Is AI easing worker shortages across OECD countries?

What the 38-country comparison actually measured

The 2026 database covers 38 countries. The most recent observation is 2024 for most of them, but it is 2022 for Chile and Norway; 2023 for Korea, the Slovak Republic, Switzerland and Türkiye; and 2025 for Australia and New Zealand. The headline tables therefore compare each country's latest available position rather than taking one simultaneous snapshot. This matters when AI adoption, hiring and economic conditions can change quickly.

For each country and year, the OECD builds an occupational imbalance score from six signals: hourly-wage growth, growth in hours worked, employment growth, the inverse unemployment rate, the annual change in underqualification and growth in online job postings. National labour-force surveys supply most inputs. Online postings generally come from Lightcast, whose coverage is available only from 2018 for most countries and is excluded where the series is too short. Inclusion requires at least wage, employment and hours-worked indicators.

The analysis operates mainly at the two-digit ISCO occupational level, merging occupations where samples are too small and mapping national classifications into the common system. Trends are separated from cyclical movements with a Hodrick-Prescott filter. Each signal is then standardised relative to other occupations in the same country and year. Signals receive equal weight except employment and online-posting growth, which receive half weights because their interpretation is more ambiguous. Positive scores mean workers are harder to find than for other occupations in that country—not that a fixed number of vacancies went unfilled.[1][2][4]

The job pattern is uneven, not a simple automation story

Across the OECD average, health, teaching and ICT professionals were in shortage from 2019 through 2023. Country paths diverged after that. Australia moved from a large ICT-professional shortage in 2021 to a surplus in 2025, and the United Kingdom moved from broadly balanced to surplus. France and the United States retained persistent ICT shortages. Those differences could reflect demand, worker supply, migration, training pipelines, industry structure or measurement; the indicator cannot assign a single cause.

The knowledge results underline the same ambiguity. Computers and electronics, mathematics, education and training knowledge are often in shortage, and administrative knowledge is in shortage in all included countries. Yet software skills such as programming and troubleshooting appear broadly in balance. The OECD warns that the skill measure aggregates across occupations, so a specialised skill used intensely by a smaller workforce can look balanced even when its core occupations are hard to fill. Employer surveys cited in the report separately suggest technical shortages can slow AI adoption, but the new indicator does not quantify that bottleneck directly.

Clerical support, food preparation, labouring and street-sales occupations are more often in surplus, and the report says automation and AI may be one contributor for routine clerical work. It also notes that diffusion may still be too limited for a clear economy-wide effect. A surplus classification is relative to other occupations and does not mean every person in that job is redundant, that every country shares the pattern or that a forecast of future job losses has been established.[1][3]

What this means for workers and employers

For workers, the most durable signal is not a narrow prompt-writing credential. Essential skills such as active listening, reading comprehension and critical thinking, together with social skills including co-ordination, persuasion, instructing and service orientation, appear in shortage across the OECD average. These capabilities travel across roles and help people question, interpret and act on automated outputs. Training that combines domain knowledge with judgement and communication is more defensible than teaching one product interface that may change quickly.

For employers, persistent shortage does not justify treating recruitment difficulty as a pure training failure. Wages, working hours, job security, location and workload all affect whether people take and keep a role. The report explicitly notes that poor job quality can be confused with scarcity. Before investing in an AI system as a labour substitute, managers should identify the tasks that are genuinely constrained, the expertise that must remain accountable and whether redesigning conditions or entry routes would solve the problem more safely.

The contrasting ICT results also argue against one global workforce plan. An organisation recruiting in France or the United States may face a different supply picture from one recruiting in Australia or the United Kingdom. Country-level dashboard data can guide a first assessment, but a local vacancy, pay and retention review is still needed. The indicator is best used to frame questions, not to set hiring quotas or justify redundancies by itself.[1][3]

Teachers, clinicians and public services face a double demand

Teaching and health shortages are especially consequential because AI adoption can add implementation work before it saves time. Teachers must decide when a tool supports learning, check outputs and help pupils build foundational skills. Clinicians must validate recommendations, explain decisions and monitor safety. If systems are introduced into already stretched teams without protected training time and clear accountability, the technology can shift work rather than remove it.

The OECD links persistent teaching demand partly to lifelong learning: even where school-age populations shrink, more adults may need upskilling and reskilling. That makes educators part of the response to AI-related labour change, not merely a workforce that technology might replace. Health demand has a different structural driver—ageing populations and rising care needs—while difficult working conditions intensify retention problems. Public policy therefore needs training capacity, flexible entry paths and better work quality alongside technology procurement.[1][3]

Why this is not a causal AI study

Every component is an indirect signal. Wage growth may reflect bargaining or inflation; employment growth can occur during both shortage and expansion; postings miss informal recruitment and can be duplicated; underqualification depends on education categories; and unemployment is tied to a person's previous occupation. Combining signals reduces reliance on any one weakness but does not turn the index into a vacancy census. Standardising within country and year also means the scores rank occupations relative to each other, so they cannot be used to compare absolute shortage levels across countries.

The skill translation adds another layer. Occupation scores are weighted with skill level and importance from O*NET, a detailed US database. Those profiles may not match job design in every country, particularly where technology, institutions and working practices differ. Survey respondents may also be skewed toward more educated workers. Country data sources, missing components, classification crosswalks and breaks in survey methods introduce further uncertainty despite the OECD's corrections and common framework.

Most importantly, the design cannot separate AI from the pandemic, demographic change, the green transition, interest rates, migration rules, education pipelines or sector-specific shocks. References to AI in the interpretation are plausible mechanisms supported by other evidence, not estimates produced by a natural experiment. The report is authoritative descriptive infrastructure; it should not be recast as proof that AI is saving or destroying a specified number of jobs.[1][2][3]

What evidence would change the assessment

The next update will be more informative if it shortens publication lags, preserves comparable online-posting coverage and separates entry-level from experienced roles. Linked employer–worker data could test whether AI adoption precedes changes in vacancies, pay, hours, retention and career progression within otherwise similar firms. Task-level measures would help distinguish the demand for an occupation from changes in the mix of work performed inside it.

For policy, the crucial test is whether specific interventions improve outcomes. Evaluations should compare alternative training and hiring routes, completion and job-placement rates, pay and retention—not merely enrolment. They should also ask whether AI tools reduce workload without weakening service quality. Evidence of sustained reductions in hard-to-fill vacancies, better working conditions and safe productivity gains would support a stronger claim that AI is easing shortages. Until then, the report's practical warning is clear: technology policy and workforce policy cannot be planned separately.[1][2][3][4]