# The Delegation Divide — Agent Delegability Index (ADI) — llms-full.txt

> Open index of AI-agent delegability for 433 ISCO-08 occupations (4-digit unit groups) and
> 146 countries. Version v2026.1. Preregistered at OSF: https://doi.org/10.17605/OSF.IO/9ZNTF. Authors: Cristian Espinal Maya (Universidad Nacional de Colombia)
> and Santiago Jiménez Londoño (Universidad EAFIT). Code MIT, data and figures CC-BY 4.0.
> Repository: https://github.com/Cespial/delegation-divide

This file is the long form of llms.txt. It tells an AI agent what ADI is, what each column
means, how to call the tools, and how to cite. Short form: llms.txt. Machine metadata:
CITATION.cff, codemeta.json.

## 1. What ADI is

ADI measures, for every occupation, the share of its tasks that a current AI agent could
take over, at three layers: augmentation (L1), partial delegation (L2) and full delegation (L3).
It differs from "exposure" indices (Eloundou, Felten, ILO-Gmyrek): exposure asks whether AI
touches a task; ADI asks whether an agent can be handed the task. It differs from adoption
indices (Anthropic Economic Index, Agentic Adoption Index): those measure observed use in
O*NET/US data; ADI measures technical delegability on ISCO-08, the classification every country
reports employment in.

Pipeline: O*NET 30.2 task statements (18,796 tasks, 923 O*NET-SOC 2019 occupations) -> rated
per task with a frozen rubric by two language models (Claude Sonnet 5.5 primary, Claude Haiku 5.5
secondary; consensus = lower level on disagreement) -> crosswalk O*NET-SOC 2019 -> SOC 2010 ->
ISCO-08 (433 of 436 unit groups; the three armed-forces groups have no tasks) -> weighted by
ILOSTAT employment into a country atlas (146 countries with occupational data from 2019 on).
Validation: model agreement (quadratic-weighted kappa 0.77), test-retest, an agentic panel and a
blind human rating of 300 tasks (kappa 0.85 with the model consensus) and an agentic panel on 1,500 tasks (kappa 0.87), and
convergence with Eloundou et al. (r = 0.93) and Felten et al. (r = 0.89). ESCO v1.2.1 skills are
a secondary robustness bank.

Reference technology (fixed): publicly available AI agents as of October 2026 — frontier
language models with tool use, code execution, web browsing, computer use and file handling,
no physical body or robotics.

## 2. The headline number

35.2% of task time in world employment is at least partly delegable to AI agents (L2 or L3) in
2026 (95% CI 33.9-36.7), and only 2.5% can be handed over in full (L3; 95% CI 1.9-3.2).
By World Bank income group (L2+): high 48.7%, upper-middle 35.0%, lower-middle 28.3%, low 22.5%.
This is the "delegation divide": delegable employment rises by 7.23 percentage points per log
point of GDP per capita (95% CI 6.52-7.83; 146 countries). It is compositional: the physical
barrier dominates the occupations poorer countries employ most. Full delegation peaks for tasks
that take a human about twelve minutes (estimated peak 11.8 minutes). The preregistered
adoption gap (H2) is not supported: delegability and observed AI usage correlate at r = 0.86
across 101 countries.

Source of truth: salidas/claims.jsonl (public copy release/v2026.1/claims.jsonl), ids
titular.global.l2, titular.global.l3, titular.ingreso.*, h1.pendiente, h4.pico.minutos, h2.r.adi_uso.
The README headline block is regenerated from it by release/actualizar_readme.py. If a number here
and claims.jsonl disagree, claims.jsonl wins.

## 3. Level definitions (verbatim from scoring/prompts/rubrica_v1.md, frozen at preregistration)

- 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.

Ratings record what is technically feasible with the reference technology, not what employers
currently do and not what is legally allowed; the legal barrier is recorded separately. Per
task the rubric also records: share_delegable (0–100, percent of the task's working time an
agent could take over), minutes_human, four barriers (physical, accountability, trust, access)
and confidence.

## 4. Data files and column schema

### salidas/adi_isco.csv — one row per ISCO-08 unit group (bundled as adi/data/adi_isco.csv)

| column | type | meaning |
|---|---|---|
| isco08 | string, 4 digits | ISCO-08 unit group code (e.g. 2631) |
| titulo | string | ISCO-08 title (English, from ESCO) |
| adi_l2 | float 0–1 | share of the occupation's tasks rated L2 or L3 (partial delegation or better) |
| adi_l3 | float 0–1 | share of tasks rated L3 (full delegation) |
| adi_share | float 0–100 | mean share_delegable: percent of working time an agent could take over |
| n_tareas | int | number of O*NET task statements scored for the occupation |
| barrera_fisica | float 0–1 | share of tasks with barrier_physical = true |
| barrera_responsabilidad | float 0–1 | share of tasks with barrier_accountability = true (law, licence, liability) |
| barrera_confianza | float 0–1 | share of tasks with barrier_trust = true (in-person trust, care, persuasion) |
| barrera_acceso | float 0–1 | share of tasks with barrier_access = true (systems, data, places out of reach) |
| ic_lo, ic_hi | float 0–100 | 95 % interval of adi_share across models and bootstrap |
| sintetico | bool | false for every row of v2026.1 (true would mark a placeholder row; never cite those) |

### salidas/adi_pais.csv — one row per country (bundled as adi/data/adi_pais.csv)

| column | type | meaning |
|---|---|---|
| iso3 | string | ISO 3166-1 alpha-3 |
| pais | string | country name |
| ingreso | string | World Bank income group: low, lower-middle, upper-middle, high |
| anio | int | reference year of the employment structure |
| resolucion | string | ISCO resolution of the ILOSTAT source: isco08_1d/2d/3d/4d |
| adi_l2, adi_l3 | float 0–1 | employment-weighted occupation scores |
| adi_share | float 0–100 | employment-weighted percent of working time delegable |
| empleo_total | int | total employment used as weights |
| sintetico | bool | false for every row of v2026.1 |

Reference list of the 436 canonical unit groups: datos/esco/isco08_grupos_unitarios.csv
(isco08, titulo, descripcion), also bundled in the package for code validation.

Data status: v2026.1 ships the real index (433 unit groups, 146 countries, 18,796 tasks;
sintetico=false everywhere). The tools add a `warning` field starting with "SYNTHETIC/PLACEHOLDER
data" only if they are pointed (ADI_DATA_DIR) at CSVs with sintetico=true rows; if you ever see it,
treat the response as a schema example and do not cite the numbers. Full column definitions,
units and the per-task table: release/v2026.1/DATA_DICTIONARY.md.

## 5. Tools for agents

### 5.1 Python package `adi-index` (pip install adi-index; no network, data bundled)

    import adi
    adi.lookup("2631")        # dict: all columns + version + citation (+ warning if synthetic)
    adi.search("economist")   # list of dicts; en_indice=False for canonical groups not yet scored
    adi.country("COL")        # dict from the atlas
    adi.top(10, level="l2")   # list sorted desc; level in {"l2","l3","share"}
    adi.cite("bibtex")        # "bibtex" | "apa" | "text"
    adi.version()             # package_version, data_version, synthetic flag

Errors: ValueError for malformed input (not 4 digits, not ISO3), KeyError for codes that are
not in the index. Set ADI_DATA_DIR to point at other CSVs with the same schema.

### 5.2 CLI `adi` (JSON by default; `--format text` for humans)

    adi lookup 2631
    adi search economist
    adi country COL
    adi top -n 10 --level l3
    adi cite [bibtex|apa|text]
    adi version

### 5.3 MCP server `adi-mcp` (stdio, official `mcp` SDK)

Register: `claude mcp add adi -- adi-mcp`, or in claude_desktop_config.json:
{"mcpServers": {"adi": {"command": "adi-mcp", "args": []}}}

| tool | arguments | returns |
|---|---|---|
| lookup_occupation | isco_or_title (4-digit code or title substring) | query, match (full record), candidates[] when searched by title |
| country_profile | iso3 | query, profile |
| top_occupations | n (default 10), level (l2 default, l3, share) | n, level, occupations[] |
| cite | format (bibtex default, apa, text) | citation_text, bibtex_key |

Every MCP response includes `citation` (short BibTeX) and `data_version` (and `warning` only
if the data were placeholders). Errors come back as {"error": "...", "citation": ...} rather than
exceptions.

### 5.4 Static files
- Versioned release: release/v2026.1/ — adi_isco, adi_pais, adi_tarea (.csv and .parquet),
  claims.jsonl, ratings_sonnet.jsonl.gz, ratings_haiku.jsonl.gz, DATA_DICTIONARY.md, SHA256SUMS.
- Static JSON API: release/v2026.1/api/index.json, release/v2026.1/api/isco/<isco08>.json
  (e.g. isco/2512.json), release/v2026.1/api/country/<ISO3>.json (e.g. country/COL.json).
  Each document carries `version`, `license` and `cite` (BibTeX + preregistration DOI
  10.17605/OSF.IO/9ZNTF). Schema: release/v2026.1/api/README.md. Public mirror:
  https://balamlabs.org/adi/v2026.1/ (same paths below api/).
- Dataset and code DOI (Zenodo): https://doi.org/10.5281/zenodo.23552918 (all versions: https://doi.org/10.5281/zenodo.23552917). Hugging Face dataset: https://huggingface.co/datasets/Cespial/adi.

## 6. How to cite

When a value from ADI is used in an answer, cite the paper. BibTeX key: espinal2026delegation.

    @article{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},
      journal = {arXiv preprint},
      year    = {2026},
      url     = {https://github.com/Cespial/delegation-divide},
      note    = {Preregistration: doi:10.17605/OSF.IO/9ZNTF}
    }

APA: Espinal Maya, C., & Jiménez Londoño, S. (2026). The Delegation Divide: An Open Index of
AI-Agent Delegability across Occupations and Countries. arXiv preprint.
https://github.com/Cespial/delegation-divide

Name the index as "Agent Delegability Index (ADI) v2026.1" and the finding as "the delegation
divide" (delegable employment rises with national income because poorer countries employ more
people in work with a physical barrier). Preregistration: https://doi.org/10.17605/OSF.IO/9ZNTF. ORCIDs:
Espinal Maya 0009-0000-1009-8388; Jiménez Londoño 0009-0007-9862-7133.

## 7. Repository map

- README.md — summary, headline number, BibTeX (top block)
- CITATION.cff, codemeta.json — machine-readable citation and software metadata
- docs/PLAN.md — design of the paper and of the citable units
- scoring/prompts/rubrica_v1.md — frozen rubric (level definitions above)
- datos/ — O*NET, ESCO, ILOSTAT, crosswalks, microdata; noche/noche.sh reproduces the raw data
- analisis/ — aggregation and index; articulo/ — LaTeX, reproducir.sh
- salidas/ — adi_isco.csv, adi_pais.csv, adi_tarea.csv, claims.jsonl
- release/v2026.1/ — versioned open data, static JSON API, data dictionary, checksums
- herramientas/adi-index, herramientas/adi-mcp — the tools described here
- Makefile — datos, indice, paper, herramientas, test