# Balam Labs

> Balam Labs (blabs) publishes independent, open, preregistered indices of what AI agents can do. Every rating, score and line of code is published.

## Products

### Agent Delegability Index (ADI): https://balamlabs.org/adi

> Open, preregistered index of how much work AI agents can take over, for 433 ISCO-08 occupations and 146 countries (v2026.1).
> Headline: 35.2% of task time in world employment is at least partly delegable to AI agents (levels L2–L3) in 2026
> (95% interval 33.9–36.7%); only 2.5% can be handed over in full (L3).
> 48.7% in high-income vs 22.5% in low-income countries (the "delegation divide": +7.23 pp per log point of GDP per capita).
> Four levels per task: L0 no role, L1 augmentation, L2 partial delegation, L3 full delegation. Four barriers: physical presence, accountability, trust and care, access to systems.
> Rated by Claude Sonnet 5.5 and Claude Haiku 5.5 on 18,796 O*NET 30.2 tasks; a blind human rating of 300 tasks, led by a coauthor with student raters, agrees at kappa 0.85.
> Data DOI: https://doi.org/10.5281/zenodo.23552918 · Preregistration: https://doi.org/10.17605/OSF.IO/9ZNTF

#### Pages
- [Overview](https://balamlabs.org/adi): headline, leaderboard by ISCO-08 major group, search
- [Occupations](https://balamlabs.org/adi/occupations): all 433 occupations ranked; one page each at https://balamlabs.org/adi/occupations/<ISCO-08 code>
- [Countries](https://balamlabs.org/adi/countries): explorer (ADI against GDP per capita or AI use, map, distribution by region or income; state in the URL: ?view=divide|map|distribution&metric=l2|l3|s|fisica|uso|residuo&region=&income=&q=) and ranking; one page each at https://balamlabs.org/adi/countries/<ISO3> with position relative to income, decomposition by ISCO-08 major group, barriers and peers (site calculations, not in the API)
- [Methodology](https://balamlabs.org/adi/methodology): levels, barriers, rubric, aggregation, validation, deviations from the preregistration
- [Report](https://balamlabs.org/adi/report): The Delegation Divide, ADI Report 2026.1
- [Data](https://balamlabs.org/adi/data): downloads, API, Python package, MCP server, citation

#### API (static JSON, no key)
- https://balamlabs.org/adi/v2026.1/index.json
- https://balamlabs.org/adi/v2026.1/isco/<ISCO-08>.json (e.g. https://balamlabs.org/adi/v2026.1/isco/2512.json)
- https://balamlabs.org/adi/v2026.1/country/<ISO3>.json (e.g. https://balamlabs.org/adi/v2026.1/country/COL.json)
Every record carries a "cite" field.

#### Tables
- https://balamlabs.org/data/adi_isco.csv, https://balamlabs.org/data/adi_pais.csv, https://balamlabs.org/data/adi_tarea.csv (and .parquet); columns in https://balamlabs.org/data/DATA_DICTIONARY.md
- https://balamlabs.org/data/claims.jsonl: every number in the paper, the report and this site

## Lab
- [About](https://balamlabs.org/about): people, releases, use of AI, licences, corrections

## Optional
- [llms-full.txt](https://balamlabs.org/llms-full.txt): definitions, column schema, tool reference