Generative AI accelerates knowledge work but quality effects depend on task: it improves packaging and creation output but undermines acquisition, while primarily boosting lower performers and narrowing skill gaps.

Main Finding

GenAI support meaningfully speeds up knowledge work across task types, but effects on output quality are task-dependent: quality improves for knowledge packaging and knowledge creation, yet declines for knowledge acquisition. GenAI also reduces performance variance (benefiting lower-performing workers) for packaging and creation, while increasing variance for acquisition tasks.

Key Points

• Study framework: task-technology-fit (TTF) used to predict heterogeneous impacts across task types and productivity dimensions (efficiency vs quality).
• Efficiency: GenAI significantly reduces task completion time for all three task types (knowledge acquisition, packaging, creation).
• Quality:
  • Knowledge packaging: quality increases with GenAI.
  • Knowledge creation: quality increases with GenAI.
  • Knowledge acquisition: quality decreases with GenAI.
• Variance / Distributional effects:
  • Packaging & creation: reduced variance in quality — largest gains accrue to lower-performing workers, narrowing performance gaps.
  • Acquisition: increased variance in quality across workers.
• Practical takeaway: benefits are not uniform — task type and worker skill interact with GenAI to produce distinct productivity trade-offs.

Data & Methods

• Design: 2 × 3 mixed lab-in-the-field experiment.
  • Between-subjects factor: GenAI support (with vs without).
  • Within-subjects factor: task type (knowledge acquisition, packaging, creation).
• Sample: 128 knowledge workers from a large multinational industrial organization; participants randomly assigned to treatment/control.
• Tasks: three representative knowledge-work tasks (one per type) completed individually in the experimental setting.
• Outcome measures:
  • Efficiency: task completion time (system-logged).
  • Output quality: expert evaluations against predefined, task-specific constructs (blind-rated).
• Key analytic contrasts: average effects on time and expert-rated quality, plus analyses of variance and heterogeneity by baseline performance.

Implications for AI Economics

• Macro productivity estimates must account for task heterogeneity. Aggregate gains from GenAI (e.g., trillion-dollar estimates) should be tempered by task-dependent quality effects — some task categories (acquisition) may show time savings but degraded quality.
• Distributional consequences: GenAI can compress within-firm performance disparities in some tasks (packaging, creation), implying potential wage and promotion impacts favoring lower-skilled workers in those tasks; but for acquisition tasks, widening variance could amplify inequality.
• Deployment strategy and ROI:
  • Firms should prioritize GenAI for packaging and creation tasks where both speed and quality tend to improve and where lower-skilled workers gain most.
  • For knowledge acquisition tasks, firms should be cautious: pair GenAI with stronger human oversight, validation processes, or tool adjustments to avoid quality losses (e.g., hallucinations, shallow retrieval).
• Policy and measurement:
  • Evaluations of GenAI-driven productivity should measure both efficiency and multiple quality dimensions (not only time or output counts).
  • Labor-market models of AI-driven automation/augmentation should incorporate heterogeneous task-level quality effects and changes in variance across workers.
• Research and modeling recommendations:
  • Incorporate TTF-like alignment parameters (task characteristics × worker skill × tool capability) into macro productivity and labor models.
  • When projecting economic impacts, adjust for the differing prevalence of task types across sectors and occupations; sectors dominated by acquisition-like tasks may see smaller net welfare gains or require investment in mitigation.
• Organizational implications: invest in targeted training, task-specific prompt engineering, monitoring and validation workflows, and adaptive allocation of GenAI to tasks where it demonstrably improves both speed and quality.

Limitations to keep in mind when applying these implications: single-organization sample, lab-in-the-field setting, and task/tool specifics (GenAI model/configuration) may limit generalizability. Future work should replicate across sectors, tools, and a broader set of task definitions.