Industries / Finance and business services / Staffing and recruiting
Matching people to jobs
Deciding which people and which openings belong together, using shared lists of occupations and skills and software that recommends jobs to job seekers or candidates to employers.
7
Milestones achieved, from 1939 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1939
A dictionary of occupations
The US Department of Labor published the first Dictionary of Occupational Titles in 1939, from data collected by occupational analysts, to meet "the demand of an expanding public employment service for standardized occupational information". New editions followed in 1949, 1965 and 1977. By its last update in 1991 it sorted jobs into over 12,000 titles, by the Bureau of Labor Statistics' count.
Credit
US Department of Labor
1998
O*NET replaces the dictionary
A 1980 National Research Council review found that nearly two-thirds of the Dictionary's job descriptions rested on observing fewer than two jobs, and that its titles leaned toward manufacturing. The Department of Labor set up an advisory panel in 1990 and published the first ONET database in 1998. A data collection program began in 2000, and its first update came in April 2003. The current release, ONET 31.0 of August 2026, lists 1,016 occupation titles.
LessonA job list built from watching a few workers fell behind the work it described, and its replacement needed a standing program of data collection.
Credit
US Department of Labor, Employment and Training Administration; National Research Council
2017
Europe's shared skills list
The European Commission published the first full version of ESCO, a multilingual list of occupations and skills for the EU job market, on 28 July 2017. It now describes 3,039 occupations and 13,939 skills in 28 languages, including Ukrainian and Arabic; the current version, 1.2.1, was last updated in 2025.
Credit
European Commission
2017
Recommended recruits fill more jobs
In a field experiment on an online labor market, employers with technical job openings who were sent algorithmic recommendations of workers to recruit filled 20% more openings than a control group. The recommended recruits were much like those employers found on their own, and the effect was largest for openings that drew few applicants.
Credit
John J. Horton (Journal of Labor Economics, 2017)
2018
Advice broadens job search
Economists invited job seekers to search for real jobs on their own website for twelve weekly sessions. For half of them, the site used labor market data to suggest other occupations and the jobs in them. This widened the jobs they considered and raised their number of job interviews, most of all for people who had searched narrowly and been out of work for a few months.
Credit
Michèle Belot, Philipp Kircher and Paul Muller (Review of Economic Studies, online 2018)
2023
Recommendations tested in Sweden
Three economists built a recommender from job seekers' histories of viewing ads and ran it on Sweden's largest online job board, in a randomized trial of 59 million recommendations that split both job seekers and vacancies into groups. It raised employment by 0.4% within a few months, and by 0.8% among unemployed job seekers, with no significant effect on how good the matches were. Workers with weaker prospects and jobs that drew less attention gained the most.
Credit
Thomas Le Barbanchon, Lena Hensvik and Roland Rathelot
2024
Degree cuts change little
Many large US employers announced that they would hire for skills rather than degrees. In February 2024 the Burning Glass Institute and Harvard Business School's Project on Managing the Future of Work reported that at the firms studied the change showed up in "not even 1 in 700 hires" in 2023. About 45% of firms had changed "in name only" and about one fifth slid back; the leading firms raised their hires without degrees by nearly 20%.
LessonRemoving a degree line from a posting did not by itself change who was hired.
Credit
Joseph Fuller, Matt Sigelman and Alex Martin
Watch for
Ranking jobs people can get
In two randomized field experiments with France's public employment service, six researchers found that deployed recommenders rank jobs by clicks, applications or hires, not by the job seeker's welfare. They report that ranking jobs by both how much the seeker wants them and the chance an application succeeds "substantially outperform[s] existing approaches".
Achieved whenThe study passes peer review with effects on hiring reported, or the employment service adopts the ranking and an outside evaluation reports its employment effect.
Claimed by
Guillaume Bied, Philippe Caillou, Bruno Crépon, Christophe Gaillac, Elia Pérennes and Michèle Sebag, in a preprint, March 2026
Open
01
Hourly job matching test
A test set of 2,000 made-up job seekers for hourly work (warehouse, retail, care, food service, driving), each with hard limits such as licenses held, shifts they can work and how far they can travel, plus stated wishes. Pair them with 500 real public postings whose requirements are already coded. An answer list, computed by fixed rules, gives the postings each seeker both qualifies for and wants. Score job recommenders on how many of those they place in their top 20, and how many of their top 20 break a hard limit.
The briefCredit · unclaimed02
Ukrainian and Arabic skill tags
500 public job ads written in Ukrainian and 500 written in Arabic, each sentence tagged by two native speakers with the ESCO skills it asks for, published with the annotators' agreement rate and scores for current skill-extraction models.
The briefCredit · unclaimed
Open 01 · the brief
Hourly job matching test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test set of 2,000 made-up job seekers for hourly work (warehouse, retail, care, food service, driving), each with hard limits such as licenses held, shifts they can work and how far they can travel, plus stated wishes. Pair them with 500 real public postings whose requirements are already coded. An answer list, computed by fixed rules, gives the postings each seeker both qualifies for and wants. Score job recommenders on how many of those they place in their top 20, and how many of their top 20 break a hard limit.
Why it is still open
The Bied team found that deployed recommenders are tuned to clicks, applications or hires rather than to what serves the job seeker. The closest open synthetic set, from the JobRec study (February 2026), covers computer science roles only, with 5,000 candidates and 100 postings, and does not model licenses or shifts. No such test for hourly work was found.
Built with what exists
Canada's Job Bank open data, which codes each posting's occupation, education and experience level, and whether it involves day, evening, night, weekend, on-call or shift work; a language model to write the seekers; WorkRB, an open benchmark toolkit for work-related AI, to run the models; Python.
First step
Take one month of Job Bank postings for one province, keep 500 hourly postings, and write 50 seekers by hand to test the answer-list rules.
Done looks like
Seekers, postings and answer lists published, with scores for at least five recommenders (keyword search, two embedding models, two language-model rankers), and a person checking 100 answer-list entries by hand.
Open 02 · the brief
Ukrainian and Arabic skill tags
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
500 public job ads written in Ukrainian and 500 written in Arabic, each sentence tagged by two native speakers with the ESCO skills it asks for, published with the annotators' agreement rate and scores for current skill-extraction models.
Why it is still open
ESCO's skill names exist in both languages. WorkRB (March 2026), an open benchmark for work-related AI, scores skill extraction against ESCO's names in 28 languages, but its code states that the job-ad sentences are "always English"; only the skill names are translated. Searches for tagged job-ad sets in Ukrainian or Arabic found none.
Built with what exists
ESCO 1.2.1 downloads in Ukrainian and Arabic; an annotation tool such as Label Studio or doccano; WorkRB to run and score models; multilingual language models.
First step
Take 50 ads from the Ukrainian part of the Djinni dataset on Hugging Face, 27,461 IT job descriptions under the MIT license, and have two annotators tag them.
Done looks like
1,000 ads tagged, agreement between annotators reported for each language, at least four models scored, and the set submitted to WorkRB as a new task.