PaperlayerView exact revision on arXiv

Source-linked AI summary

Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce

Yijia Shao, Humishka Zope, Yucheng Jiang, Jiaxin Pei, David Nguyen, Erik Brynjolfsson, Diyi Yang

arXiv:2506.06576v3cs.CYcs.AIcs.CLcs.HCcs.LG

TL;DR

The paper addresses the lack of systematic, worker-centered evidence about which occupational tasks AI agents should automate or augment and how those preferences align with technical capability. It introduces a task-level audit and WORKBank, finding positive automation preferences for many repetitive tasks, substantial desire-capability mismatches, diverse human-agency expectations, and shifting skill emphasis.

  • Problem

    Research lacks a systematic, forward-looking account of AI-agent automation and augmentation preferences across diverse occupational tasks and worker values.

    Source: §1 · p. 2

  • Method

    The study combines O*NET task auditing, audio-enhanced worker surveys, the H1-H5 Human Agency Scale, and AI-expert capability assessments in WORKBank.

    Source: §1 · p. 2

  • Results

    Workers express positive attitudes toward automation for 46.1% of tasks, while desire-capability comparisons reveal four task zones and 45.2% of occupations favor H3 equal partnership.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Aligning AI-agent development with worker desires requires attention to human involvement, capability mismatches, and emerging interpersonal and organizational skill needs.

    Source: §1 · pp. 2–3

    The audit is bounded by existing O*NET tasks, workers’ incomplete awareness of evolving AI capabilities, and a snapshot of agent systems as of early 2025.

    Source: §5 · pp. 16–17

Abstract

from arXiv · show

The rapid rise of compound AI systems (a.k.a., AI agents) is reshaping the labor market, raising concerns about job displacement, diminished human agency, and overreliance on automation. Yet, we lack a systematic understanding of the evolving landscape. In this paper, we address this gap by introducing a novel auditing framework to assess which occupational tasks workers want AI agents to automate or augment, and how those desires align with the current technological capabilities. Our framework features an audio-enhanced mini-interview to capture nuanced worker desires and introduces the Human Agency Scale (HAS) as a shared language to quantify the preferred level of human involvement. Using this framework, we construct the WORKBank database, building on the U.S. Department of Labor's O*NET database, to capture preferences from 1,500 domain workers and capability assessments from AI experts across over 844 tasks spanning 104 occupations. Jointly considering the desire and technological capability divides tasks in WORKBank into four zones: Automation "Green Light" Zone, Automation "Red Light" Zone, R&D Opportunity Zone, Low Priority Zone. This highlights critical mismatches and opportunities for AI agent development. Moving beyond a simple automate-or-not dichotomy, our results reveal diverse HAS profiles across occupations, reflecting heterogeneous expectations for human involvement. Moreover, our study offers early signals of how AI agent integration may reshape the core human competencies, shifting from information-focused skills to interpersonal ones. These findings underscore the importance of aligning AI agent development with human desires and preparing workers for evolving workplace dynamics.

1 Introduction

The paper introduces a worker-centered, task-level audit of AI-agent automation and augmentation, combining worker preferences with expert capability assessments. It identifies mismatches between desired human involvement, technical feasibility, and emerging workplace skill needs.

Motivation

  • Existing research lacks a systematic, forward-looking, worker-centered account spanning diverse occupational tasks and stakeholder values.Prior work often concentrates on limited domains, profitable tasks, or observed chatbot usage.Source: §1 · p. 2

Audit framework

  • The framework combines O*NET task coverage, audio-enhanced worker input, the H1-H5 Human Agency Scale, and AI-expert capability assessments.WORKBank contains responses from 1,500 workers across 104 occupations and annotations from 52 AI experts covering 844 tasks.Source: §1 · p. 2

Worker preferences

  • 46.1% of tasks receive positive worker attitudes toward AI-agent automation, especially for repetitive and low-value work, with freeing time for high-value work the primary motivation.Preferences vary significantly by sector and persist after workers reflect on potential job loss and work enjoyment.Source: §1 · p. 2

Desire-capability landscape

  • The desire-capability landscape divides tasks into Green Light, Red Light, R&D Opportunity, and Low Priority zones, exposing mismatches between worker demand and current capabilities.41.0% of Y Combinator company-task mappings fall in the Low Priority or Red Light zones, while promising tasks remain under-addressed.Source: §1 · pp. 2–3

Human agency

  • 45.2% of occupations have H3, equal partnership, as the dominant desired agency level, while workers generally prefer higher human involvement.These patterns suggest potential friction as AI capabilities advance.Source: §1 · p. 3

Changing skills

  • AI-agent integration is associated with a shift from information-processing skills toward interpersonal and organizational competence, alongside broader skill requirements.The paper presents these patterns as early signals of changing workplace competencies.Source: §1 · p. 3

2 Auditing Framework

The auditing framework evaluates computer-compatible occupational tasks through both worker preferences and expert assessments, spanning automation, augmentation, and human involvement. It operationalizes these perspectives in WORKBank using task-level sampling, the Human Agency Scale, audio-supported worker surveys, and dual assessments of desire and capability.

2.1 Defining Audit Granularity and Scope

  • The framework audits complex, multi-step, computer-compatible occupational tasks at the task level, enabling more contextual analysis than occupation-level assessment.Tasks are sourced from O*NET and represent recurring job responsibilities rather than isolated low-level activities.Source: §2.1 · p. 3

2.2 Emphasizing the Spectrum of Automation and Augmentation

  • HAS defines five human-involvement levels, from H1 autonomous execution to H5 essential continuous involvement, providing a shared language for automation versus augmentation.H1-H2 favor automation approaches, whereas H3-H5 support augmentation strategies; higher levels are not inherently better.Source: §2.2 · pp. 3–5, Figure 2 · p. 5

2.3 Constructing A Worker-Centric Auditing Framework

  • Worker-centered surveys collect automation desire and desired HAS ratings through audio-enhanced reflection, task-familiarity filtering, and guided consideration of relevant task factors.Workers rate each task’s automation desire and required collaboration on five-point Likert scales.Source: §2.3 · pp. 5–6

2.4 Ensuring Dual Perspectives from Both Domain Workers and AI Experts

  • The framework pairs worker desire ratings with AI-expert assessments of automation capability and feasible HAS, revealing alignments, gaps, and development priorities.Experts use the same rubrics while accounting for current systems’ strengths and limitations.Source: §2.4 · p. 6

2.5 Instantiating the Audit Framework to Derive WORKBank

  • WORKBank begins with 2,131 filtered tasks across 287 occupations, sourced from recurring computer-compatible O*NET workflows.The filtering process retains complex, multi-step tasks central to the audit’s focus.Source: §2.5 · p. 6
  • WORKBank contains ratings from 1,500 workers across 104 occupations and assessments from 52 AI experts, with sector comparisons suggesting broad workforce coverage.Experts independently assessed tasks, and the database combines complementary worker and capability data.Source: §2.5 · p. 7, Figure 3 · p. 7

3 Results

The results audit worker preferences, technical feasibility, human agency, and skill implications across occupational tasks. They reveal mismatches between desired and feasible automation, heterogeneous expectations for human involvement, and a potential shift toward interpersonal and organizational skills.

3.1 Worker-centered Views on Occupational Task Automation

  • 46.1% of tasks received positive automation-desire ratings, with freeing time for high-value work the most cited motivation at 69.38%.Preferences were mixed across tasks and sectors, while workers also reported concerns about trust, replacement, and loss of human qualities.Source: Figure 4 · p. 8, §3.1 · p. 9
  • The top 10 occupations by automation desire represented only 1.26% of Claude.ai usage, indicating that usage logs may miss latent workplace demand.This supports directly soliciting worker input rather than relying solely on observed chatbot use.Source: Figure 4 · p. 8, §3.1 · p. 9

3.2 Desire-Capability Landscape for AI Agents in the Workplace

  • The desire-capability landscape distinguishes Green Light, Red Light, R&D Opportunity, and Low Priority zones according to worker desire and expert-assessed capability.These zones separate deployment candidates, caution areas, research opportunities, and less urgent tasks.Source: §3.2 · pp. 9–10
  • Worker desire and expert capability were weakly aligned, with Spearman ρ=0.17, and 41.0% of YC company-task mappings fell in Low Priority or Automation “Red Light” zones.Research papers were more concentrated in the R&D Opportunity Zone but remained focused largely on computer science and engineering tasks.Source: §3.2 · pp. 10–11, Figure 5 · p. 10

3.3 Human Agency Scale (HAS) Spectrum

  • 47.5% of tasks fell below the diagonal comparing worker-desired and expert-assessed Human Agency Scale levels, while only 26.9% received matching levels.Disagreement was especially pronounced in lower-HAS occupations, including several occupations rated H1-dominant by experts.Source: §3.3 · p. 11
  • H3 was the dominant worker-desired level in 47 of 104 occupations, while H5 tasks were associated with Interpersonal Communication and Domain Expertise.Experts rated 16 occupations predominantly H1, whereas workers predominantly desired H5 only for Editors.Source: §3.3 · pp. 12–13, Figure 6 · p. 12
  • Workers most commonly envisioned role-based AI support at 23.1%, supportive assistants at 23.0%, and pure automation for some workflow aspects at 16.5%.These responses indicate substantial interest in collaboration alongside automation.Source: §3.3 · p. 13

3.4 The Potential Shift of Core Human Skills

  • High-human-agency tasks were less associated with information processing and more associated with interpersonal, organizational, decision-making, and quality-judgment skills.The analysis compared O*NET skill mappings, average wages, and expert-assessed human agency to identify these emerging patterns.Source: §3.4 · pp. 14–15, Figure 7 · p. 14

4 Related Work

Prior research has examined AI agents and generative AI, but often within constrained domains or capability benchmarks. This paper extends that literature by auditing worker-centered workplace integration across a broader occupational task landscape.

Digital AI Agents

  • Recent AI agents use foundation models, tool access, and multi-step execution to perform workflows across software engineering, analytical writing, and customer support.Prior work also shows that human-agent collaboration can outperform fully autonomous agents on some tasks.Source: §4 · p. 15
  • Existing agent benchmarks often use constrained task collections and therefore provide limited worker-centered coverage of broader occupational integration.The paper addresses this limitation by sourcing tasks from the U.S. Department of Labor’s O*NET database.Source: §4 · p. 15

The Economic Impacts of Generative AI

  • Prior labor-impact research estimated that approximately 80% of the U.S. workforce has some tasks exposed to LLM capabilities but did not incorporate worker desire or the broader scope of AI agents.Real-user chatbot data and customer-support field studies provide complementary evidence on usage and productivity.Source: §4 · pp. 15–16

5 Conclusion

The paper presents WORKBank as a worker-centered audit of AI-agent automation and augmentation, combining worker preferences with expert capability assessments. Its findings suggest diverse human-agency expectations and potential shifts in workplace skills, while the audit remains bounded by its task, participant, and time coverage.

5 Conclusion

  • The conclusion identifies diverse Human Agency Scale patterns across occupations, including a dominant inverted-U trend that supports human–agent collaboration.The findings extend beyond a binary automate-or-not view and distinguish automation from augmentation.Source: §5 · p. 16
  • WORKBank combines worker perspectives with AI-expert assessments to inform AI-agent research, deployment, and workforce-development strategies.The framework is implemented through an audio-enhanced survey with task ratings and mini-interviews, alongside expert annotations.Source: §5 · p. 16, A Survey Details · p. 21
  • The authors propose examining worker reskilling and retraining as a future direction because AI agents may reshape demand for core human skills.The conclusion frames this as an implication for future research rather than a tested intervention.Source: §5 · p. 16
  • The audit is limited because its quantitative assessments use existing O*NET tasks, may miss emerging work, and cover only a subset of occupations as of early 2025.Worker awareness, incentive-related response biases, and evolving AI capabilities further constrain interpretation.Source: §5 · pp. 16–17

A Survey Details

  • The audit uses structured task ratings for automation desire and desired human agency, supplemented by open-ended accounts of work processes and AI use.Workers rate task familiarity, automation concerns and preferences, collaboration requirements, and human-agency factors on structured items.Source: A Survey Details · pp. 21–22

B.1 Annotation Agreement of AI Expert Assessments

  • AI-expert annotations were collected from 52 qualified experts, with every task independently assessed by at least two experts and additional reviews enforcing rating consistency.Inter-annotator agreement was Krippendorff’s α=0.539 for automation capability and 0.511 for feasible human-agency levels.Source: §B.1 · p. 22

B.2 Mixed-Effects Model Regression on Worker Responses

The analysis tests whether automation-desire ratings contain task-level signal after accounting for worker characteristics, using mixed-effects regression and an O*NET-based task dataset. It finds meaningful between-task variation and stable four-zone distributions under a high-signal robustness check.

B.2 Mixed-Effects Model Regression on Worker Responses

  • The regression models worker i’s automation-desire rating for task j using demographic and attitude covariates, fixed effects, a task-specific random intercept, and residual error.The model was estimated by REML using statsmodels MixedLM.Source: §B.2 · pp. 22–23
  • Roughly 5% of total variance in automation desire is attributable to between-task differences, indicating that task identity carries a real signal after controlling for individual differences.The model uses task-specific random intercepts alongside demographic and attitude covariates.Source: §B.2 · p. 23
  • Higher education, longer work experience, income, and agreement that AI relieves tedious work were associated with greater automation desire, whereas stronger AI-suffering attitudes were associated with lower desire.Reported coefficients include β̂=0.236 for doctorate versus bachelor’s, β̂=0.229 for over 10 versus 1–2 years, β̂=0.685 for relieving tedious work, β̂=−0.438 for AI suffering, and β̂=0.687 for $529K+ versus $0–30K.Source: §B.2 · pp. 23–24
  • The high-signal subset produced an ICC of 0.405, while zone shares remained similar to the full dataset across Green Light, Red Light, Low Priority, and R&D Opportunity categories.The retained 238 tasks yielded 41.60%, 31.09%, 15.13%, and 12.18% in those zones, respectively.Source: §B.2 · p. 24

C.1 Occupational Information Network (O*NET)

  • The study starts with 18,796 O*NET task statements spanning 923 occupations and filters them to computer-centered occupations, computer-completable tasks, complete Core/Supplementary annotations, and tasks performed more than monthly.Source: §C.1 · p. 24

C.2 Occupational Employment and Wage Statistics by U.S. Bureau of Labor Statistics

  • BLS employment and annual mean-wage statistics are combined with WORKBank data to contextualize the analysis of task skills, wages, and human agency.The source is the BLS May 2024 Occupational Employment and Wage Statistics Query System.Source: §C.2 · p. 25

C.3 Claude.ai Usage Data

The study compares WORKBank’s task-level automation desires with Claude usage from the Anthropic Economic Index. The occupations with the highest average automation desire account for only a small share of observed Claude usage.

C.3 Claude.ai Usage Data

  • The top 10 occupations with the highest average automation desire represent only 1.26% of total Claude usage.The comparison is possible because both datasets use standardized O*NET task definitions.Source: §C.3 · p. 25

C.4 Y Combinator (YC) Company and AI Agent Research Paper Data

The study maps AI-related companies and agent research papers to WORKBank tasks to compare current investment and research activity with the desire–capability landscape. It also documents the WORKBank coverage used for these analyses.

C.4 Y Combinator (YC) Company and AI Agent Research Paper Data

  • An LLM-assisted pipeline systematically maps relevant YC startups and academic agent publications to specific occupational tasks in WORKBank.The mapping uses binary company-to-task and paper-to-task classifications, defaulting to False for incidental, indirect, or speculative links.Source: §C.4 · pp. 25, 27
  • The YC collection began with 5,156 companies and identified 1,723 AI-related companies after LLM-based filtering.Company descriptions were collected from official YC detail pages retrieved on April 28, 2025.Source: §C.4 · p. 25
  • The research-paper pipeline screened 17,064 arXiv papers and retained 1,222 papers meeting agent-system, implementation, task-evaluation, and realistic-task criteria.Each retained paper received an extracted representative task statement for subsequent mapping.Source: §C.4 · pp. 25–26

D WORKBank Statistics

  • The WORKBank statistics section reports 1,500 individual assessments across 104 occupations, retaining occupations with at least ten worker responses collected from January to May 2025.The database includes occupation-level participant counts for the covered occupations.Source: D WORKBank Statistics · p. 28, Table 2 · p. 28

D.1 Full List of Included Occupations

  • The demographic distribution of worker participants is compared with U.S. workforce demographics for the same covered occupations using Bureau of Labor Statistics data.This comparison is presented in Figure 8.Source: Figure 8 · p. 29

D.2 Coverage of the U.S. Workforce

WORKBank is evaluated against U.S. workforce data for sector coverage and participant demographics, using workers recruited through multiple platforms and AI experts from varied career stages.

D.2 Coverage of the U.S. Workforce

  • WORKBank captures a broad and representative cross-section of the U.S. workforce at the sector level.The comparison covers both the full U.S. workforce and the 104 occupations included in the database.Source: §D.2 · p. 29

D.3 Domain Worker Demographic Information

  • The domain-worker pool has comprehensive demographic coverage, with a tendency toward younger age groups.Demographics are compared with 2024 BLS Current Population Survey data restricted to the same 104 occupations.Source: §D.3 · p. 29

D.4 Recruitment Sources

  • Worker assessments were recruited mainly through Prolific, Upwork, and LinkedIn, while expert assessments came from 52 AI experts recruited through word of mouth.The experts included PhD students, postdocs, professors, and industry practitioners across career stages.Source: §D.4 · p. 29, Table 4 · p. 30
  • Recruitment-source counts are reported before filtering, and post-filter source breakdowns are unavailable because participant identities were not linked to WORKBank data.The privacy-preserving design prevents reporting recruitment-source breakdowns after filtering.Source: Table 3 · p. 30

E.1 Top 20 Tasks Workers Want Automated

The paper reports the 20 occupational tasks with the highest average worker desire for automation.

E.1 Top 20 Tasks Workers Want Automated

  • Table 5 lists the top 20 tasks with the highest average automation desire.The supplied passage identifies the table’s scope but does not provide the task names or values.Source: Table 5 · p. 31
  • The ranking is based on average worker automation desire.No further comparison or numerical result is provided in the supplied passage.Source: Table 5 · p. 31
  • The table provides a task-level view of which work activities workers most want automated.This task-level scope is indicated by the table title.Source: Table 5 · p. 31

E.2 Bottom 20 Tasks Workers Want Automated

The paper identifies tasks with the lowest average worker desire for automation and situates task zones within broader AI-agent investment trends.

E.2 Bottom 20 Tasks Workers Want Automated

  • Table 6 lists the bottom 20 tasks with the lowest average automation desire.The supplied passage does not provide the task names or their values.Source: Table 6 · p. 32
  • The supplied evidence does not specify which individual tasks comprise the lowest-automation-desire ranking.Only the scope of Table 6 is identified.Source: Table 6 · p. 32
  • AI-agent startup growth has not produced disproportionate concentration in the Automation “Green Light” or R&D Opportunity Zones.Across 2006–2025, newly funded Y Combinator companies were distributed relatively uniformly across task zones.Source: Figure 9 · p. 33

E.3 Y Combinator Investment Patterns Across Task Zones

The paper examines investment patterns, human-agency mismatches, task-to-skill mappings, transcript topics, and worker examples to characterize AI-agent integration across work.

E.3 Y Combinator Investment Patterns Across Task Zones

  • Y Combinator investment remained relatively uniform across the four task zones despite exponential AI-agent startup growth.The authors identify no disproportionate concentration in the Automation “Green Light” or R&D Opportunity Zones and characterize this as a possible investment misalignment.Source: §E.3 · p. 33

E.4 Full Human Agency Scale Results

  • 45.2% of occupations have H3, or equal partnership, as the dominant worker-desired Human Agency Scale level.Workers generally prefer higher human-agency levels than experts deem technologically necessary, potentially foreshadowing friction as capabilities advance.Source: Figure 10 · p. 34

E.5 Top 10 Occupations By Worker-Expert Discrepancies in HAS Ratings

  • Worker-desired and expert-assessed Human Agency Scale distributions are compared using Jensen–Shannon divergence at the occupation level.Table 7 ranks the 10 occupations with the largest discrepancies using distributions computed from all annotated tasks within each occupation.Source: Table 7 · p. 35

E.6 Task to Skill Mapping

  • O*NET maps tasks through Detailed Work Activities to exactly one Generalized Work Activity, enabling direct task-to-skill matching.The audit excludes several physical skills and “Working with Computers” because it focuses on tasks likely exposed to digital AI agents.Source: §E.6 · p. 35

F.1 LLM-based Topic Modeling

  • LLooM analyzes worker audio transcripts by distilling relevant quotes, clustering summarized quotes with an LLM, and manually inspecting or merging clusters.The resulting concept groups support analyses associated with the paper’s worker-desire and human-agency sections.Source: §F.1 · pp. 35–36

F.2.1 Full Transcripts for Direct Quotes in §3.1

  • Workers describe using AI for research, writing revision, task summaries, and workflow efficiency while retaining human control over creative content and design decisions.The interviews include explicit resistance to AI replacing artists, brainstorming on workers’ behalf, or performing content creation.Source: §F.2.1 · pp. 36–38

F.2.2 Transcripts from Occupations Exhibiting High Human Agency (HAS) Levels

Transcripts from high-human-agency occupations show that workers describe AI as limited, selectively useful, or requiring careful human oversight. These accounts span editing, mathematics, and aerospace engineering, illustrating different boundaries around acceptable assistance.

Cross-occupation comparison

  • Across the section, the examples align with the finding that only a small number of occupations have dominant H5 preferences, with editors the sole worker-identified occupation.AI expert assessments instead place mathematicians and aerospace engineers in this category, showing that worker and expert judgments need not coincide.Source: §F.2.2 · p. 38

Editors

  • Editors resist routine AI adoption and would limit it to basic grammar editing, checking every suggestion carefully before accepting changes.Their work involves repeated passes for grammar, clarity, flow, consistency, and design, with responsibility for accurate, print-ready materials.Source: §F.2.2 · pp. 38–39

Mathematicians

  • Mathematicians describe current AI tools as largely unhelpful because their work depends on developing novel problems, understanding lengthy papers, and applying mathematical intuition.One mathematician distinguishes solving crafted problems from producing genuinely new ideas, while another reports that AI is currently useless for the work.Source: §F.2.2 · pp. 40–41
  • Mathematical formalization creates more concrete opportunities for AI assistance in theorem search, proof completion, coding, debugging, and auto-formalization.The described workflow uses Lean to state theorems, search libraries, fill proofs, and revise formalizations, with Python and DeepSeek supporting training and debugging.Source: §F.2.2 · p. 41

Aerospace Engineers

  • Aerospace engineers see AI as potentially useful for detecting engine problems and reducing manual diagnostic work, while current simulation workflows use programming, testing, and validation.The reported tasks include aircraft design, maintenance, and safety analysis, as well as programming simulations for orbital-rocket performance.Source: §F.2.2 · p. 42
Loading 2506.06576v3…