In plain words
The paper introduces the Agent Delegability Index, which estimates how much of work AI agents could take over. Two language models rate written task descriptions, from “can only help” to “could finish the task,” and note barriers such as needing a person physically present. The authors combine those ratings with employment data to compare occupations and countries.
For the countries covered, the index estimates that 35.2% of task time could be handed over at least partly, but only 2.5% fully. These are estimates of what agents could do under a dated technology description, not records of what workplaces actually delegate.
The research
How much work could AI agents take over, rather than merely help with, and can that question be compared across countries? Existing measures often track exposure or current use, while task-level measures commonly rely on US occupation categories.
The index addresses both gaps with task ratings mapped to an international occupation classification.
The Agent Delegability Index rates tasks for partial or full handover to AI agents, then weights those ratings across occupations and countries.
The Agent Delegability Index (ADI) asks how much occupational work AI agents could take over, a question different from whether AI can speed up a task or is already used. It rates tasks for levels of handover and barriers, then combines the ratings with occupational and country employment data.
Across 146 countries, the employment-weighted estimate is 35.2% of task time at least partly delegable and 2.5% fully delegable.
The share of employment in delegable tasks also rises with national income; the authors interpret this pattern as reflecting differences in occupational mix, not different feasibility for the same task.
These are estimates of technical feasibility for publicly available AI agents as of October 2026, not measures or forecasts of actual adoption, job loss, or employment effects.
The index carries US task descriptions and weights to other countries, where the same occupations may involve different tasks and tools.
Analysis
Deep
- Original paper authors
- Cristian Espinal Maya, Santiago Jiménez Londoño
- Submitted to arXiv
- License
- CC BY 4.0