Datasets:

Cespial
/

adi

license: cc-by-4.0
pretty_name: Agent Delegability Index (ADI) v2026.1
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-classification
  - text-classification
tags:
  - ai-agents
  - delegation
  - automation
  - labor-economics
  - occupations
  - isco-08
  - onet
  - ilostat
  - economics
  - llm-as-a-judge
  - delegation-divide
configs:
  - config_name: adi_isco
    default: true
    data_files:
      - split: train
        path: adi_isco.parquet
  - config_name: adi_pais
    data_files:
      - split: train
        path: adi_pais.parquet
  - config_name: adi_tarea
    data_files:
      - split: train
        path: adi_tarea.parquet

Agent Delegability Index (ADI) v2026.1

35.2 % of task time in world employment is at least partly delegable to AI agents (L2 or L3) in 2026; only 2.5 % can be handed over in full (L3). Delegable employment rises with income — 22.5 % in low-income, 28.3 % in lower-middle, 35.0 % in upper-middle and 48.7 % in high-income countries — the delegation divide (7.23 percentage points per log point of GDP per capita, 146 countries).

Dataset of the paper The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries (Espinal Maya & Jiménez Londoño, 2026). Preregistered at OSF: doi:10.17605/OSF.IO/9ZNTF. Explore every occupation and country, with charts and citations: https://balamlabs.org/adi (Balam Labs). Paper: arXiv:2610.30752. Dataset DOI: 10.5281/zenodo.23552918. Code, rubric, MCP server and full pipeline: https://github.com/Cespial/delegation-divide.

Configs

config	rows	unit
`adi_isco` (default)	433	ISCO-08 unit group (4 digits)
`adi_pais`	146	country (ISO 3166-1 alpha-3)
`adi_tarea`	18,796	O*NET 30.2 task statement, with each model's level

from datasets import load_dataset
occ = load_dataset("cespial/adi", "adi_isco", split="train")
countries = load_dataset("cespial/adi", "adi_pais", split="train")
tasks = load_dataset("cespial/adi", "adi_tarea", split="train")

The repository also carries the same tables as CSV, claims.jsonl (every number in the paper), the raw model outputs (ratings_sonnet.jsonl.gz, ratings_haiku.jsonl.gz), a static JSON API (api/isco/<code>.json, api/country/<ISO3>.json, api/index.json), DATA_DICTIONARY.md and SHA256SUMS.

Rating scale (verbatim from the frozen rubric)

L0 — No meaningful role. The agent cannot perform or materially speed up any substantial part of the task.
L1 — Augmentation. A human performs the task; the AI assists step by step (drafts, searches, suggestions). The human stays in the loop at each step and produces the output.
L2 — Partial delegation. An agent can execute substantial sub-steps of the task end to end from a goal specification. A human still has to perform at least one essential step or review and correct the work before use.
L3 — Full delegation. An agent can complete the whole task end to end, to the quality a competent worker would accept, given only the goal and access. The human's role is limited to assigning it and accepting the result.

Reference technology: publicly available AI agents as of October 2026 (tool use, code execution, browsing, computer use, file handling; no physical body). Ratings measure technical feasibility, not adoption or legality. Each task also carries the share of working time an agent could take over, the minutes a human needs, and four barrier flags (physical, accountability, trust, access).

Main columns

adi_isco: isco08, titulo, adi_l2 (share 0–1 of task weight at L2+), adi_l3 (share at L3), adi_share (% of working time delegable), n_tareas, barrera_{fisica,responsabilidad,confianza,acceso}, ic_lo/ic_hi (95 % bootstrap interval of adi_share).
adi_pais: iso3, pais, ingreso (World Bank FY2026), anio, resolucion (ISCO level of the employment data), adi_l2, adi_l3, adi_share (employment-weighted), empleo_total.
adi_tarea: task_id (O*NET), titulo, texto, consensus nivel, share, minutos, barrier fractions, nivel_sonnet, nivel_haiku.

Types, units and definitions for every column: DATA_DICTIONARY.md.

How it was built

Raters: Claude Sonnet 5.5 (primary) and Claude Haiku 5.5 (secondary), one task per call, neutral configuration, rubric hash checked before every run. Consensus: lower level when the models disagree (quadratic-weighted κ = 0.77 between them). Tasks are weighted by O*NET importance × frequency and carried to ISCO-08 through O*NET-SOC 2019 → SOC 2010 → ISCO-08 (433 of 436 unit groups). Country scores weight unit groups by ILOSTAT employment (latest year ≥ 2019, 2-digit or 1-digit ISCO) with within-group four-digit weights from BLS OES (high income) or Colombia's GEIH (other groups).

Limitations

Ratings come from two models of one developer; a blind human rating of 300 tasks agrees with the model consensus at kappa = 0.85, and an agentic panel on 1,500 tasks at kappa = 0.87 (see the paper, section 7).
O*NET describes work in the United States; the same task mix is carried to every country, so cross-country differences are compositional.
Technical delegability is an upper bound on delegation, not a forecast of employment effects.

Licence and attribution

CC BY 4.0. Includes information from the O*NET 30.2 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license (task statements are reproduced verbatim in adi_tarea). ESCO, ILOSTAT, World Bank, BLS and DANE inputs are used only as derived aggregates.

Citation

@misc{espinal2026delegation,
  title         = {The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries},
  author        = {Espinal Maya, Cristian and Jim{\'e}nez Londo{\~n}o, Santiago},
  year          = {2026},
  eprint        = {2610.30752},
  archivePrefix = {arXiv},
  primaryClass  = {econ.GN},
  doi           = {10.48550/arXiv.2610.30752},
  url           = {https://arxiv.org/abs/2610.30752}
}