Employees in moderately complex jobs are most ready to adopt generative AI: both low- and high-complexity roles report lower AI confidence and fewer adoption behaviors, challenging the assumption that highly skilled roles naturally lead AI integration.

Main Finding

Employees’ confidence in using generative AI (AI self-efficacy) follows a curvilinear relationship with job complexity: both low- and high-complexity jobs are associated with lower AI self-efficacy and therefore lower AI adoption readiness and behaviors, while moderately complex jobs show higher self-efficacy and greater adoption. This contradicts the simple assumption that higher-skilled (more complex) roles will automatically lead AI integration and highlights task-technology fit as a key determinant of adoption.

Key Points

• Theoretical framing: combines Task-Technology Fit and self-efficacy theory to explain how task characteristics shape confidence and adoption of generative AI.
• Curvilinear effect: job complexity → AI self-efficacy is inverted-U shaped: moderate complexity maximizes self-efficacy; extremes (very low or very high complexity) reduce it.
• Mediation: AI self-efficacy mediates the relationship between job complexity and both AI adoption readiness (attitudinal) and observable adoption behaviors.
• Empirical challenge: findings counter the expectation that highly skilled roles will be early and eager adopters; high complexity can reduce confidence if tasks are poorly aligned with AI capabilities.
• Practical implication: fostering adoption requires aligning task structure with AI strengths and investing in targeted job design and training — not just deploying tools across high-skill roles.
• Boundary condition: introduces a non-linear boundary to standard technology adoption models (which often assume monotonic effects of skill/complexity).

Data & Methods

• Design: Two employee survey studies testing the proposed hypotheses.
• Key constructs: job complexity, AI self-efficacy, AI adoption readiness, and AI adoption behaviors (self-reported).
• Analysis: structural equation modeling to estimate relationships and test mediation and curvilinear effects.
• Samples: employee respondents across a range of occupations/roles (multi-role/organizational sampling; details in original paper).
• Robustness/limitations:
  • Strength: replication across two studies strengthens internal consistency.
  • Limitation: surveys are cross-sectional and rely on self-report measures, limiting causal inference and maybe subject to common-method bias.
  • Future work: longitudinal, experimental, or administrative usage data would help establish causality and generalize findings across industries and AI tool types.

Implications for AI Economics

• Diffusion and productivity effects:
  • Adoption will not be monotonically higher in more complex (higher-skill) occupations; moderate-complexity roles may see the fastest uptake and near-term productivity gains.
  • Models of AI diffusion should include task complexity as a non-linear moderator when forecasting adoption rates and aggregate productivity growth.
• Labor demand and wage dynamics:
  • Redistribution of gains may be heterogeneous: moderately complex roles could gain complementary benefits (raising productivity and perhaps wages), while very high- and low-complexity roles may either lag or require costly reskilling.
  • Policy expectations that high-skill occupations automatically capture AI gains may be misplaced; targeted interventions could be necessary to avoid uneven labor-market effects.
• Human capital and training investments:
  • Firms and policymakers should prioritize training and job redesign that improve task-technology fit, especially for high-complexity roles where lack of fit undermines confidence.
  • Subsidies, certifications, and practical hands-on training that build AI self-efficacy could increase adoption and the effective complementarity between workers and AI.
• Automation vs augmentation strategy:
  • Strategic decisions about which tasks to automate or augment should account for the non-linear adoption propensity across task complexity; initially deploying generative AI in moderately complex tasks may yield the highest uptake and ROI.
• Inequality and transition policy:
  • Because adoption and benefits are uneven by task structure, targeted safety nets and retraining for workers in extremes of complexity may be required to manage transition costs and distributional effects.

Overall, the paper suggests that economic models and organizational strategies around generative AI must move beyond simple skill-level proxies and explicitly model task-technology fit and non-linear adoption responses when predicting adoption, productivity, and labor-market outcomes.