AI Agents Can Fully Take Over Only 2.5% of World Work Time, Index Finds

About 35.2% of the task time in world employment can be at least partly handed to an AI agent in 2026, and only 2.5% can be handed over in full. [Fact] Those are the headline numbers of the Agent Delegability Index (ADI), a preprint posted to arXiv on October 22, 2026 by Cristian Espinal Maya of Universidad Nacional de Colombia and Santiago Jiménez Londoño of Universidad EAFIT. [Fact] The index rates all 18,796 O*NET task statements, maps them to 433 ISCO-08 occupation groups, and weights them by employment in 146 countries. [Fact]

Delegation, not exposure

Most indexes ask whether AI can speed up a task. ADI asks whether a worker can hand the task to an agent that browses, runs code, and edits files. [Fact] Each task gets one of four levels: no role, augmentation, partial delegation (the agent does substantial steps but a human still checks or finishes), and full delegation (a human only assigns and accepts). [Fact] Four barriers are recorded separately: physical presence, legal accountability, in-person trust, and access to systems an agent would not normally reach. [Fact]

At the occupation level, the result tracks the older exposure scores closely: the correlation with the OpenAI-led "GPTs are GPTs" measure is 0.93, above the 0.80 ceiling the authors preregistered as the point where ADI would add little. [Fact] The authors concede that the ranking of jobs is not new; the levels and barriers are. [Fact]

Clerks reach the top rung

Thirteen groups score the maximum, led by coding and proofreading clerks and contact centre salespersons; payroll clerks are among them. [Fact] The gap between partial and full delegation is where the jobs differ. Coding and proofreading clerks have 67.7% of task weight at full delegation; payroll clerks have 16.1%. [Fact] For payroll work, roughly five-sixths of the delegable task weight still ends with a human check. [Estimate]

Clerical work carries the fewest accountability flags of any major group (10.8%), which is why it is the one group where whole tasks go to the agent. [Fact] Managers score 81.8% at partial delegation or above but have very little weight at full delegation, because the models leave decisions with a person. [Fact] Full delegation peaks for tasks that take a human about 11.8 minutes. [Fact]

At the bottom, 16 groups score zero, including hand launderers, civil engineering labourers, stonemasons, and sewing machine operators. [Fact] The physical barrier flags 94.1% of task weight in elementary occupations. [Fact]

A divide by income

Delegable employment is 48.7% in high-income countries and 22.5% in low-income ones. [Fact] The United States sits at 51.9%, and India at 27.8%. [Fact] Every country gets the same US-built task scores, so the gap comes entirely from which jobs people hold. [Fact] The authors write that low delegability in poor countries "is not protection." [Claim]

The number moves with the rater

The two raters are Anthropic's Claude Sonnet 5.5 and Claude Haiku 5.5. [Fact] Rated by Sonnet alone, the world headline is 35.3%; by Haiku alone, 51.4%, about 46% higher. [Estimate] When Sonnet was addressed through the default Claude Code system prompt, which tells it that it is an agent with tools, its share of tasks at partial delegation or above rose from 37.6% to 56.8% on the same 1,500 tasks. [Fact] A blind human rating of 300 tasks was more generous than the models: of 50 tasks humans placed at full delegation, the models put 42 one level lower. [Fact]

What to prepare

For payroll clerks, data entry keyers, and receptionists, the line between partial and full delegation is the review step. Learning to audit an agent's batch of entries, spot the exceptions, and sign off on the result is the part of the job the index leaves with people. [Claim]

For proofreaders and telemarketers, whole tasks already rate as hand-over work, so moving toward roles that carry accountability or in-person trust is the safer direction. [Claim]

In office groups, access is the most frequent barrier, flagging between 64.2% and 74.0% of task weight. [Fact] Workers who control system permissions and data connections hold the gate that keeps a task below full delegation. [Claim]

For stonemasons and sewing machine operators, the barrier is physical and the index puts none of their task weight at partial delegation or above. [Fact]

Limits

ADI measures what an agent could technically do in October 2026, not what firms deploy, and the paper is a preprint without peer review. The raters, the agentic check panel, and the usage data used for comparison all come from Anthropic, and the human validation was led by a coauthor without per-rater records; the US task mix is applied unchanged to every country.

Sources

• Espinal Maya, C. and Jiménez Londoño, S., "The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries," arXiv:2610.30752v1, October 22, 2026 (CC BY 4.0): https://arxiv.org/abs/2610.30752

About this analysis

This article was written with AI assistance (Anthropic's Claude) from the full text of the arXiv paper. The index itself was rated by Claude models from the same developer. Ratios marked [Estimate] are our arithmetic on the paper's figures; occupation links point to our job pages.