Agent Delegability Index (ADI): how much work can AI agents take over? · Balam Labs

· Balam Labs

3 min read Original article ↗

Agent Delegability Index · v2026.1 · updated October 2026

How much of the world’s work can AI agents take over?

The index rates every one of 18,796 occupational task statements for how much of it a current AI agent could carry out, from helping a person to finishing the job alone, and carries the ratings to 433 occupations and 146 countries.

35.2%

of task time in world employment is at least partly delegable to AI agents in 2026. Only 2.5% can be handed over in full.

95% interval

33.9

–

36.7

% · employment-weighted

The delegation divide. Each log point of GDP per capita adds 7.23 percentage points of delegable work.Open the country atlas

Full delegation2.5%

of task time can be handed over to an agent in full, given only the goal and access.

Most likely to be handed over11.8 min

Full delegation is most likely for tasks that take a person about twelve minutes.

Human checkκ 0.85

agreement of a blind human rating of 300 tasks, led by a coauthor with student raters, with the models.

Observed user 0.86

correlation between delegability and AI use across 101 countries (Anthropic Economic Index).

Leaderboard · ISCO-08 major groups

Office work is the most delegable; manual work the least

Share of task weight an agent can at least partly take over, averaged over unit groups with world employment weights. Whiskers: 95% bootstrap interval.

All 433 occupations

Look up an occupation or country

Find your job or country in the index

The delegation divide

Richer countries have more work agents can take over

Each flag is one of 146 countries. The ADI rises 7.23 points per log point of GDP per capita; income alone explains 71% of the spread.

Explore 146 countries

Ctrl/⌘ + scroll to zoom · drag to pan · Shift + drag to zoom to a box

ADI against GDP per capita, 146 countries0102030405060$1k$2k$5k$10k$20k$50k$100kGDP per capita, PPP (log scale)ADI: task time at least partly delegable, %SingaporeLuxembourgMozambiqueBurundiDashed line, the delegation divide: +7.2 pts per log point of GDP pc (R² 0.71)

Balam Labs · Agent Delegability Index v2026.1 · balamlabs.org/adi

Task length · Hypothesis H4

Full delegation peaks for tasks of about twelve minutes

The chance that an agent can finish a task end to end on its own rises with task duration up to 11.8 minutes of human work and falls for longer, multi-step tasks.

Methodology

Observed share at L3 (20 task-duration bins)Preregistered quadratic logit fit (H4)

Tasks rated L3 (full delegation) by human task duration, peaking at 11.8 minutes0%1%2%3%4%5%6%7%8%1 min5 min15 min1 h4 h16 h80 hMinutes a competent person needs for one instance of the task (log scale)Tasks rated L3 (full delegation), %Model peak at 11.8 min (4.4% of tasks at L3)

Hover or focus a dot to inspect one of the 20 task-duration bins. Very short tasks are often physical or interpersonal moments; long tasks bundle steps that still need human review.

Download

Balam Labs · Agent Delegability Index v2026.1 · balamlabs.org/adi

Latest

  • 22 Oct 2026Paper
  • 22 Oct 2026Report
  • 22 Oct 2026

    Dataset v2026.1 on Zenodo

    Ratings for every task, scores for every occupation and country, the rubric, and the code that builds them. doi:10.5281/zenodo.23552918

    Data
  • Oct 2026

    Preregistration on OSF

    Hypotheses, rubric and analysis plan were registered before the main scoring run. Deviations are listed in the methodology.

    Changes

For AI agents and scripts

Query and cite the index from code

Every record carries its citation. No key, no rate limit.

JSON/adi/v2026.1/isco/2512.json

One static file per occupation and per country.

Pythonpip install "adi-index @ git+https://github.com/Cespial/delegation-divide#subdirectory=herramientas/adi-index"

Then adi lookup 2512.

MCP serverclaude mcp add adi -- adi-mcp

Four tools: lookup, country profile, top occupations, cite.

llms.txt/llms.txt

A plain-text map of the index for language models. Source on GitHub.