How the index is built

The index asks one question of every task a worker does: how much of it could a current AI agent carry out on its own, given only the goal and access? It answers on a fixed scale, for every task statement in O*NET, and adds the answers up for occupations and countries.

Motivation

Existing measures score exposure: whether a language model could make a task faster for the person who does it. Agents change the question. A system that runs software, reads and writes files and calls other systems can take a task over, not just assist with it. Delegation is a change of owner, and exposure indices do not measure it. They also cover mostly the United States, while most of the world’s workers live elsewhere, in countries whose jobs look very different.

What it measures

Each task gets one of four levels. The levels form a ladder: at L1 the machine contributes, at L2 it executes, at L3 it finishes.

L0 No meaningful role

The agent cannot perform or materially speed up any substantial part of the task. 25.9% of task statements.

L1 Augmentation

A human performs the task while the AI assists step by step with drafts, searches or suggestions; the human produces the output. 24.1% of task statements.

L2 Partial delegation

An agent executes substantial sub-steps end to end from a goal, while a human still performs at least one essential step or reviews the work before use. 47.6% of task statements.

L3 Full delegation

An agent completes the whole task end to end, to the quality a competent worker would accept, given only the goal and access. 2.4% of task statements.

Alongside the level, the rubric flags four barriers that keep a task from moving up the ladder:

• Physical presence: the task needs presence or manipulation of objects, people or places.
• Accountability: law, licence or liability requires a human to perform or sign.
• Trust and care: the task depends on in-person trust, care or persuasion.
• Access to systems: the task depends on systems, data or places an agent would normally not reach.

It also records the share of the task’s working time an agent could take over and the minutes a competent person needs for one instance. Three indices follow, each a weighted share of an occupation’s task time:

Results

35.2% of task time in world employment is at least partly delegable to AI agents in 2026 (95% interval 33.9–36.7%), across 2,336.3 million workers in 146 countries. Only 2.5% can be handed over in full. The index goes from 22.5% in low-income countries to 48.7% in high-income countries, and from 14.7% in Burundi to 64.2% in Singapore. Explore it by occupation and by country.

Key takeaways

• A delegation divide. Each log point of GDP per capita adds 7.23 percentage points of delegable work (95% interval 6.5–7.8, R² = 0.71, 146 countries).
• The divide is compositional. Poorer countries employ more people in farming, crafts and elementary jobs, where the physical barrier is flagged on 94.1% of task weight, against 21.3% for managers.
• Delegation tracks exposure. Across occupations the index correlates with Eloundou et al. at r = 0.93 and with Felten et al. at r = 0.89, above the preregistered bound of 0.80. What it adds is how far delegation goes, what stops it, and global reach.
• Full delegation peaks for tasks of about 11.8 minutes of human time and falls for longer ones.
• Use follows delegability. Across 101 countries observed AI use correlates with the index at r = 0.86; the preregistered adoption gap in middle-income countries is not supported.

Method

Rubric and raters

A single rubric, frozen and hashed at registration, asks for one JSON object per task: level, share, human minutes, four barrier flags and a confidence label. It contains no example tasks, so no task can anchor another. Two models rate all 18,796 O*NET 30.2 task statements, one task per call: Claude Sonnet 5.5 as primary rater and Claude Haiku 5.5 as secondary, in a neutral configuration with no tools. The consensus level is the level both give when they agree and the lower one when they do not.

Aggregation

Task weights are importance times frequency, normalised within each O*NET occupation and carried through a SOC-to-ISCO-08 bridge to 433 unit groups. A country’s index is the employment-weighted mean over the occupational groups it reports to ILOSTAT, at two digits for 126 countries and one digit for 20. Within a reported group, unit groups are weighted with microdata: United States OES for high-income countries, Colombia’s GEIH for the rest, a stated loan.

Country decomposition and position relative to income

Country pages split the index into the nine ISCO-08 major groups: each group’s share of employment, its index with that country’s mix of unit groups, and its contribution in points, which sum to the country’s ADI. The “vs income” figure is the country’s ADI minus the delegation-divide line (ADI regressed on log GDP per capita, slope 7.23) at its GDP per capita; within ±1.5 points it reads as in line with income. Barriers are the employment-weighted share of task weight carrying each flag. These are site calculations from the published v2026.1 figures and are not part of the dataset on Zenodo or the API.

Uncertainty

Every interval is a 95% percentile interval from 1,000 bootstrap draws over tasks, resampled within occupation, that recompute the indices, the headline and the regressions. A placebo that shuffles ratings across occupations confirms that the index carries occupation-specific information.

Validation

Values above 0.80 indicate strong agreement.

Limits

The ratings are judgements by language models about what agents can do, checked against people on a sample; they are not observations of agents at work. The human rating was led by a coauthor, and because individual contributions were not recorded it is analysed as one rating. O*NET describes United States jobs, and the same occupation may be organised differently elsewhere. The within-group weights for middle- and low-income countries come from one country. The breakdowns by sex, age and formality exist only for Colombia and are exploratory.

Updates

• 22 Oct 2026

  v2026.1 · first release

  Index, atlas, report and dataset (doi:10.5281/zenodo.23552918). Hypotheses, rubric and analysis plan were preregistered on OSF (doi:10.17605/OSF.IO/9ZNTF) before the main scoring run.

• D4 · 22 Oct

  Household microdata: Colombia only

  Breakdowns by sex, age and formality are computed only for Colombia with GEIH 2024–2026. Mexico’s ENOE codes occupation at one digit; Brazil’s PNADC and Peru’s ENAHO are not yet processed. The comparisons with the ILO 2025 index and the Agentic Adoption Index are deferred.

• D3 · 21 Oct

  Four country-years excluded for data quality

  Suriname, Senegal, Comoros and Hong Kong report occupations covering less than 75% of employment or fewer than six major groups. The rule was added after seeing results; with all 150 countries the headline is 35.2% and the slope 7.30, against 7.23.

• D2 · 21 Oct

  Validation redesigned before any result was opened

  Three human raters on 1,500 tasks became a blind rating of 300 tasks, led by a coauthor with student raters, plus an agentic panel reported as non-human validation.

• D1 · 20 Oct

  Within-group employment weights

  No country publishes four-digit employment in ILOSTAT, so unit groups are weighted with US OES for high-income countries and Colombia’s GEIH for the rest. Equal weights give 34.9%.

Full record of deviations: docs/DESVIACIONES.md (Spanish). The rubric is reproduced in Annex A of the report.