Verified demand makes the first move, with the worker’s consent.
Fromapplication→Todemand-led discovery
Today the worker searches for demand and applies to it. The proposal is that demand, described as capability requirements, finds verified capability and makes the first contact, with the worker’s consent. Matching technology already makes candidates visible and recommends them, sometimes with real gains. The stronger claim is not yet evidenced: that demand can originate real work without workers applying again and again. Whether it holds depends on design choices. Discovery should be opt-in, people without a work history need a way in, and rate and availability should stay the worker’s information.
Four different things
The concept proposes one of the largest changes in the architecture: from job application to capability discovery [CN]. The evidence review’s first discipline is to separate four things that often get blurred together.
| Stage | What happens | What the evidence shows |
|---|---|---|
| Visibility | A profile can be seen | Widespread |
| Recommendation | A system suggests a candidate or a job | Faster search in public portals, and more hiring and better-performing hires in some field studies [1] |
| Discovery | A buyer identifies relevant capability and makes contact | Partial, mostly through recruiters and platforms |
| Transaction | Both sides agree to paid work | Weak evidence that demand originates transactions without repeated applications |
Visibility is not recommendation, recommendation is not discovery, and discovery is not a transaction.
When demand describes outcomes
In the concept, an employer does more than post a job description. Its demand agent breaks an outcome into the capabilities required and the evidence that would show them, then looks for matching identities. On the worker’s side, the agent weighs fit, availability, rate, interest and where the person is trying to go [CN].
Public employment services already do parts of this. Flanders’ VDAB suggests skills missing from profiles and vacancies, and shows job seekers alternative occupations along with the gaps they would need to close. France’s public employment service built a tool that predicts which companies are likely to hire in the next six months. Singapore’s national jobs portal scores job fit on skills, and Malta runs its matching through VDAB’s engine as a shared service [1, 2]. Korea’s public portal cut the time users spent searching for job information from about ten minutes to a few seconds [1].
Where the value would come from
For employers, the savings would come in sourcing, screening, vacancy duration and mismatch. For workers, they would come in application effort, search time and repeated testing, along with wider access to demand. The review judges employer savings the more plausible commercial value, and treats worker-side gains as something to test rather than assume.
The field evidence on recommendation is encouraging, within limits. On one online labour market, recommending workers to employers substantially increased hiring, because acting on a recommendation was cheap and the suggested candidates resembled those employers would have chosen anyway [1]. At one large company, candidates selected by an algorithm were more likely to pass interviews, accept offers and perform well once hired [1]. Independent estimates of such tools’ effects remain rare, so figures that vendors publish are treated here as claims rather than evidence [1].
Go deeper
The risk of automating applications
More visibility can mean more work, not less. One platform’s own analysis finds that a skills-first approach expands talent pools nearly tenfold on average, a self-reported figure whose method the page does not give [5]. But a bigger pool is not the same as more hires, and it can add to employers’ review burden. AI screening can also reproduce bias through unrepresentative data, past human decisions, proxies for protected traits and feedback from recruiters’ clicks [1]. Applicants are wary too. In one U.S. survey, three in four people said they would not want to apply for a job where a computer program made the hiring decision [1].
An agent that only ranks candidates is automated recruiting. The proposal only differs if it reduces the total effort both sides spend reaching an agreement.
A network that only rewards history
Discovery built on evidence has a cold-start problem. Test credentials barely help workers win a first engagement [3]. And hiring algorithms trained on past hires tend to keep choosing the kinds of people chosen before. In one study, a screening algorithm designed to explore, giving weight to candidates it knew less about, improved both the quality and the diversity of the candidates it selected for interview [1].
So discovery needs deliberate entry routes. A share of every shortlist can be reserved for exploration. New entrants can carry a visible label instead of being excluded. And paid work samples or apprenticeships can serve as the first rung that produces evidence (Shifts 2 and 4).
A network that only rewards history can’t help people who don’t have one yet.
The worker keeps the first and last word
The concept keeps the worker in control of availability, rate, evidence and acceptance [CN]. As discovery becomes continuous, that control has to become concrete. Discoverability should be opt-in, tiered and time-limited. The worker should have a veto on any contact and rate floors held privately. And the worker should be able to say find me work without also saying make everything about me visible to everyone. The review rates the claim that continuous discoverability carries little risk as materially challenged by the evidence.
Referral networks show what is at stake. They carry trusted information and produce durable matches, but they also follow social lines [4]. Discovery built on open evidence could widen access to demand, or it could simply rank people who are already well connected even faster.
Demand as a formation signal
Once demand is visible, it can shape what people learn, because rising demand for a capability tells agents what is becoming valuable. The risk is herding. If every agent reads the same signal, many people may train for the same capability at once, and uniform algorithms can reduce match quality across a whole market [1]. Shift 7 takes this up.
Not our invention
Public AI matching in Europe and Asia, skills-based search, recruiter outreach, internal talent marketplaces, freelance marketplaces and referral networks already move information between workers and demand [1, 2, 4]. This series builds on them. The addition here is discovery built on verified evidence, on terms the worker controls, and aimed at completed transactions rather than more applications.
What could go wrong
- Spam.Contacts rise but transactions don’t.
- Pressure on pay.Visible rates and wider competition push pay down.
- Bias carried forward.Matching reproduces the patterns of past hiring [1].
- The invisible majority.Workers who don’t opt in lose out to those who do.
- Thin liquidity.A capability family may have too little demand for discovery to work (Shift 1).
How we’d test it
| The test | What we would measure | What would end the idea |
|---|---|---|
| Demand-originated discovery (Experiment 6) | For about 2,000 workers and 100–200 employers in one sector: who initiates, contacts, interviews, applications avoided, accepted paid work, search hours, pay, spam and complaints | Contacts rise but accepted work, earnings or worker welfare don’t. The review’s planning threshold is a 20% cut in worker search time or employer acquisition cost, with no fall in pay and no material rise in privacy complaints |
| Share of hires without an application | Share of accepted engagements with no application started by the worker | Close to zero after a full hiring cycle |
| Exploration | Hiring and later performance of new entrants surfaced through exploration slots | New entrants hired this way do worse than comparable hires |
How strong is the evidence?
- Established
- Algorithmic recommendation can speed search and increase hiring in bounded settings [1]
- Evidence-supported
- Automated matching can reproduce bias, and applicants distrust automated decisions at high-stakes stages [1]
- Hypothesis
- Verified capability lets demand originate transactions with less total effort on both sides
- Unknown
- Whether continuous discoverability raises earnings without harm to workers’ bargaining position or privacy
Sources
- [1]Broecke, Artificial Intelligence and Labour Market Matching, OECD Social, Employment and Migration Working Papers (2023)Checked; the field studies it summarises were not re-checked
- [2]World Bank, Solutions for Youth Employment, draft note on digital job-matching platforms in public agencies (2023)Checked (draft discussion note)
- [3]Kässi and Lehdonvirta, Do Microcredentials Help New Workers Enter the Market? Evidence from an Online Labor Platform, Journal of Human Resources 59(4) (2024)Checked
- [4]Schmutte, How do social networks affect labor markets?, IZA World of Labor (2016)Checked
- [5]LinkedIn Economic Graph, Skills-First report (2023)Checked; self-reported
- [CN]Open Commerce for Human Capital: concept document consolidated from the discussion (internal)Provided
Spotted a source we missed or a rule we misread? Tell us what we got wrong.