PaperlayerView exact revision on arXiv

Source-linked AI summary

The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries

Cristian Espinal Maya, Santiago Jiménez Londoño

arXiv:2610.30752v1econ.GNcs.AIcs.CYq-fin.EC

TL;DR

Existing measures largely score AI exposure, while the paper asks how much work AI agents can take over and extends measurement beyond the United States. ADI aggregates task ratings into occupational and country measures. It finds that 35.2% of task time in world employment is at least partly delegable, while only 2.5% is fully delegable.

  • Problem

    Existing measures chiefly assess AI exposure, and task-level measurement beyond the United States is limited.

    Source: §1 · p. 1

  • Method

    ADI aggregates task-level consensus ratings using task weights and carries them through occupational groups to countries.

    Source: §4.3 · p. 9

  • Results

    35.2% of task time in world employment is at least partly delegable to AI agents, while 2.5% is fully delegable.

    Source: §9 · p. 22

  • Takeaways & Limitations

    ADI adds delegation levels and barriers to exposure measures rather than providing a new ranking of occupations.

    Source: §9 · p. 22

    ADI measures technical capability, not deployment, and is therefore an upper bound on delegation.

    Source: §8.1 · p. 20

Abstract

from arXiv · show

35.2% of task time in world employment is at least partly delegable to AI agents in 2026, and only 2.5% can be handed over in full. Existing measures score exposure to generative AI, mostly for the United States. We build the Agent Delegability Index (ADI): two language models rate all 18,796 O*NET 30.2 task statements on a frozen, preregistered rubric of delegation levels and barriers; tasks are mapped to 433 ISCO-08 unit groups and weighted by ILOSTAT employment in 146 countries. Four findings. First, delegable employment rises with income, from 22.5% in low-income to 48.7% in high-income countries, a slope of 7.23 percentage points per log point of GDP per capita. This delegation divide is compositional: the physical barrier dominates the jobs poorer countries employ most. Second, at the occupation level ADI tracks exposure closely (r = 0.93 with Eloundou et al. and 0.89 with Felten et al.), above the preregistered bound of 0.80; ADI adds delegation levels and barriers on top of exposure, for the world, not a new ranking of occupations. Third, full delegation is rare and peaks for tasks that take a human about twelve minutes (11.8 minutes). Fourth, the preregistered adoption gap is not supported: across 101 countries, delegability and observed AI usage correlate at r = 0.86, and the gap between them is not wider in middle-income countries (p = 0.27). Model agreement is kappa = 0.77; a blind human rating of 300 tasks agrees with the model consensus at kappa = 0.85, and an agentic panel with web search at kappa = 0.87. We release the index, the atlas, the prompts and an MCP server at https://github.com/Cespial/delegation-divide.

1 Introduction

The paper introduces ADI to measure whether AI agents can take over tasks, addressing conceptual and geographic gaps in exposure-focused research. It finds that delegable employment rises with country income, largely because poorer countries employ more people in physically constrained work.

1 Introduction

  • ADI rates how much of each O*NET task an agent can execute, maps ratings to ISCO-08, and weights occupations by employment across 146 countries.Its rubric distinguishes delegation levels and records barriers separately from the level.Source: §1 · pp. 1–2
  • Prior measures emphasize AI exposure, while the paper argues that agents make whether workers can hand tasks over a distinct question from whether AI can speed them up.Existing cross-country measures also rely mainly on US occupational classifications or coarse bridges, while ADI maps tasks into ISCO-08.Source: §1 · pp. 1–2
  • ADI’s task-to-ISCO-08 bridge preserves 403 distinct task sets and supports an atlas for 146 countries, with demographic and informality breakdowns where microdata allow.The earlier matrix described in the passage collapsed occupations into nine task sets.Source: §1 · pp. 2–3
  • Occupation-level ADI correlates with exposure indices at r = 0.93 and r = 0.89, above the preregistered 0.80 bound; its added value is delegation levels, barriers, and global coverage.The comparisons are with Eloundou et al. (2024) and Felten et al. (2021), respectively.Source: §1 · p. 3
  • Delegable employment rises from 22.5% in low-income to 48.7% in high-income countries, with physical barriers in poorer countries’ occupational mix explaining the divide.The country-level regression slope is 7.23 percentage points per log point of GDP per capita.Source: §1 · p. 3

2 From exposure to delegation

ADI distinguishes assistance from partial or full task handoff and records barriers that can keep a task from full delegation. It differs conceptually from exposure measures, although occupation-level scores largely coincide with them.

2.1 Three layers

  • The four rubric levels run from no meaningful agent role through augmentation and partial delegation to full end-to-end completion; physical, accountability, trust, and access barriers are recorded separately.The reference technology is publicly available agents as of October 2026, without a body or robotics.Source: §2.1 · pp. 3–4, Figure 1 · p. 4

2.2 Why delegation is not exposure

  • Exposure measures how much AI touches work, whereas delegation asks whether an agent can do the work; a task can be exposed without being delegable, or delegable without much exposure.For example, a radiology report may be sped up while a licensed radiologist must still read and sign it; an agent with system access could reconcile a bank statement end to end.Source: §2.2 · p. 5

2.3 Comparison with existing measures

  • ADI is the only task-level measure described as natively in ISCO-08, rated for agents, validated against human and agentic panels, and released with agent-query tools.The comparison table also records each measure’s construct, unit, rater, classification, country coverage, and data openness.Source: §2.3 · p. 5

3 Data

The data combine O*NET task statements, a documented bridge to ISCO-08, and employment statistics from countries reporting occupational data. Supplementary sources support robustness checks and demographic breakdowns where available.

3.1 Tasks: O*NET 30.2

  • O*NET 30.2 supplies 18,796 task statements for 923 occupations, weighted within occupations by importance and frequency, with missing values imputed by the median.A task appearing under several occupations receives one rating and several weights.Source: §3.1 · p. 5

3.2 Bridge to ISCO-08

  • A documented three-step crosswalk maps O*NET-SOC to ISCO-08, covering 433 unit groups with 403 distinct task sets; the earlier matrix described collapsed 440 Colombian occupations into nine task sets.Three military groups lack O*NET tasks, while seven residual groups inherit tasks from their three-digit parent and are flagged as imputed.Source: §3.2 · pp. 6–7

3.3 Secondary task bank: ESCO

  • ESCO skills and competences form a secondary robustness bank, with 8,229 items across 424 unit groups rated using the same rubric.The paper treats ESCO as a robustness check because its units are skills rather than tasks.Source: §3.3 · p. 7

3.4 Employment by occupation: ILOSTAT

  • ILOSTAT employment data provide the country weights; 146 countries with data from 2019 onward enter the atlas, using the finest sufficiently complete ISCO-08 level each reports.The coverage table groups countries by World Bank income group and reports their finest ISCO-08 resolution.Source: §3.4 · p. 7, Table 2 · p. 8

3.5 Income and adoption

  • Income groups use the World Bank’s fiscal-year 2026 classification, while observed adoption is measured with country-level Claude usage per working-age adult from the Anthropic Economic Index.Occupation-level comparison scores are also carried to ISCO-08 where necessary.Source: §3.5 · p. 7

3.6 Household microdata

  • Four household surveys from Colombia, Brazil, Mexico, and Peru support sex, age, and informality breakdowns; Colombia’s monthly series is also used to relate delegability to wage dynamics.These surveys code occupation at four digits in national adaptations of ISCO-08.Source: §3.6 · pp. 7–8

4 Method

The preregistered method applies a fixed rubric to occupational tasks, aggregates model ratings into occupation and country indices, and tests validation, convergence, and hypotheses. Human and agentic panels provide additional validation, while bootstrap and placebo procedures assess uncertainty and occupational specificity.

4 Method

  • The pipeline moves from model task ratings to consensus, occupation weights, and country weights, with validation or bootstrapping at each step.The design was preregistered before the main pass.Source: §4 · p. 8

4.1 Rubric

  • The registered design uses one frozen rubric to rate each task on a four-level delegation scale and record task time, takeover share, barriers, and confidence.The rubric contains no example tasks, is hash-checked before the main pass, and pilot ratings are discarded.Source: §4.1 · pp. 8–9

4.2 Raters: language models

  • Two preregistered models rate each task separately, and repeated ratings measure stability because the interface does not expose a temperature setting.In the pilot, batching fell below the preregistered agreement floor, while the default tool-agent wrapper raised the mean level; the main pass therefore uses neutral, one-task calls.Source: §4.2 · p. 9

4.3 Aggregation

  • Task consensus uses the shared level when models agree and the lower level otherwise, then weights task ratings into occupations and employment-weights occupational groups into countries.Country estimates borrow within-group occupational structure from U.S. data for high-income countries and Colombian survey data for middle- and low-income countries; the latter is a stated loan where four-digit microdata were unavailable.Source: §4.3 · pp. 9–10

4.4 Human and agentic validation

  • Human validation compares model consensus with blind ratings of 300 stratified tasks, while a separate web-search panel rates all 1,500 validation tasks.The human rating was led by a coauthor, does not permit human–human agreement estimates, and has wider alpha confidence intervals than the registered 1,500-task design; the panel is explicitly not human validation.Source: §4.4 · pp. 10–11

4.5 Convergent validity

  • ADI correlates with Eloundou et al. (2024) at r = 0.93 and Felten et al. (2021) at r = 0.89, exceeding the preregistered 0.80 bound.The comparison tests convergence at the occupation level; the two correlations are above the preregistered range of 0.40–0.80.Source: §4.5 · p. 11

4.6 Hypothesis tests

  • The preregistered tests assess the income–delegability slope, whether the adoption gap is larger in middle-income countries, and a nonlinear relationship between task length and full delegation.The task-length model uses log human minutes and its square, with standard errors clustered by occupation and minutes winsorised at 1 and 4,800.Source: §4.6 · p. 11

4.7 Inference and placebo

  • Rating uncertainty is propagated with 1,000 task-level bootstrap draws, and a 1,000-shuffle placebo tests whether occupational indices retain occupation-specific information.Robustness checks vary the model, use ESCO skills or uniform task weights, and omit seven imputed groups.Source: §4.7 · pp. 11–12

5 The index

Partial delegation is common across tasks, while full handoff is rare. Delegability varies across occupations and barriers, peaks at intermediate task lengths, and is more sensitive to model choice than to weighting choices.

5 The index

  • Partial delegation is the most common task rating, full delegation is rare, and clerical and administrative occupations rank highest.The index reports task-weighted shares at L2 or above and at L3, plus a weighted mean share of working time delegable.Source: §5 · p. 12, Table 3 · p. 13

5.1 Tasks

  • 47.6% of valid task statements are rated L2 and 2.4% L3, so 50.0% reach at least partial delegation before weighting.L2 is the modal level; the models judge that agents can execute substantial parts of many tasks but can finish very few without a human step.Source: §5.1 · p. 12

5.3 Barriers

  • Physical barriers flag 72.2% to 94.1% of task weight across manual and service groups, while access is the most frequent flag in office groups.Accountability and trust flags are more frequent among professionals and managers; barriers can stack, and the figure reports each separately rather than assigning a single cause.Source: §5.3 · p. 13

5.2 Occupations

  • Managers lead L2+ delegation at 81.8%, but clerical support workers have the largest L3 share; manual groups range from 12.3% for elementary occupations to 18.4% for operators.Figure 3 aggregates unit groups using implied world employment weights.Source: §5.2 · p. 12, Figure 3 · p. 14

5.4 Task length (H4)

  • Full delegation peaks at about 11.8 minutes of human work, following an inverted U in task length.The minutes are model estimates recorded in the same call as the delegation level, not measured durations.Source: §5.4 · pp. 13–14

5.5 Robustness

  • The occupational index is not an assignment artefact: its unit-group standard deviation is 0.33 versus 0.09 under the permutation null (p < 0.001).The world headline is 38.6% with uniform task weights and 35.6% without imputed unit groups.Source: §5.5 · p. 14
  • Model choice matters more than weighting: the world headline is 35.3% with Sonnet and 51.4% with Haiku, while consensus is 35.2%.The preregistered consensus takes the lower level when models disagree, keeping its headline close to Sonnet; the difference is systematic rather than noise.Source: §5.5 · p. 14

6 The atlas: a delegation divide

The atlas estimates that 35.2% of task time in world employment is at least partly delegable, with full delegation rare. Delegability rises with national income, reflecting differences in the occupations countries employ.

6.1 The headline

  • 35.2% of world employment task time is at least partly delegable, while 2.5% is fully delegable, across 146 countries and 2,336.3 million workers.The headline is 34.9% under equal within-group weights, 35.6% with Colombian weights, and 37.7% without employment weighting of unit groups.Source: §6.1 · p. 15

6.2 Delegability and income (H1)

  • Delegable employment rises from 22.5% in low-income countries to 48.7% in high-income countries, with a slope of 7.23 percentage points per log point of GDP per capita.The divide reflects employment composition: poorer countries employ more workers in agriculture, crafts, and elementary occupations, where physical barriers constrain delegation.Source: §6.2 · pp. 15–16, Figure 4 · p. 15
  • With ESCO essential skills replacing O*NET tasks, the exploratory index reports a 38.0% world headline and a 4.66 slope, and correlates with the O*NET index at r = 0.88.This is an exploratory Sonnet-only analysis using the same country weights.Source: §6.2 · p. 16

6.3 Countries

  • The most delegable countries include Singapore at 64.2%, Luxembourg at 63.5%, and the United Kingdom at 54.9%; Burundi, Mozambique, and Madagascar are among the least delegable.Among large economies, Colombia is 38.3%, Mexico 36.8%, Brazil 39.9%, the United States 51.9%, and India 27.8%.Source: §6.3 · pp. 16–17

7 Validation, convergence and adoption

The ratings show substantial agreement, while occupation-level delegability closely tracks prior exposure measures. The preregistered adoption-gap hypothesis is not supported, and the interface wrapper materially affects ratings.

7.1 Agreement between models and over time

  • The two models agree at quadratic-weighted κ = 0.77 over 18,796 tasks, but Haiku’s L2-or-above headline is 51.4% against 35.3% for Sonnet.The preregistered consensus uses the lower level when the models disagree.Source: §7.1 · p. 17

7.2 Human and agentic validation

  • Human–model agreement is quadratic-weighted κ = 0.85 over 300 tasks; an agentic panel with web search agrees with the model consensus at κ = 0.87.The human rating was not an independent validation equivalent to the preregistered plan, and the agentic panel came from the same developer as the raters.Source: §7.2 · p. 18

7.3 Convergence with exposure indices (H3)

  • ADI correlates with exposure at r = 0.93 for Eloundou et al. (2024) and r = 0.89 for Felten et al. (2021), exceeding the preregistered 0.80 bound.ADI’s distinct contributions are delegation levels, barriers, and coverage beyond the United States, rather than a new occupation ranking.Source: §7.3 · p. 19

7.4 Observed usage and the adoption gap (H2)

  • The preregistered adoption-gap hypothesis is not supported: across 101 countries, delegability and usage correlate at r = 0.86, and the middle-income gap is not distinguishable from zero (p = 0.270).Usage measures conversations with one developer’s product, not delegated work.Source: §7.4 · p. 19

7.5 Sensitivity to the interface wrapper

  • The default system-prompt wrapper moves Sonnet’s L2-or-above share from 37.6% to 56.8%, with agreement between configurations of κ = 0.60.Haiku’s share moves from 54.9% to 56.2%, with κ = 0.76.Source: §7.5 · pp. 19–20

8 Limitations and ethics

ADI measures technical delegability, not actual deployment, and its country comparisons transfer U.S. task descriptions across countries. The authors also flag rater, framing, and ethical limits, with preregistration making the analysis checkable rather than making the measurement correct.

8.1 Technical delegability is not deployment

  • 35.2% of world task time is at least partly delegable, but ADI is an upper bound, not a measure or forecast of deployment.Deployment also depends on costs, system access, and accountability; employment-effect forecasts require adoption, prices, and institutions.Source: §8.1 · p. 20

8.2 O*NET describes work in the United States

  • The country comparison carries U.S. task mixes to other countries without adapting them, a limitation that is sharpest for agriculture.Replacing O*NET tasks with ESCO yields a GDP-per-capita slope of 4.66 for all groups and 6.92 for major group 6 only, against 7.23 with O*NET.Source: §8.2 · p. 20

8.3 Language models as raters

  • Two raters and the usage data come from one developer, while the 300-task human check is not a sample of workers or a full independent validation.The authors say three independent human raters on all 1,500 tasks remain the registered standard.Source: §8.3 · pp. 20–21
  • Ratings can vary across days and with framing: the default wrapper raised the mean level from 1.33 to 1.64 in the pilot.The authors use a neutral configuration for the main pass and report the wrapper comparison as a preregistered secondary result.Source: §8.3 · p. 21

8.4 Preregistration

  • Preregistration constrains confirmatory claims and makes deviations checkable, but does not establish that the measurement is right.The registration predates the main pass, and deviations are dated.Source: §8.4 · p. 21

8.5 Ethics

  • The authors warn that machine-centered delegability can reduce workers to tasks agents cannot do and treat care or trust as obstacles; the atlas raises distributional questions without settling them.They distinguish high delegability in rich countries from low delegability in poorer countries, which they say is not protection.Source: §8.5 · pp. 21–22

9 Conclusion

The conclusion reports that partial delegation is far more common than full handoff, and that delegability varies with countries’ employment composition. The appendix rubric operationalizes agent capability through levels, barriers, and task-time estimates.

9 Conclusion

  • 35.2% of world task time is at least partly delegable, while only 2.5% can be handed over in full.The authors say whole-task handoff occurs mostly for tasks taking a human about twelve minutes.Source: §9 · p. 22
  • Delegable employment ranges from 22.5% in low-income countries to 48.7% in high-income countries, reflecting poorer countries’ greater employment in work with a physical barrier.The authors caution this is not evidence that the same task is less delegable in poorer countries, nor that low delegability is protection.Source: §9 · p. 22
  • ADI tracks exposure closely, but its contribution is to add delegation levels and barriers across countries rather than provide a new occupation ranking.The conclusion reports correlations of 0.93 with Eloundou et al. and 0.89 with Felten et al., both above the preregistered bound.Source: §9 · p. 22

A Scoring rubric

  • The rubric rates technical feasibility with publicly available agents as of October 2026, assuming software access but no physical body or robotics.It distinguishes no meaningful role, augmentation, partial delegation, and full delegation, and records physical, accountability, trust, and access barriers separately.Source: A Scoring rubric · p. 28
  • For each task, the rubric requests a delegation level, the share of task time delegable, typical human minutes, four barrier flags, and confidence.It asks raters to assess technical feasibility, not current employer practice or legal permission.Source: A Scoring rubric · p. 28

B Crosswalk O*NET to ISCO-08

The appendix documents the reproducible bridge from O*NET 30.2 task statements to ISCO-08 unit groups.

B Crosswalk O*NET to ISCO-08

  • The O*NET-to-ISCO-08 bridge is implemented in a script, with counts printed in its log and reproducible from the listed input files.The appendix identifies the bridge script as datos/construir_crosswalk.py.Source: B Crosswalk O*NET to ISCO-08 · p. 28

B.1 Inputs

The crosswalk appendix identifies O*NET task and rating inputs, an O*NET-SOC version crosswalk, and a SOC-to-ISCO-08 crosswalk.

B.1 Inputs

  • O*NET 30.2 supplies 18,796 task statements for 923 occupations, alongside importance and frequency ratings and occupation titles.The cited O*NET materials are the Task Statements, Task Ratings, and Occupation Data.Source: §B.1 · p. 28
  • The O*NET-SOC 2010-to-2019 crosswalk contains 1,110 links from 2010 codes to 1,012 2019 codes.Source: §B.1 · p. 28
  • The Hardy et al. (2018) SOC 2010-to-ISCO-08 crosswalk contains 1,131 rows linking 839 six-digit SOC codes to all 436 ISCO-08 unit groups.Source: §B.1 · p. 29

B.2 Steps

The task-to-ISCO chain weights O*NET tasks within occupations, bridges taxonomies, and aggregates the results into unit-group task sets. It documents exclusions and imputations where the mapping chain leaves gaps.

B.2 Steps

  • Each task’s within-occupation weight is its importance multiplied by its frequency, with missing scale values imputed by the median.The weights are normalised to sum to one within each O*NET-SOC 2019 occupation.Source: §B.2 · p. 29
  • The chain reverses the official O*NET-SOC crosswalk, dividing a 2019 occupation’s mapping weight equally when it descends from multiple 2010 codes.This applies to 97 cases.Source: §B.2 · p. 29
  • Two absent SOC codes use documented fallback matches, while four post-2010 occupations are assigned to SOC 15-1199 and split evenly across two ISCO groups.The surgical-technologist fallback is a group-level approximation rather than a conceptual match.Source: §B.2 · p. 29
  • 30,549 ISCO–task pairs cover 18,796 distinct tasks and 433 unit groups, with a median of 48 tasks per group.The groups contain 403 distinct task sets.Source: §B.2 · p. 29

B.3 Exclusions and imputations

  • The three armed-forces groups are excluded because they have no O*NET tasks; seven other groups inherit the task set of their three-digit minor group.The seven imputed groups are flagged so they can be dropped in the robustness check.Source: §B.3 · pp. 29–30

B.4 Why not the earlier matrix

The paper uses a finer-grained occupation-to-ISCO bridge than the authors’ earlier matrix, which could not distinguish occupations such as surgeons and statisticians. The release includes versioned data and tools for accessing and citing the index.

B.4 Why not the earlier matrix

  • The bridge preserves 403 distinct task sets across 433 groups, unlike the earlier matrix’s 9 task sets across 440 occupations, which could not separate surgeons from statisticians.The paper applies the same chain to occupation-level comparison indices.Source: §B.4 · p. 30

C.1 Citation

  • The paper describes index version v2026.1; later versions re-rate tasks against a later reference technology and are not interchangeable.Direct data use should cite the dataset DOI together with the paper.Source: §C.1 · p. 30

C.2 Files

  • The release provides occupation-, country-, and task-level files, plus model ratings and a claims file linking each paper number to its value and producing script.The occupation file includes index measures, bootstrap intervals, barrier shares and task counts; the country file includes employment and reference-year information.Source: §C.2 · pp. 30–31

C.3 Static JSON

  • Static JSON documents let tools or agents fetch a single occupation or country record without downloading the dataset.Each document includes the version, licence and citation string.Source: §C.3 · p. 31

C.4 Python package

  • The Python package ships the index as data, works without network access, and exposes lookup, search, country, ranking and citation operations.It provides the same operations as the server.Source: §C.4 · p. 31

C.5 MCP server

  • The MCP server provides occupation lookup, country profiles, top-occupation lists and citation retrieval, with citation strings included in responses.Occupation lookup accepts an ISCO-08 code, an ESCO identifier or a free-text title.Source: §C.5 · p. 31

C.6 llms.txt

  • The repository and dataset site provide llms.txt and llms-full.txt summaries intended for language-model readers.The files cover the paper, headline number, file layout and citation.Source: §C.6 · p. 32

C.7 Licences

  • Code is MIT-licensed, while data, figures and the rubric use CC BY 4.0; input datasets retain their own licences.O*NET task statements are redistributed verbatim, while other inputs are provided only as derived aggregates.Source: §C.7 · p. 32
Loading 2610.30752v1…