; an AI agent baseline adapted from InvAgent's open code on GitHub.
First step
Ask two local distributors or a buying group whether they would release two years of branch sales with customer names removed and item numbers replaced by product categories and codes.
Done looks like
The data released under an open licence with a data card that says what was removed, and three stocking policies (an IMPACT-style reorder point, a statsforecast policy and an AI agent) scored on the same 52-week holdout for fill rate and average inventory value.
Open 02 · the brief
Find stock that can return
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the return and stock rotation terms that makers publish for their distributors (stock rotation means sending unsold stock back to the maker for credit): how many rotation returns a year, the cap as a share of last year's purchases, the handling fee, the age limit on returned goods and the kind of credit, for 100 makers of electrical and industrial products. With it, a checker that reads a branch's list of slow-moving stock and marks which items can still go back, to which maker, and by when.
Why it is still open
The terms sit in each maker's own terms of sale. Hubbell's 2016 distributor terms, for example, allow two stock rotation returns a year, up to 5% of the previous year's purchases by brand, with a 25% handling fee, only for goods shipped to the distributor within the previous 18 months, and only for merchandise credit. STEMCO, a maker of truck and trailer parts, allows one return a year of up to 2% of the previous year's net purchases, with a 10% restocking fee, and refuses goods with a date code more than three years old. I found no public compilation of such terms across makers (searches in November 2026).
default.assets-stateelectric-prod.roccommercecloud.comstemco.com
Built with what exists
Makers' terms-of-sale and returned-goods PDFs; pdftotext; an AI model to pull the terms into fixed fields, with every value checked by hand; a short Python script for the checker that reads an inventory export (item, maker, date received, value).
First step
Collect the published terms of sale of 30 makers that sell through electrical distributors and code ten of them by hand.
Done looks like
Terms for 100 makers in the table, each field linked to its dated PDF and checked by a second person; on a sample stock list of 1,000 lines, the checker's flags confirmed by hand with at least 95% agreement.