Industries / Health / Medical supply distribution
Ordering and buying
Hospitals, clinics and pharmacies choose what to buy and at what price, often through buying groups, and send their orders to wholesalers by phone, terminal or computer link; for controlled drugs, the wholesaler must check each order before it ships.
7
Milestones achieved, from 1910 to 2021.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1910
Hospitals buy as a group
A group purchasing organization (GPO) negotiates contracts with sellers on behalf of many hospitals. By the account of the trade association HSCA, reported by the US Government Accountability Office (GAO), the first was set up in 1910 by the Hospital Bureau of New York; there were more than 100 by the 1980s and over 600 in 2014, and 96% to 98% of US hospitals buy through GPO contracts, typically using two to four groups each.
Credit
the Hospital Bureau of New York
1960s
Hospitals order by wire
American Hospital Supply Corporation built ASAP (Analytical Systems Automated Purchasing), an electronic order link between hospitals and the supplier. N. Venkatraman, writing in 1992, dates its development to the early 1960s and calls it one of the best-known and most often cited "strategic information systems", systems a company uses to gain an edge over its rivals. Its later phase, under Baxter, moved away from one supplier's own order link toward systems that connect a hospital to many suppliers.
Lessona link that tied hospitals to one supplier gave way to links that served many.
Credit
American Hospital Supply Corporation
1970s
Pharmacies order by handset
McKesson's Economost let pharmacies key in orders on handheld pads and send them by telephone instead of writing them out; McKesson dates it to the 1970s and calls it the trade's first electronic order transmission system. Eric Clemons and Michael Row found in 1988 that nearly 100% of McKesson's orders were entered electronically by customers, that rivals offered similar systems, and that such systems had a significant effect on the consolidation of the trade.
Lessonrivals copied it, and the two authors found the lasting advantage to McKesson itself hard to establish.
Credit
McKesson
1986
Sellers may pay the buyer
US law bars payments that induce purchases paid for by Medicare. In 1986 Congress made an exception that lets sellers pay GPOs a fee, and a 1991 rule requires each GPO to cap the fee at 3% of the price or state the amount in writing, and to disclose the fees it gets to each customer at least once a year. GAO found the five largest GPOs collected about $2.3 billion in such fees in 2012 and passed nearly 70% to their customers or owners.
Lessona buying agent paid by the seller may gain from higher prices; GAO found little evidence either for or against that concern 28 years on.
Credit
the US Congress (Omnibus Budget Reconciliation Act of 1986); HHS Office of Inspector General
2017
Two invoices at most
On 9 January 2017 eight Chinese government bodies, the National Health and Family Planning Commission and seven others, published the two-invoice system for drugs bought by public hospitals: at most one invoice from the maker to a distributor and one from the distributor to the hospital, to cut layers of markup and "money-laundering by invoice fraud". China had more than 13,100 drug distributors in 2017, and the top four held 37.6% of the market. A study of listed Chinese drug makers from 2014 to 2020 found that their selling expenses rose after the rule, an effect that had faded by the third year.
Lessonremoving middle layers moved their selling work onto the makers, at least for a time.
Credit
the National Health and Family Planning Commission, with the State Council's medical reform office, the China Food and Drug Administration, the National Development and Reform Commission and other ministries; study by Yi Ran and colleagues (2022)
2020
Hospital drug tenders rigged
On 9 December 2020 Japan's Fair Trade Commission (JFTC) filed criminal accusations against the drug wholesalers Alfresa, Suzuken and Toho and seven employees, who were indicted the same day; a fourth wholesaler, Mediceo, reportedly applied for leniency. In March 2022 the JFTC ordered the three to stop and to pay surcharges of 423.85 million yen in all, finding that in 2016 and 2018 they had agreed in advance which of them would win each drug category in the purchasing contracts for the 57 hospitals of the Japan Community Health Care Organization.
Lessonwhen a few large wholesalers bid for a hospital group's whole drug list, the work can be shared out among them before the bids are opened.
Credit
the Japan Fair Trade Commission and the Tokyo prosecutors
2021
Paying for unchecked opioid orders
Since a rule of 24 April 1971, every US distributor of controlled drugs must run a system that spots suspicious orders, those of unusual size, pattern or frequency, and report them to the Drug Enforcement Administration (DEA). A House committee found in 2018 that distributors shipped 20.8 million opioid pills over ten years to pharmacies in Williamson, West Virginia, a town of about 3,000, and nearly 9 million in two years to one pharmacy in Kermit, population 406, and that from 2013 the DEA brought no immediate suspension orders against distributors for nearly six years. In 2021 McKesson, Cardinal Health and AmerisourceBergen settled the states' and local governments' opioid suits for up to $21 billion over 18 years, with rules lasting until about 2032 on how they screen orders.
Lessonin focus groups with 46 pharmacists, Thomas Varisco and colleagues found that the order limits distributors set afterwards led to rationing of buprenorphine, a medicine that treats opioid addiction.
law.cornell.edud1dth6e84htgma.cloudfront.netnationalopioidsettlement.comebi.ac.uk
Credit
the state attorneys general and the plaintiffs' committee in the federal opioid litigation; the House Energy and Commerce Committee
Watch for
Shared data on opioid shipments
Under their 2021 settlement, McKesson, Cardinal Health and Cencora must set up and run, until about April 2032, an independent clearinghouse that gives all three and state regulators combined data on where opioids go and how often, so that each sees the others' shipments to the same pharmacy. McKesson's page on its monitoring program still describes the clearinghouse as something that "will be used".
Achieved whenthe settlement's independent monitor, or a state regulator, reports that the clearinghouse is running and that the three distributors use it to stop orders.
Claimed by
McKesson, on its program page as of November 2026; terms summarized by the Congressional Research Service, February 2025
Open
01
Pill-mill flags, scored alike
A public benchmark on the released ARCOS shipment records for 2006 to 2014. Its labels are the pharmacies that the DEA later acted against, taken from the DEA's published decisions, each matched to its ARCOS buyer number, with a fixed split by date. Every method, from a simple trailing-average limit to published machine-learning models, is scored on how many months before the DEA's action it flags each labeled pharmacy, and how many other pharmacies it flags each month.
The briefCredit · unclaimed02
A units-of-measure answer key
A public test set of at least 1,000 order lines for medical supplies written as hospitals and clinics write them ("2 BX exam gloves L", "1 CS syringes 3 mL"), each with the correct number of single units, worked out from the package levels (each, box, case and how many of one sit in the next) that makers report to the FDA's device database, GUDID. A script scores ordering systems and AI models on converting each line.
The briefCredit · unclaimed
Open 01 · the brief
Pill-mill flags, scored alike
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark on the released ARCOS shipment records for 2006 to 2014. Its labels are the pharmacies that the DEA later acted against, taken from the DEA's published decisions, each matched to its ARCOS buyer number, with a fixed split by date. Every method, from a simple trailing-average limit to published machine-learning models, is scored on how many months before the DEA's action it flags each labeled pharmacy, and how many other pharmacies it flags each month.
Why it is still open
Several methods exist, each scored on its own labels. Dimitris Bertsimas and colleagues published a semisupervised model for spotting over-ordering pharmacies in 2025; a Washington University team reported in 2022 that its anomaly detector was 100% precise on a set of known illicit transactions; the Washington Post's arcos tools walk through a trailing six-month limit on one Walgreens store that the DEA sanctioned in 2011; thresh, an open-source plugin posted on GitHub in August 2026, scores its own detectors on 18 months of made-up orders with 12 planted anomalies. None publishes a shared label set and split, and false flags matter: pharmacists told Varisco and colleagues that order limits led to rationing of buprenorphine.
Built with what exists
the ARCOS records through the Washington Post's arcos tools and the reconstructed dataset of Amico, Verginer and Schweitzer (Scientific Data, 2024); the Federal Register API for DEA decisions on pharmacy registrations; Python with pandas and scikit-learn; a language model such as Claude to pull pharmacy names and DEA numbers out of the decisions, checked by hand.
First step
Pull the DEA's decisions on pharmacies published from 2010 to 2016 and match 20 of them to ARCOS buyer records.
Done looks like
At least 100 labeled pharmacies, each with a link to its DEA decision; at least three methods scored on the same split for lead time and for flags per 1,000 pharmacies per month; code that reruns from the public data.
Open 02 · the brief
A units-of-measure answer key
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 at least 1,000 order lines for medical supplies written as hospitals and clinics write them ("2 BX exam gloves L", "1 CS syringes 3 mL"), each with the correct number of single units, worked out from the package levels (each, box, case and how many of one sit in the next) that makers report to the FDA's device database, GUDID. A script scores ordering systems and AI models on converting each line.
Why it is still open
GHX, which runs a large hospital ordering network, writes that errors in units of measure are "rampant" in the healthcare supply chain and estimates that 40% of a provider's item master data, the hospital's list of everything it buys, is inaccurate. A data-cleaning firm, Symmetric Health Solutions, notes that once an item has transactions, many ordering systems lock its packaging and unit of measure, so early mistakes last for years. A web search in November 2026 found vendor clean-up services but no public test set.
Built with what exists
the AccessGUDID API, which gives each device identifier's package type, quantity and the identifier it contains; Python; open ordering systems with unit conversion, such as ERPNext or Odoo; language models such as Claude or Gemini to read free-text order lines.
First step
Pull the GUDID records for 50 common exam gloves and syringes and draw each one's package tree by hand.
Done looks like
1,000 lines drawn from at least 200 device records, each answer agreed by two people; at least three systems or models scored on exact unit counts, with errors sorted by type (box read as case, each read as box, and so on).