The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries
B · Useful, limited Preregistered, open index from two Colombian university researchers with global reach, but its ratings come from two language models of one developer and the human check was weakened from the registered design.
Summary based on the full text·Original: about 60 minutes to read
In short
An arXiv preprint (dated 23 October 2026) by Cristian Espinal Maya (Universidad Nacional de Colombia) and Santiago Jiménez Londoño (Universidad EAFIT). It builds an Agent Delegability Index: two Claude models rate every O*NET task on four levels (no role, augmentation, partial delegation, full delegation) and four barriers (physical, accountability, trust, access), then weight results by employment in 146 countries. About a third of world task time is partly delegable and very little fully. Delegable work rises with national income because poorer countries employ more people in physical work.
Key numbers
- 35.2% / 2.5%share of world task time at least partly delegable vs fully delegable to AI agents in 2026
- 22.5% vs 48.7%delegable employment in low-income vs high-income countries
- 18,796O*NET 30.2 task statements rated, mapped to 433 ISCO-08 unit groups in 146 countries
- r = 0.93occupation-level correlation of the index with the Eloundou et al. exposure measure
- 11.8 minuteshuman task duration at which full delegation peaks
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- World: 35.2% of task time at partial delegation or above; 2.5% at full delegation.
- By income: 22.5% in low-income, 31.0% in middle-income, 48.7% in high-income countries; slope of 7.23 percentage points per log point of GDP per capita. Physical barriers drive the gap.
- At occupation level the index tracks exposure closely (r = 0.93 with Eloundou et al., 0.89 with Felten et al.), so its added value is the levels, the barriers and the global coverage, not a new ranking.
- The preregistered "adoption gap" hypothesis failed: across 101 countries, delegability and observed AI usage correlate at r = 0.86, and the gap is not wider in middle-income countries (p = 0.270).
- Model agreement κ = 0.77; a blind human rating of 300 tasks agrees with model consensus at κ = 0.85; an agentic panel with web search at κ = 0.87.
- Released: index, country atlas, prompts, a Python package and an MCP server.
How they got it
Claude Sonnet 5.5 (primary) and Claude Haiku 5.5 rated all tasks under a frozen, hashed rubric via the Claude Code command line. Tasks were weighted by O*NET importance and frequency, mapped to ISCO-08 through documented crosswalks and weighted by ILOSTAT employment. Usage comparisons use the Anthropic Economic Index.
Limits
- The index measures technical delegability as of October 2026, an upper bound, not deployment or forecast employment effects.
- Raters and usage data come from one developer, so agreement may reflect shared blind spots; the authors flag possible circularity.
- The human validation was cut from three independent raters on 1,500 tasks to a coauthor-led group on 300, with no human-human agreement measure.
- O*NET describes US work; it is especially thin for agriculture, which matters for low-income countries. An ESCO substitution lowers the income slope to 4.66.
Why it matters for companies
The split between partial and full delegation is the useful frame for workforce planning: most exposed work changes shape rather than disappearing, and the residual human part is often checking and accountability. The barrier codes (accountability, trust, access) give a checklist for why a task that looks automatable stays with people. Treat country and occupation figures as model opinion with a documented method, a starting point for local analysis rather than evidence of what firms are doing.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Work and AI, Occupations