Industries / Government, education and community / Government
Procurement
How governments buy goods, services, works and IT: advertising tenders, choosing suppliers, and signing and publishing contracts.
6
Milestones achieved, from 1981 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1981
Public buying opened to trade
Countries in the GATT, the trade agreement that came before the World Trade Organization (WTO), signed a code on government procurement in 1979, and it took effect in 1981. Its successor, the WTO Agreement on Government Procurement, has 22 parties covering 49 WTO members. It opens public purchasing worth more than US$1.7 trillion a year to international competition, by the WTO's estimate.
Credit
GATT members of the Tokyo Round; World Trade Organization
2002
Korea buys online
South Korea's Public Procurement Service had its online system, KONEPS, fully running by December 2002, with notices, bids, contracts and payment in one place. In its first year about US$20 billion of goods and services went through it, from more than 25,000 public agencies and 87,000 companies, and payments usually reached suppliers within four hours, against 14 days allowed by law, according to a World Bank note. By the agency's own count, 66.8% of Korea's public contracts by value, KRW 78.1 trillion, went through KONEPS in 2016.
LessonOne virus infection once shut the whole system down, the World Bank note reports, so a single channel for all public buying is also a single point of failure.
Credit
Public Procurement Service of Korea
2013
A website launched untested
The US federal health insurance marketplace, HealthCare.gov, opened on 1 October 2013, and many users could not get in. GAO found that the agency in charge, the Centers for Medicare & Medicaid Services, issued work before key requirements were known, used contracts that left cost overruns to the government, and launched the site "without verification that it met performance requirements." Money committed to the main contract grew from $56 million to more than $209 million between September 2011 and February 2014.
LessonA launch date is no substitute for a test that the system works.
Credit
Centers for Medicare & Medicaid Services (the buyer); findings by GAO
2015
A standard for contract data
The Open Contracting Partnership released the Open Contracting Data Standard (OCDS), a common format for publishing data on each stage of a public contract, as a draft on 18 November 2014 and as version 1.0 on 29 July 2015. Ukraine's ProZorro system is built on it, according to the European Commission.
standard.open-contracting.orginteroperable-europe.ec.europa.eu
Credit
Open Contracting Partnership
2016
All purchases in one system
Ukraine required central government buyers to use ProZorro, an online procurement system, from 1 April 2016, and all public buyers from 1 August 2016. Anti-corruption activists started it in May 2014. One central database holds the data, and buyers and suppliers work through commercial marketplaces linked to it. In 2022 KSE Institute, part of the Kyiv School of Economics, compared prices for the 40 kinds of goods public buyers buy most often. It found buyers paid more than 5% less in ProZorro tenders, and more than 10% less in its online catalog, than in purchases made without competition.
interoperable-europe.ec.europa.euopen-contracting.orginstitute.kse.ua
Credit
the Ukrainian activists who started it, guided by the e-procurement experts Tato Urjumelashvili and David Marghania
2023
One form for EU notices
Since October 2023, public buyers publishing notices on Tenders Electronic Daily (TED), the EU's online journal of public tenders, must use eForms. These are digital forms with the same fields and terms in every country, set by an EU regulation. The European Commission says a common standard and common terms can significantly improve the quality and analysis of the data.
Credit
European Commission
Watch for
AI review of bids
In a test, three AI agents reviewed a modeled US$8.5 million proposal for a fictitious federal agency against federal purchasing rules and flagged gaps, such as missing small-business subcontracting papers, while people made every decision. The authors call the results "striking" but give no figures for time saved or accuracy.
Achieved whenan agency uses such agents on a real competition and an outside reviewer publishes how often their findings matched the human evaluators'.
Claimed by
KJ Lian (Amazon Web Services) and Anil Chaudhry (US Department of Transportation), co-chairs of the ATARC Agentic AI Lab, writing in their own names, May 2026
Open
01
Brand names without equivalents
A screen that reads a tender's technical specifications and flags every brand, trade mark or proprietary product named without the words "or equivalent", together with a public test set to measure it. Use published tenders only, and do not bid on them.
The briefCredit · unclaimed02
Notice against documents
A check of whether the award criteria and weights a buyer enters in its eForms notice match those in its tender documents, across EU languages, with AI reading the documents and people checking every disagreement.
The briefCredit · unclaimed
Open 01 · the brief
Brand names without equivalents
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A screen that reads a tender's technical specifications and flags every brand, trade mark or proprietary product named without the words "or equivalent", together with a public test set to measure it. Use published tenders only, and do not bid on them.
Why it is still open
EU law bars specifications from naming a make or trade mark unless the subject requires it, and then only with the words "or equivalent". Anti-corruption guides list brand-name specifications as a sign of a rigged tender. The Spanish vendor LicitadIA says its AI detects brand mentions in drafts (April 2026), but I found no published accuracy figure for any such tool and no public count of how often unqualified brand names appear.
Built with what exists
the TED API and the document link in each eForms notice; language models that read scanned PDFs, such as Claude; a sample labeled by hand. (see the Note "Seven more flagged")
First step
Download 50 specifications for one kind of purchase, such as vehicles or office software, and label their brand references by hand.
Done looks like
1,000 technical specifications from at least five EU countries, every brand reference labeled by two people, a published share of brand references found and false alarms raised for at least two AI screens, and the share of specifications that name a brand without "or equivalent".
Open 02 · the brief
Notice against documents
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A check of whether the award criteria and weights a buyer enters in its eForms notice match those in its tender documents, across EU languages, with AI reading the documents and people checking every disagreement.
Why it is still open
eForms has fields for each award criterion's type, description and weight, which can serve as an answer key. Published tests of AI reading award criteria each cover one country and language: 167 Swiss calls for tenders in German (Bern University of Applied Sciences, 2024) and Spanish documents in the pAtChWoRK system (Universidad Carlos III de Madrid, 2026). I found no test that scores the documents against the notice across languages.
Built with what exists
the TED API; the eForms SDK field list (BT-539 criterion type, BT-540 description, BT-541 weight, BT-15 documents link); language models that read PDFs and tables; the Swiss team's GATE-CH dataset and code as a starting point. (see the Note "Seven more flagged")
First step
Pull 50 notices with weighted award criteria from TED, open their documents, and compare them by hand.
Done looks like
1,000 lots from notices in at least six EU languages, the share where the model's criteria and weights match the notice exactly, and every mismatch checked by a person and sorted into a model error or an error in the notice.