AI agents may handle parts of a third of global task time—but finish very little alone
A new preprint maps model ratings of 18,796 US task statements into employment data for 146 countries. It estimates 35.2% of task time is at least partly delegable, while 2.5% is fully hands-off; both figures describe technical potential, not jobs lost or work already automated.
The short version
- What happened
- Two Anthropic models rated 18,796 O*NET task statements; the researchers mapped the consensus to 433 international occupation groups and employment data covering 146 countries and 2.34 billion workers.
- Who it affects
- For managers and workers, the study argues for examining individual tasks rather than labelling an entire occupation automatable. Most of the estimated potential still leaves a person responsible for an essential step or review.
- What remains unproven
- This is an unreviewed preprint. Its headline is an index produced by model judgments, not observed delegation, causal employment evidence or a forecast of jobs lost.
Keep in mind
This is an unreviewed preprint. Its headline is an index produced by model judgments, not observed delegation, causal employment evidence or a forecast of jobs lost.
O*NET describes US work. The crosswalk carries those task descriptions to international occupations, so it cannot capture how the same job differs across countries, employers or informal settings.
The two Anthropic models produced materially different ratings; the consensus rule is conservative but does not establish which model is right.
The human check covered 300 tasks, was led by a coauthor and did not preserve individual rater identities, preventing a human-to-human reliability estimate.
Country employment data have different reference years and levels of occupational detail. The paper reports no separate funding or competing-interest statement.
Source analysis · 1 linked source. Primary analysed source date: 22 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 5 min The evidence, methods, implications and limitations
Partial handoff dominated; fully hands-off work was rare
The researchers asked two language models to classify every task statement in O*NET 30.2 as no meaningful role, augmentation, partial delegation or full delegation. They then weighted those tasks within occupations, translated 923 US occupation codes into 433 ISCO-08 groups and combined them with the latest post-2018 occupational employment data available for 146 countries. Together, those country files represent 2.34 billion workers, although their reference years and occupational detail vary.
The employment-weighted result was 35.2% of global task time at partial delegation or above, with a 95% bootstrap interval of 33.9% to 36.7%. Full delegation accounted for 2.5%, with an interval of 1.9% to 3.2%. That distinction is the useful finding: an agent might draft, search or execute digital sub-steps while a person still checks the output, handles a physical step or carries legal responsibility. The paper does not show that employers have removed those human roles.[1]
The country gap reflects different job mixes, not different workers
Model-rated delegability rose from 22.5% of task time in low-income countries to 48.7% in high-income countries. The authors attribute most of that gap to employment composition: lower-income economies have more work in agriculture, crafts and elementary occupations where physical presence blocks software agents. Every country receives the same occupation score, so the atlas cannot show that a farmer, clerk or engineer in one country is inherently more or less delegable than the same occupation elsewhere.
For workers, low exposure is not automatically protection and high exposure is not a forecast of redundancy. Access to systems, regulation, employer choices, wages, trust and the cost of human review determine what is actually deployed. A comparison with Claude usage covered 101 countries and found a strong correlation with the index, but it measures one vendor's conversations per working-age adult—not delegated tasks, output quality or employment change.[1]
The distinction that matters
Three different claims about AI and a task
The study separates assistance from delegation so a headline about exposure is not mistaken for a hands-off job replacement.
Augmentation
AI drafts, searches or suggests while a person performs the task
Human stays involved at every step
Partial delegation
An agent completes substantial sub-steps from a goal
A person still performs or checks an essential step
Full delegation
An agent finishes the task to an acceptable standard
Human only assigns and accepts the result
The rating method is transparent—and still the largest uncertainty
The two model raters agreed reasonably well, but they did not produce the same world. Claude Sonnet alone yielded 35.3% at partial delegation or above; Haiku yielded 51.4%. The preregistered consensus used the lower level when they disagreed, pulling the headline toward Sonnet. On a 1,500-task sensitivity set, changing Sonnet's system prompt from a neutral rater to its default coding-agent wrapper lifted the share at partial delegation or above from 37.6% to 56.8%. The framing of the judge is therefore part of the measurement.
A blind human check agreed with the consensus at weighted kappa 0.85 across 300 tasks, but it was led by a coauthor with student help, did not record which person rated each item and cannot estimate human-to-human agreement. The index also closely tracks older AI-exposure measures, limiting its novelty as a ranking. Independent raters, country-specific task descriptions and observed workplace outcomes would show whether the partial-versus-full distinction predicts safe, useful delegation in practice.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
UNESCO · 11 min 27 sec · Background explainer
UNESCO Talks: Interview with 2024 Nobel laureate Professor Daron Acemoglu
YouTube loads only when you choose; playback starts when you press play.
Economist Daron Acemoglu discusses how choices about AI affect workers, inequality and society. Pair the interview with our comparison of task exposure, hiring and skills evidence.
Evidence check: An expert interview from 2024, offering a perspective on policy and work rather than a measured forecast of job losses.
Published 2024-10-14. Source checked 14 October 2026. Check captions on the publisher’s video page.
Read the AI and jobs guidePhoto: Dylan Gillis / Unsplash · Licence
Continue the story
Read next
Continue exploring ai agents: compare the finding and its evidence limits.
Do agentic tools now make up most business AI use in the EU?
Within OpenAI's Enterprise and Business customer sample, yes: Codex and ChatGPT Work produced 55% of included output tokens at the end of September. That is not a measure of all EU firms, completed work or productivity.
Read the next reportExplore the full source and evidence record
What this means for people
- For managers and workers, the study argues for examining individual tasks rather than labelling an entire occupation automatable. Most of the estimated potential still leaves a person responsible for an essential step or review.
Global context
The index extends task-level analysis to the occupational classification used by most countries and highlights a divide driven by economic structure. It should not be read as a country league table of worker replaceability.
What the evidence does not yet show
- This is an unreviewed preprint. Its headline is an index produced by model judgments, not observed delegation, causal employment evidence or a forecast of jobs lost.
- O*NET describes US work. The crosswalk carries those task descriptions to international occupations, so it cannot capture how the same job differs across countries, employers or informal settings.
- The two Anthropic models produced materially different ratings; the consensus rule is conservative but does not establish which model is right.
- The human check covered 300 tasks, was led by a coauthor and did not preserve individual rater identities, preventing a human-to-human reliability estimate.
- Country employment data have different reference years and levels of occupational detail. The paper reports no separate funding or competing-interest statement.
What to watch next
- Independent human rating of the full preregistered validation sample
- Country-specific task descriptions and better coverage of informal work
- Field evidence linking rated delegation levels with output quality, review time, wages and employment
- Re-estimation with models from different developers and stable system prompts
Living evidence record
Impact record IAI-095ANCB
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
23 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 23 October 2026
Research question, topics and regions
How much occupational task time current software agents could partly or fully complete, and how that model-rated potential varies with countries' employment mix
Topics: AI agents · Work · Task delegation · Global labour markets · Occupational measurement
Regions: Global · Latin America · Colombia
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.