The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries
Most studies measure how much AI can speed up a task—exposure—rather than whether a worker can actually hand it off. This paper introduces the Agent Delegability Index (ADI), which rates all 18,796 O*NET tasks on a frozen, preregistered four-level delegation ladder, then maps them to ISCO-08 occupations and weights them by employment across 146 countries. By distinguishing partial from full delegation and tagging physical, accountability, trust, and access barriers, the authors reveal a steep income gradient in who can delegate: high-income countries have roughly twice the delegable employment share of low-income ones, driven mainly by the composition of work.
The paper is a methodologically ambitious and unusually transparent attempt to shift the AI-and-work literature from exposure to delegation, with open data, preregistered hypotheses, and an MCP server setting a high bar. However, two of its five preregistered hypotheses fail or invert: occupation-level delegability converges with existing exposure indices far more tightly than anticipated ($r=0.93$ with Eloundou et al.), and the predicted adoption gap in middle-income countries is not found ($p=0.270$). Moreover, the human validation fell short of the registered standard, leaving the model ratings with only a coauthor-led blind check on 300 tasks rather than independent raters on 1,500.
The preregistration and transparency are exemplary: a frozen rubric locked by SHA-256 hash, bootstrap inference, and full release of model outputs and crosswalk code. The conceptual framework distinguishing four levels of delegation and four named barriers is clear and policy-relevant, and the global employment weighting via ILOSTAT natively in ISCO-08 fills a real geographic gap. Model stability is high—Sonnet repeats at quadratic-weighted $\kappa = 0.97$ on a 100-task subsample—and the blind human rating agrees with the model consensus at $\kappa = 0.85$, comfortably above the preregistered floor of $0.60$.
The strongest concern is that ADI does not produce a new occupational ranking. The preregistration set an upper bound of $r = 0.80$ against exposure indices, explicitly warning that correlations above it would mean ADI 'adds little beyond exposure'; the actual correlations are $0.93$ with Eloundou et al. and $0.89$ with Felten et al. The authors report this honestly, but it means the index's novel contribution is restricted to the level scale, barrier flags, and geography—not to reallocating risk across occupations.
A second concern is the validation downgrade: the registered plan of three independent human raters on 1,500 tasks was replaced by a single coauthor-led sheet on 300 tasks, preventing any measure of inter-rater reliability and weakening claims of bias control. Third, the usage test for the adoption gap relies on Anthropic's own Economic Index, creating circularity risk since the same model family rates delegability and generates the usage data; the null result is therefore difficult to interpret without independent adoption metrics.
The evidence supports the headline country-level gradient and the rarity of full delegation ($2.5\%$ of world task time), and the ESCO robustness check and permutation placebo strengthen occupational-level claims. However, the high convergence with exposure indices suggests the ratings are largely consistent with prior work rather than revealing a distinct task ordering. Table 1 offers a fair comparison, correctly noting that ADI is the only measure natively in ISCO-08 and rated for agents. The logistic relationship between full delegation and estimated task length—peaking near $11.8$ minutes—is an intriguing regularity, though it rests on model-estimated minutes rather than measured durations. Against observed usage, ADI correlates lower ($r = 0.69$) than against exposure, supporting the authors' claim that capability and adoption are not identical rankings, even if the usage source is non-independent.
Reproducibility is structurally strong but practically constrained. All prompts, raw outputs, code, and a Python package are released, and the rubric is cryptographically fixed. Yet independent rerating is blocked because the pipeline depends on proprietary Claude models (Sonnet and Haiku 5.5) accessed through a specific CLI version that exposes no temperature control, and the neutral system prompt was essential to avoid an inflationary framing effect. Employment microdata for within-group weights are borrowed from the United States or Colombia for most countries, documenting a known fidelity limit rather than solving it. The released MCP server and JSON endpoints make the index easy to query, but rerunning the full pipeline from scratch requires API access to exact model snapshots that may drift or deprecate.
35.2% of task time in world employment is at least partly delegable to AI agents in 2026, and only 2.5% can be handed over in full. Existing measures score exposure to generative AI, mostly for the United States. We build the Agent Delegability Index (ADI): two language models rate all 18,796 O*NET 30.2 task statements on a frozen, preregistered rubric of delegation levels and barriers; tasks are mapped to 433 ISCO-08 unit groups and weighted by ILOSTAT employment in 146 countries. Four findings. First, delegable employment rises with income, from 22.5% in low-income to 48.7% in high-income countries, a slope of 7.23 percentage points per log point of GDP per capita. This delegation divide is compositional: the physical barrier dominates the jobs poorer countries employ most. Second, at the occupation level ADI tracks exposure closely (r = 0.93 with Eloundou et al. and 0.89 with Felten et al.), above the preregistered bound of 0.80; ADI adds delegation levels and barriers on top of exposure, for the world, not a new ranking of occupations. Third, full delegation is rare and peaks for tasks that take a human about twelve minutes (11.8 minutes). Fourth, the preregistered adoption gap is not supported: across 101 countries, delegability and observed AI usage correlate at r = 0.86, and the gap between them is not wider in middle-income countries (p = 0.27). Model agreement is kappa = 0.77; a blind human rating of 300 tasks agrees with the model consensus at kappa = 0.85, and an agentic panel with web search at kappa = 0.87. We release the index, the atlas, the prompts and an MCP server at https://github.com/Cespial/delegation-divide.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.