Source-linked AI summary
Who Delegates to AI? Evidence from 53,000 Agent Configurations
Hyeongjae Lee, Jihyang Cheon, Lanu Kim
TL;DR
Existing AI-exposure measures describe what AI could do or what workers use, but not whether workers have committed tasks to AI workflows. The paper constructs the Agentic Adoption Index by matching practitioner-built agent skills to O*NET tasks, finding that delegated exposure differs from pre-AI risk rankings, tracks capability more than conversational use, and peaks below the highest wage and education levels.
Problem
Existing exposure measures capture AI capability or use but do not directly show whether workers have committed tasks to AI through workflows.
Method
The paper constructs the Agentic Adoption Index by matching roughly 53,000 practitioner-built agent skill specifications with about 18,000 O*NET task statements and aggregating similarities by occupation.
Results
Delegated exposure differs from pre-AI automation-risk rankings, aligns more closely with technical capability than conversational use, and follows an inverted-U across wages while peaking at the bachelor’s level.
Takeaways & Limitations
Technical feasibility is necessary but insufficient for agentic adoption, because the highest-paid and most-educated occupations adopt agents least despite high technical exposure.
Takeaways & Limitations
The data cannot distinguish whether lower adoption in top occupations reflects difficulty specifying work in advance or professionals’ choice to delay or decline delegation.
Abstract
from arXiv · showhide
A growing literature measures how far occupations are exposed to AI, but these measures capture where AI could perform tasks, not whether workers have adopted it. We propose a new layer of exposure, delegated exposure, which records whether a worker has committed a task to AI by building it into a workflow. We operationalize it as the Agentic Adoption Index (AAI), which measures how closely an occupation's tasks match the agentic routines practitioners have already built and shared. We embed roughly 53,000 agent skill specifications from the Manus Skills Marketplace, compute their semantic similarity to about 18,000 O*NET task statements, and aggregate to the occupation level. Three findings follow. First, the occupations where delegation concentrates differ sharply from those pre-AI frameworks identified as most at risk. Second, the AAI tracks what AI could do more closely than what workers currently use it for. Third, the AAI peaks below the top of the wage distribution and at the bachelor's level, declining at both extremes. Technical availability explains most of this variation, but not the shortfall among the most educated occupations, so feasibility alone cannot account for who adopts. That shortfall may reflect work that resists advance specification, or professional discretion over the pace of codification. Distinguishing the two, and tracking how these measures diverge over time, will require repeated measurement.
1 Introduction
The paper introduces delegated exposure to measure whether workers have committed tasks to AI through agentic workflows, addressing limits in capability, availability, and observed-use measures. Using the Agentic Adoption Index, it finds that adoption patterns differ from earlier exposure frameworks, align more with technical capability than conversational use, and peak below the highest-wage and most-educated occupations.
- Motivation and contribution: Delegated exposure records whether workers have committed tasks to AI by building it into agentic workflows.The paper distinguishes deliberate task delegation from merely using AI or exposing work to technically capable systems.
- Method: 53,000 agent skill specifications are matched semantically to 18,000 O*NET task statements to construct the Agentic Adoption Index.The index aggregates task-level similarities to the occupation level, weighting tasks by their importance.
- Findings: Agent adoption concentrates in occupations that differ sharply from those earlier automation research identified as most exposed.The finding challenges the assumption that current agentic adoption follows pre-AI automation-risk rankings.
- Findings: The AAI aligns more closely with technical capability than with observed conversational LLM use.Building an agent requires deliberate configuration, whereas conversational use can spread without regard to whether AI can reliably perform the work.
- Findings: Adoption follows an inverted-U across wages and peaks at the bachelor’s level, declining among the highest-paid and most-educated occupations.Technical feasibility explains much of the variation, but it does not explain the shortfall at the top of the wage and education distributions.
- Contribution: The AAI measures revealed practitioner choices and can be recomputed as the continuously growing repository of agent configurations changes.This distinguishes the measure from exposure estimates anchored to periodic expert assessments.
2 Background and Related Work
Prior AI-exposure research distinguishes technical capability, tool availability, and observed use, but none directly shows whether workers have delegated tasks into workflows. The paper adds delegated exposure by using practitioner-configured agents, which provide more direct evidence of intended delegation than conversational logs.
- Exposure measurement: Occupational exposure varies across tasks within a job, so O*NET’s task-level structure helps identify which parts of work are affected.Treating occupations as indivisible units can obscure substantial within-occupation differences in task exposure.
- Four layers of AI exposure: Capability, availability, and observed exposure form three approximately nested layers covering what AI can do, what tools make usable, and what workers use.The layers differ by how closely they approach realized adoption.
- Capability exposure: Capability exposure measures which tasks AI could technically perform, often identifying language-heavy and cognitively demanding occupations.Results depend on the technology assessed; reinforcement-learning measures can highlight monitoring and control occupations overlooked by language-model measures.
- Availability exposure: Availability exposure narrows capability to what software scaffolding or production-ready tools make accessible to workers.A technically capable model does not become usable until incorporated into an interface or software tool.
- Observed exposure: Observed exposure captures actual LLM use, but usage logs may cover fewer activities than capability estimates and cluster in narrow task ranges.Conversational sessions do not map neatly onto tasks, and prompts reveal little about whether workers rely on AI or merely experiment.
- Delegated exposure: Configured agents provide clearer evidence of delegation because users specify goals, procedures, and required tools for repeated execution.Agents shift practitioners from performing each step to supervising a sequence of work, motivating the AAI’s linkage to O*NET tasks.
3 Data
The data combine practitioner-authored agent skill files with standardized O*NET occupational task statements to measure how agentic workflows correspond to work. The construction emphasizes reusable, task-like descriptions and task importance when aggregating matches to occupations.
- Agent skill data: Practitioner-authored skill files are intended to provide direct evidence of users’ configuration choices and describe agent tasks at an occupationally comparable level.The data source is selected to capture what practitioners intend agents to do, rather than platform usage volume alone.
- Agent skill data: The Marketplace contains reusable skill.md resources whose descriptions state what a skill does and when it should be invoked.These descriptions are treated as a natural semantic-comparison unit because they operate at a similar abstraction level to O*NET task statements.
- Data processing: 117,887 distinct skills are reduced through deduplication and filters that remove entries lacking execution instructions or sufficient descriptive context.The remaining skills are then filtered for work-relatedness before constructing the analysis sample.
- Occupational task data: O*NET 30.2 supplies standardized occupation-specific descriptions of the tasks, knowledge, and skills required across occupations.The database is maintained by the U.S. Department of Labor and supports task-level occupational analysis.
- Occupational task data: Task importance scores on a five-point scale weight task-level exposure when scores are aggregated to occupations.After excluding tasks without valid importance ratings, 17,951 of 18,797 O*NET task statements remain.
- Comparison measures: The analysis compares delegated exposure with capability, availability, observed-use, and pre-AI computerization-risk measures.These comparisons test whether agentic diffusion follows earlier routine-task patterns or differs from existing AI-exposure measures.
4 Methods
The paper constructs and validates the Agentic Adoption Index by matching agent skill descriptions to O*NET tasks, aggregating similarities by occupation, and comparing the index with existing exposure measures and occupational characteristics.
- Constructing the AAI: Occupation-level AAI scores average task-level similarity across skills, then use importance-weighted averaging across each occupation’s tasks.Task importance weights are normalized so occupations with more tasks are not scored higher merely because they contain more tasks.
- Validating the AAI: A different embedding model produces occupation-level AAI indices with a Pearson correlation of 0.91, indicating robustness to model choice.The robustness check replaces all-MiniLM-L6-v2 with all-mpnet-base-v2.
- Validating the AAI: High-AAI occupations are information-intensive, while low-AAI occupations more often require manual dexterity or direct physical intervention.The AAI is positively correlated with the share of cognitive abilities used in occupations.
- Comparative analyses: The analysis compares AAI with pre-AI computerisation risk, capability, availability, and observed-use measures to assess what the index captures.It also examines employment-weighted relationships between AAI, log wages, squared log wages, and education categories relative to bachelor’s degree.
- Comparative analyses: The study controls for differences in occupation coverage because the exposure sources use different O*NET releases and levels of measurement.These differences produce slightly varying occupation counts across measures.
5 Results
The AAI diverges from pre-AI automation risk, aligns more closely with capability and availability than observed use, and follows an inverted-U across wages while peaking at the bachelor's level. Availability explains much of these patterns, but not the persistent adoption gap among the most-educated occupations.
- 5.1 Agentic adoption does not follow pre-AI automation risk: 0.05 Jaccard similarity indicates that high-AAI occupations are largely distinct from those classified as high-risk by Frey–Osborne automation risk.The weak overlap also holds at sector and state levels.
- 5.2 The AAI aligns with capability more than with observed use: ρ = 0.673*** for capability and ρ = 0.623*** for availability exceed ρ = 0.374*** for observed usage in correlations with AAI.The AAI aligns more closely with technical capability and availability than with where AI is already deployed in practice.
- 5.3 Availability explains adoption, except at the top: AAI follows an inverted-U across wages, with β = 0.447*** for the linear term and β = −0.019*** for the quadratic term.The index peaks below the top of the wage distribution and declines at both extremes.
- 5.3 Availability explains adoption, except at the top: AAI is highest at the bachelor's level: high-school-or-below occupations have β = −0.011*** and master's-or-above occupations β = −0.026*** relative to bachelor's occupations.The master's-or-above gap remains significant after controlling for availability, while the lower-education gap loses significance.
- 5.3 Availability explains adoption, except at the top: R2 = 0.578 for availability alone exceeds R2 = 0.349 for wages and education together, while adding socioeconomic variables raises fit only modestly after availability is included.Availability explains the lower AAI of less-educated occupations, but not the persistent shortfall among the most-educated occupations.
6 Discussion and Conclusion
The AAI adds delegated exposure as a distinct layer of occupational AI exposure, capturing where practitioners are configuring AI to automate their own work. Its patterns resemble technical capability but diverge at the highest wage and education levels, where adoption falls below availability-based predictions.
- Delegated exposure: Delegated exposure captures where practitioners have begun configuring AI systems to automate their own work, rather than where automation has already occurred.The AAI matches agent skill descriptions to O*NET task statements and measures active attempts at delegation.
- Occupational pattern: AAI patterns differ from pre-AI computerization-risk estimates and concentrate in information-intensive work rather than tasks requiring manual dexterity or physical intervention.The measure also aligns more closely with technical capability and application availability than with prompt-level conversational use.
- Adoption mechanism: Technical availability accounts for most variation in agent adoption, consistent with the short distance between recognizing automatable work and configuring natural-language agents.The observed skill files were created and published by individuals without the procurement, integration, or specialized expertise usually associated with workplace technologies.
- Adoption gradient: AAI adoption follows an inverted-U across wages, peaks at the bachelor’s level, and declines among the highest-paid and most-educated occupations.Availability explains lower adoption among less-educated occupations but not the shortfall at the top, which may reflect either specification constraints or professional choice.
- Future research: Longitudinal evidence could distinguish whether the adoption shortfall among top occupations reflects improving feasibility constraints or persistent professional choice.A narrowing shortfall as agentic systems improve would favor constraint; persistence alongside rising capability would strengthen the case for choice.
- Future research: Publicly shared skill files offer a renewable basis for tracking diffusion without platform cooperation, though their availability depends on norms of open sharing.Conversational data are concentrated among a small number of platform firms, while expert task annotation is costly and slow to update.