Industries / Technology / Software and SaaS / ERP and CRM software
Supply chain
Deciding what to buy and make, and when: forecasting demand, planning materials and factory capacity, placing purchase orders with suppliers and keeping the stock records the plan depends on.
6
Milestones achieved, from 1960s to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1960s
Parts needs calculated, not guessed
Joseph Orlicky, at the tractor maker J.I. Case and at IBM, and George Plossl and Oliver Wight, at The Stanley Works, each wrote programs for material requirements planning (MRP). MRP works back from the production plan through each product's bill of materials, the list of its parts, to how many of each part are needed and when. It replaced reorder rules that assume steady demand; demand for parts comes in uneven batches as products are built, but it can be calculated. The three met at a conference of the American Production and Inventory Control Society (APICS) in 1966, ran an "MRP Crusade" through the society in the early 1970s, and Orlicky published his book on the method in 1975.
Credit
Joseph Orlicky, George Plossl and Oliver Wight; APICS
1981
Factory capacity enters the plan
Oliver Wight's 1981 book set out manufacturing resource planning (MRP II). It added checks of factory capacity to MRP, extended the plan to people and budgets, and tied it to control of work on the shop floor. In the early 1990s these systems grew into enterprise resource planning (ERP), which also covers accounting, finance, sales and human resources.
Credit
Oliver Wight
1996
Retailer and supplier plan together
In a five-month pilot from April 1996, Wal-Mart and Warner-Lambert planned stock of Listerine mouthwash jointly, in what was named collaborative planning, forecasting and replenishment (CPFR). Warner-Lambert reported that Wal-Mart's in-stock rate on the line rose from 87% to 98%, while its inventory of the 12 Listerine products fell by about 20%. The pilot was run on paper, not over the internet as first planned. In the late 1990s a retail standards committee, VICS, turned the process into a nine-step model.
Lessonin 2007 the consultant Rich Sherman wrote that more than 300 companies had implemented some form of CPFR but that it had been "relegated to the strategic few", the largest customer–supplier pairs; the head of VICS disputed it.
Credit
Wal-Mart; Warner-Lambert; VICS
1997
Orders are not demand
Hau Lee, V. Padmanabhan and Seungjin Whang described the "bullwhip effect". Procter & Gamble had found that orders for Pampers diapers swung more at each step up the chain, from stores to distributors to P&G's own orders from suppliers such as 3M, though babies used diapers at a steady rate; Hewlett-Packard saw the same with a printer. The authors named four causes: forecasts updated from the next company's orders, ordering in batches, price changes such as promotions, and rationing in shortages. One remedy they described is passing stores' sales data back up the chain to suppliers.
Lessona plan built on customers' orders inherits their swings, so planning needs the end sales figures.
Credit
Hau L. Lee, V. Padmanabhan and Seungjin Whang, in Sloan Management Review
2001
A planning system hits earnings
On 26 February 2001 Nike cut its forecast for the quarter to 34 to 38 cents a share, from 50 to 55 cents. Its chairman, Philip Knight, blamed weak US footwear sales and "complications arising from the impact of implementing our new demand and supply planning systems and processes which resulted in product shortages and excesses as well as late deliveries." The software's maker, i2 Technologies, then reported shareholder lawsuits alleging it had hidden "software implementation difficulties" with Nike.
Lessona new planning system can cause shortages and surpluses at the same time, within one quarter.
Credit
Nike; i2 Technologies
2026
Agents tested on factory work
ERP-Bench set AI agents 300 long procurement and manufacturing tasks in Odoo 19, an open-source ERP, each starting from a fresh database seeded with stock, customers, suppliers, bills of materials and work centers, and each with a best answer certified by a solver. Across agents that work by writing code, by driving a browser and by using a computer screen, frontier models followed every rule and instruction in 25.9% of trials and reached the best decision in 17.4%.
Credit
Abhijay Rana (University of California, Berkeley) and Maksim Ivanov (Agentic Labs)
Watch for
A simulated year, no stockouts
Researchers at KTH Royal Institute of Technology in Stockholm built Agentic ERP, a set of AI agents with separate business roles that work through Odoo 19's sales, purchasing, inventory and accounting modules, with a person signing off risky steps. They report that it ran a simulated trading company for 365 days with zero stockouts. A rule-based automation baseline built up hundreds of stockouts under the same demand. The scenario and data are the authors' own.
Achieved whenanother group reruns the test on released data, or a company runs the system on its own records, and reports stockouts and costs.
Claimed by
Zhihao Liu, Tianyu Wang, Xi Vincent Wang and Lihui Wang, KTH, July 2026
Supply problems in one click
SAP says its supply chain assistants let planners "turn multi-hour disruptions into one-click resolutions across planning, manufacturing and logistics". The chief financial officer of Morgan Foods, a customer, said agents would connect demand signals, inventory levels and supplier status "faster than ever before". SAP gave no measurements.
Achieved whena named customer publishes the time from a supply problem to an agreed new plan, before and after, over at least a quarter, measured by someone other than SAP.
Claimed by
SAP, October 2026
Open
01
Supplier emails into orders
A public test set of supplier emails about purchase orders, such as confirmations, new delivery dates, short shipments, price changes, cancellations, split deliveries and replies that promise an answer later, each labeled with the order, line and field it changes. It scores how often AI agents sort each email correctly and propose the right change to the order.
The briefCredit · unclaimed02
A year in the factory
A public, year-long test in which AI agents plan a made-up factory in an open-source ERP, with bills of materials several levels deep, limited machine capacity and a stream of setbacks such as late suppliers, broken machines and rush orders. The agents are compared with the ERP's own MRP run and with a human planner on the same stream.
The briefCredit · unclaimed
Open 01 · the brief
Supplier emails into orders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of supplier emails about purchase orders, such as confirmations, new delivery dates, short shipments, price changes, cancellations, split deliveries and replies that promise an answer later, each labeled with the order, line and field it changes. It scores how often AI agents sort each email correctly and propose the right change to the order.
Why it is still open
Microsoft's Procurement Agent for Dynamics 365 reads supplier emails, decides whether each is a confirmation or a change request and finds the purchase orders it refers to. Microsoft says only that it "underwent substantial testing with demo data examples", and that generated content "should never be used without manual review or supervision". Published extraction work covers purchase-order documents, not the emails about them; one 2023 study reported results on a private set of invoices and purchase orders. A search found no public labeled set of such emails.
Built with what exists
made-up purchase orders in ERPNext or Odoo Community; emails written from templates and by AI models, then checked by someone who has worked in purchasing; a model that returns probabilities for fixed-answer questions, such as Gemini Decision 1, to sort the emails by type (see the Note "A thousand questions, one pass"); agents built on models such as Claude or GPT to propose the field changes.
First step
Write 100 emails of seven kinds against 30 ERPNext purchase orders, label them, and run one model on them.
Done looks like
2,000 emails in at least five languages against 500 open purchase orders, with each model's share of emails sorted correctly, share of changes applied to the right line and field, and count of changes it made that the email did not ask for.
Open 02 · the brief
A year in the factory
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, year-long test in which AI agents plan a made-up factory in an open-source ERP, with bills of materials several levels deep, limited machine capacity and a stream of setbacks such as late suppliers, broken machines and rush orders. The agents are compared with the ERP's own MRP run and with a human planner on the same stream.
Why it is still open
ERP-Bench scores single tasks by their end state. KTH's year-long simulation ran a trading company with only sales, purchasing, inventory and accounting installed, so no parts were made from other parts. No test was found that runs agents over time in a factory.
Built with what exists
Odoo Community's manufacturing and MRP modules, or ERPNext's production planning; Python to generate the factory, its demand and its setbacks; the ERP-Bench authors' task setup as a model for seeding and scoring; agents built on current models, each run several times (see the Note "All twenty times").
First step
Build the factory's bills of materials and routings in Odoo, run its MRP for four simulated weeks with one supplier delay, and record which orders come in late.
Done looks like
One made-up factory with at least 200 parts in three or more levels, run for 52 simulated weeks under 50 different streams of setbacks; for each planner, the share of customer orders delivered late, the number of stockouts and the average inventory value, with code and data under an open license.