Future of Work with AI Agents

Analysis

Why This Paper Matters

As compound AI systems (AI agents) rapidly enter the workplace, there is growing concern about job displacement, loss of human agency, and overreliance on automation. Yet, until now, there has been no systematic way to understand which tasks workers actually want AI to automate versus augment, and how those desires align with what AI can currently do. This paper fills that gap with a rigorous, human-centered auditing framework that goes beyond simplistic automation-or-not debates.

The significance lies in its direct relevance to practitioners building AI agents for enterprise and labor markets. By grounding the analysis in worker preferences and expert capability assessments, the paper provides a data-driven map of where AI development should focus (e.g., tasks workers want automated and AI can do) and where it should proceed cautiously (tasks workers want to retain control over). This is critical for avoiding backlash, ensuring adoption, and designing AI that complements rather than replaces human skills.

Technical Contributions

• Audio-Enhanced Mini-Interview: A novel method to capture nuanced worker desires about AI involvement, going beyond simple survey questions to elicit richer qualitative data.
• Human Agency Scale (HAS): A shared, quantifiable metric for preferred level of human involvement in tasks, enabling cross-occupation comparisons.
• WORKBank Database: A large-scale resource with 1,500 workers and AI experts covering 844 tasks across 104 occupations, built on the U.S. Department of Labor's O*NET taxonomy.
• Four-Zone Taxonomy: Tasks are classified into Automation Green Light (desired and feasible), Automation Red Light (desired but not feasible), R&D Opportunity (feasible but not desired), and Low Priority (neither desired nor feasible), highlighting mismatches.
• HAS Profiles: Reveals that different occupations have distinct patterns of desired human involvement, challenging one-size-fits-all automation strategies.

Results

The paper does not report traditional quantitative metrics like accuracy or F1 scores, as it is a survey and framework paper. Instead, its key results are qualitative and structural: (1) identification of critical mismatches between worker desires and AI capabilities across 844 tasks; (2) demonstration that HAS profiles vary significantly across 104 occupations, indicating heterogeneous expectations; (3) early evidence that AI agent integration may shift valued human competencies from information-processing skills (e.g., data analysis) toward interpersonal skills (e.g., communication, empathy). The WORKBank database itself is a major output, enabling future research.

Significance

This work has broad implications for AI practitioners, policymakers, and workforce developers. For AI developers, it provides a roadmap for prioritizing agent capabilities that align with worker desires, potentially increasing adoption and reducing resistance. For organizations, it offers a framework to assess which tasks to automate, augment, or leave to humans, balancing efficiency with human agency. The shift toward interpersonal competencies suggests that training and education should emphasize soft skills as AI takes over information tasks. Overall, the paper moves the conversation from abstract fears about AI replacing jobs to a nuanced, data-driven approach for human-AI collaboration.