AI exposure is much broader than observed job loss. The ILO estimates one in four workers has an occupation with some generative AI exposure, but only 3.3% of global employment is in its highest exposure category. Firm surveys through 2026 show limited realized headcount reductions, while emerging evidence suggests more pressure on some entry-level, highly automatable roles.
AI
AI exposure is much broader than observed job loss. The ILO estimates one in four workers has an occupation with some generative AI exposure, but only 3.3% of global employment is in its highest exposure category. Firm surveys through 2026 show limited realized headcount reductions, while emerging evidence suggests more pressure on some entry-level, highly automatable roles.
Key findings
• ILO estimates one in four workers globally is in an occupation with some GenAI exposure.
• The highest ILO exposure category contains 3.3% of global employment, including 4.7% of female and 2.4% of male employment.
• Exposure varies by economy: 34% of employment in high-income countries has some exposure, compared with 11% in low-income countries.
• US Census research found AI-related employment decreases in 2% of firms during its 2025-2026 reference period; 66% of AI-using firms used AI only to augment tasks.
• OECD found 83.0% of GenAI-using SMEs reported no change in staff need, 9.1% a decrease, and 5.5% an increase.
• UK ONS found 4% of AI-using businesses reported a headcount decrease in late 2025.
• Stanford payroll research found a 16% relative employment decline among workers aged 22 to 25 in highly AI-exposed occupations after firm-level controls, but broader labor indicators do not show an economy-wide AI jobs collapse.
Methodology and definitions
This review separates four concepts:
Exposure estimates whether AI could perform or assist tasks within an occupation. It is a technical and task-composition measure, not an observed employment outcome.
Augmentation means AI changes how a worker performs tasks while the role remains. Automation means AI substitutes for more of the task.
Displacement is an observed loss of a role or employment associated with AI. Establishing causality requires separating AI from interest rates, demand, outsourcing, restructuring, and other changes.
Hiring effect includes slower hiring, vacancy nonreplacement, or a shift toward experienced workers. It can occur without a layoff and may not appear immediately in headcount surveys.
The sources include the ILO’s 2025 refined occupational exposure index, US Census firm data, OECD’s representative SME survey, UK ONS business statistics, and Stanford labor-market research. Forecasts and employer expectations are excluded from observed job-loss totals.
One in four jobs is exposed, not eliminated
The International Labour Organization’s Generative AI and Jobs refined index, published May 20, 2025, combines task-level data, worker input, expert review, and model-assisted classification.
It found that one in four workers globally was in an occupation with some GenAI exposure. Only 3.3% of global employment fell in the highest exposure category. Women were more represented in that highest category, at 4.7% of female employment versus 2.4% of male employment, reflecting occupational concentration in clerical and administrative work.
Total exposure reached 34% of employment in high-income countries and 11% in low-income countries. Digitized professional and clerical work is more common in high-income economies, while many physical tasks remain outside current generative AI capabilities.
The ILO’s main conclusion is transformation rather than automatic replacement. Most occupations contain a mix of exposed and unexposed tasks. Technical capability does not establish that deployment is affordable, reliable, lawful, accepted, or able to perform the whole job.
What US firms report
The US Census Bureau’s Microstructure of AI Diffusion study uses nationally representative Business Trends and Outlook Survey data. During the November 2025 to January 2026 reference period, 18% of firms used AI in a business function, or 32% when weighted by employment.
Among AI-using firms, 66% reported using AI solely to augment tasks. AI-related employment decreases occurred in 2% of firms.
The paper found that broader functional deployment and operational investment were positively associated with employment decreases, while worker-level task integration was not significantly linked to headcount reduction after accounting for those factors. These are regression associations, not proof that AI caused each decrease.
The evidence suggests that firm-level restructuring can matter more for employment than individual chatbot use. A worker drafting faster does not automatically remove a position. A redesigned operation spanning several functions can change staffing decisions.
What SMEs report
The OECD’s 2025 SME workforce study surveyed 5,232 SMEs in seven countries in late 2024. Among firms using generative AI, 83.0% said it had no effect on the total number of staff needed. Nine-point-one percent reported a decrease and 5.5% an increase.
The survey did not ask the size of the staffing change, so the 9.1% and 5.5% figures cannot be netted into jobs lost or gained.
Task structure was associated with different outcomes. SMEs using GenAI mostly for recurring tasks were more than 50% more likely to report decreased staff need than those using it for one-off tasks. SMEs using it mostly for complex tasks were almost twice as likely to report increased staff need as those using it for simple tasks.
One interpretation is that recurring-task automation substitutes for capacity, while complex-task augmentation can expand what the firm offers. This remains an interpretation of survey associations.
The UK showed the largest split in the OECD sample: 18.3% of GenAI-using SMEs reported decreased staff need and 8.6% an increase. Differences in labor-market flexibility, sector mix, adoption behavior, and reporting could contribute.
UK official business evidence
ONS reported in October 2025 that 23% of businesses used some form of AI technology in late September, up from 9% when the question was introduced in September 2023.
Among AI-using businesses, 4% reported that headcount had decreased because of the technology. Seven percent of businesses planning adoption within three months expected a decrease.
By late December 2025, ONS reported 25% adoption. Again, 4% of adopters reported lower headcount due to AI, while 5% of planned adopters expected a decrease.
Expected reductions exceeded observed reductions in these snapshots. Expectations can influence hiring plans, but they are not realized job-loss counts.
The entry-level warning signal
Stanford Digital Economy Lab’s Canaries in the Coal Mine research uses high-frequency payroll data. It found a 16% relative employment decline among workers aged 22 to 25 in the most AI-exposed occupations after controlling for firm-level shocks. More experienced workers in the same occupations were stable or growing.
Young software developers showed a decline approaching 20% from their late-2022 peak. Declines were more concentrated in occupations where AI usage was classified as automative rather than augmentative.
This is evidence consistent with AI affecting entry-level demand. Causality remains contested. Stanford’s 2026 policy review notes that aggregate unemployment rose 0.77 percentage points since 2022 among the most AI-exposed quintile and 0.85 points among the least exposed. It also discusses interest rates, pandemic-era overhiring, and remote work as confounders.
Other research found deterioration in some exposed occupations began before ChatGPT. The most defensible conclusion is that aggregate effects remain limited while entry-level disruption in selected occupations is a credible emerging risk.
Practical SMB implications
SMBs should plan workforce changes by task, not by headline exposure. For each workflow, identify:
• Tasks AI can prepare, recommend, or complete.
• Tasks requiring context, accountability, trust, or physical action.
• New review, integration, and exception tasks.
• Capacity likely to be released.
• Demand that could absorb that capacity.
• Skills employees need to supervise or improve the workflow.
Avoid announcing headcount savings before a controlled workflow proves net capacity after review and correction. Use attrition or hiring plans only when evidence is stable and the service impact is understood.
A business AI operator can own the translation from task evidence to workflow design, controls, training, and outcome measurement. That role is different from declaring an occupation automatable.
Protect the entry-level learning ladder. If AI performs first drafts, research, and routine analysis, juniors still need structured opportunities to practice judgment, receive feedback, and learn exceptions. Removing all foundational work can weaken the future expert pipeline.
Limitations
ILO exposure is modeled capability, not displacement. Census, OECD, and ONS outcomes are self-reported at firm level. They may miss contractor changes, vacancy nonreplacement, or employment effects outside the surveyed period.
Stanford’s entry-level evidence is observational even with firm-level controls. Macro conditions and industry restructuring complicate causal attribution. All findings describe an early period in a rapidly changing technology cycle.
What would change our view
We would conclude that broad displacement was accelerating if representative payroll, vacancy, and firm data showed persistent declines concentrated after AI adoption and strongest in automative uses, with alternative causes ruled out. We would favor augmentation more strongly if exposed firms consistently expanded output and employment while worker transitions remained limited.
FAQs
How many jobs has AI eliminated?
No authoritative global count exists. Firm surveys show limited observed reductions, while some occupation-level research finds pressure on selected entry-level roles.
Does one in four jobs being exposed mean one in four jobs will disappear?
No. ILO exposure means at least some occupational tasks could be affected. Most jobs combine exposed tasks with work that still requires people.
What percentage of firms reduced staff because of AI?
US Census research found AI-related employment decreases at 2% of firms. UK ONS snapshots found 4% of AI-using businesses reported a headcount decrease. Definitions and periods differ.
Are entry-level jobs at greater risk?
Some evidence says yes. Stanford found a 16% relative employment decline among workers aged 22 to 25 in highly exposed occupations. Other economic factors make precise attribution difficult.
Should businesses plan layoffs from AI productivity estimates?
No. First measure net capacity after adoption, review, errors, and demand. Task speedups do not directly equal removable roles.
Get a 20-Minute AI Workflow Audit
AI Operator can map which tasks should be augmented, automated, or kept human and show the real capacity and skill implications before staffing decisions.