WORKBank Database: Auditing AI in US Workforce

• WORKBank Database is a comprehensive resource analyzing the automation and augmentation potential across 844 tasks in 104 U.S. occupations.
• It integrates dual-stakeholder ratings from workers and AI experts using audio-enhanced surveys and structured scales to assess technical feasibility and desired human agency.
• The framework categorizes tasks into actionable zones, revealing both opportunities for AI deployment and areas prone to workforce friction.

The WORKBank Database is a comprehensive, empirically validated resource and methodological framework designed to audit the automation and augmentation potential of occupational tasks in the context of AI agent deployment across the U.S. workforce. Building on foundational occupational data from the U.S. Department of Labor’s O*NET database, WORKBank systematically cross-references worker preferences and expert capability assessments for over 844 distinct tasks spanning 104 occupations. This enables a rigorous mapping of the alignment and disjunctions between worker desires for AI support, the technical feasibility of current AI agents, and the broader implications for human agency and labor market dynamics.

1. Auditing Framework and Methodology

At the core of WORKBank is a novel, mixed-method auditing framework focused on eliciting both nuanced worker attitudes and structured ratings. The framework centers on a dual-stakeholder model:

• Audio-Enhanced Mini-Interview: Workers first engage in an audio-supported survey segment, offering natural, spoken reflections on their work practices, attitudes toward tool use, and their openness or resistance to AI integration. This step contextualizes subsequent quantitative ratings.
• Worker Ratings: For each nonphysical, recurring task relevant to their occupation (sampled from O*NET v29.2), workers provide two principal responses:
  • Desire for Automation ($A_w(t)$): Measured on a 5-point Likert scale indicating their willingness for a task to be fully automated by AI.
  • Desired Human Agency Scale ($H_w(t)$): Using a five-level Human Agency Scale (HAS), workers indicate the level of collaboration or control preferred when AI is involved.
  • H1: AI agent performs task entirely
  • H2: Minimal human input needed
  • H3: Equal human–AI partnership
  • H4: Human input required for effective completion
  • H5: AI requires constant human involvement
• Expert Ratings: Independently, AI experts (52 with practical agent R&D backgrounds) rate both the current technical automation capability ($A_e(t)$) and the expert-feasible HAS ($H_e(t)$) for the same set of tasks.

Additional design features ensure robustness: participants only rate tasks they actually perform, are guided to reflect on relevant skill and contextual factors, and data collection includes careful quality controls and demographic quotas for national representativeness.

2. Database Scope and Construction

The WORKBank Database encompasses:

• Tasks and Occupations: 2,131 screened tasks (of which 844 are rated with sufficient frequency) spanning 287 (104 with complete data) distinct, computer-using occupations.
• Participants: 1,500 domain workers and 52 AI experts completed the framework, yielding 7,016 structured ratings alongside demographic and qualitative data.
• Data Structure:
  • For each task $t$:
  • $A_w(t)$: Mean worker automation desire
  • $H_w(t)$: Mean worker-desired HAS
  • $A_e(t)$: Mean expert-rated automation capability
  • $H_e(t)$: Mean expert-feasible HAS
  • Ratings are stored at the task–occupation level, allowing aggregation and comparative analyses.

Data acquisition leveraged a custom IRB-approved online platform with recruitment through Prolific, Upwork, and LinkedIn, and covered the vast majority of major U.S. labor sectors as aligned with Bureau of Labor Statistics weighting.

3. Occupational Task Zone Taxonomy

A central analytic contribution is the stratification of tasks into four zones based on the crossing of worker automation desire and expert technical capability:

Zone	Worker Desire	Expert Capability	Interpretation
Automation Green Light	High	High	Viable for immediate, accepted AI agent deployment
Automation Red Light	Low	High	Technologically feasible but socially resisted; caution needed
R&D Opportunity	High	Low	Strong demand but limited technical feasibility; R&D focus
Low Priority	Low	Low	Neither wanted nor feasible; minimal automation interest

Distributional analysis demonstrates weak correlation (Spearman $\rho$ = 0.17) between desire and capability, indicating that workflow alignment is nontrivial and that large fractions of tasks fall into either opportunity or frictional misalignment.

4. Human Agency Scale (HAS) Profiles and Divergence

The Human Agency Scale (HAS) enables systematic quantification of the preferred and feasible levels of human involvement in task execution with AI agents. Key findings include:

• Inverted-U Distribution: Most occupations display a peak in H3 (“equal partnership”) in worker preference, signifying a general ideal of collaborative AI, rather than passive automation.
• Worker–Expert Discrepancy: Only 26.9% of tasks reach HAS agreement between workers and experts. Discrepancies often occur where experts rate full automation feasible but workers desire significant human control, highlighting potential for friction or dissatisfaction.
• Skill and Contextual Correlates: High-agency (H4/H5) tasks, by worker ratings, are most associated with interpersonal communication; expert high-agency assessments are also marked by required domain expertise.
• Occupational Extremes: Editor roles, by worker assessment, most require continuous human involvement; mathematician and aerospace engineer tasks, by expert assessment, dominate the technically high-agency end.

HAS scores enable multidimensional mapping of workforce expectations, supplementing binary automate-or-not decision frameworks.

5. Implications for Workforce Dynamics

The empirical audit supports several significant signals regarding future workforce development and AI agent integration:

• Shift in Essential Competencies: Analysis reveals a transition away from information-processing skills to interpersonal and organizational skills as the locus of high-human-agency work (see skill rank shift figures in the paper). This suggests reskilling initiatives should emphasize collaboration, management, and communication.
• Augmentation Preference: Workers frequently express desire for role-based, collaborative support rather than full replacement, aligning with trends toward augmentation rather than pure automation.
• Potential for Friction: Mismatches between worker-desired agency and what experts deem technically necessary foreshadow possible resistance or need for careful change management where full automation is feasible but not socially desired.
• Sectoral Specificity: Creative and communicative fields display higher resistance to AI-driven automation; routine and computational domains are generally more automation positive.

6. Analytical Metrics and Mathematical Models

WORKBank employs statistical modeling to assess and interpret the structure of collected ratings:

• Mixed-Effects Model for Automation Desire: Models automation desire as a function of fixed effects (demographic and attitudinal) and task-level random effects:

$$y_{ij} = \beta_0 + \sum_{k=1}^{K} \beta_k X_{k,ij} + u_j + \varepsilon_{ij}$$

• Intraclass Correlation Coefficient (ICC):

$$\mathrm{ICC} = \frac{\widehat{\sigma}_u^2}{\widehat{\sigma}_u^2 + \widehat{\sigma}_{\varepsilon}^2}$$

quantifies variance attributable to task-level differences.

• Jensen–Shannon Divergence (JSD): Used to measure the distance between worker and expert HAS distributions by occupation, identifying where misalignments are greatest.

Key worker/expert scoring functions per task $t$ include $A_w(t)$, $H_w(t)$, $A_e(t)$, $H_e(t)$.

7. Significance and Research Utility

WORKBank provides a multidimensional, workforce-scale empirical reference for stakeholders analyzing, designing, or deploying AI agents in occupational contexts.

• For policymakers and R&D organizations, WORKBank highlights both high-impact (Green Light, R&D Opportunity) and high-friction (Red Light) zones, supporting nuanced, socially-informed AI targeting.
• For labor market researchers and sociologists, the database documents a rich landscape of human preference heterogeneity, collaborative ideals, and the evolution of valued skill sets under AI transformation.
• For system designers, the dual agency scoring system and task-level breakdown guide responsible, context-sensitive, and worker-inclusive agent development practices.

In sum, WORKBank serves as both a methodological paradigm and a substantive research asset for auditing and analyzing the evolving intersection of human labor, technical advancement, and AI agent deployment in the modern economy.