Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
Citation. Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, Articles in Advance. https://doi.org/10.1287/orsc.2025.21838
Summary. A preregistered randomized field experiment with 758 Boston Consulting Group consultants tested GPT-4 assistance on realistic management consulting tasks. The paper introduces the “jagged technological frontier”: AI improves performance on some knowledge-work tasks but worsens it on others that appear similarly difficult to humans. Inside the frontier, AI users completed more tasks, worked faster, and produced higher-quality outputs; outside it, AI users were less likely to solve the task correctly. The design implication is that AI adoption depends on task-level capability mapping and calibrated human judgment, not tool access alone.
Key claims
- On 18 tasks selected to be inside the AI capability frontier, consultants using AI completed 12.2% more tasks and completed them 25.1% more quickly on average than consultants without AI.
- Inside-the-frontier AI use also improved output quality; the authors report significant positive treatment effects on expert-rated performance across creative, analytical, writing, and persuasiveness dimensions.
- On a complex managerial task selected to be outside the frontier, consultants using AI were 19 percentage points less likely to produce correct solutions than consultants without AI.
- AI assistance improved the subjective coherence of outside-the-frontier responses even when correctness fell, showing how fluent AI-supported reasoning can make wrong answers look more persuasive.
- The same professional domain can contain both AI-augmentable and AI-degrading tasks, making the practical adoption problem one of task diagnosis, workflow mapping, and calibrated human verification.
Evidence & limitations
- Design: preregistered randomized laboratory-in-the-field experiment with 758 BCG consultants. Participants first completed baseline tasks, then were randomly assigned to no AI, GPT-4 access, or GPT-4 access plus a prompt-engineering overview. Tasks were realistic consulting exercises; AI interaction logs were captured through a GPT-4 platform.
- Evidence strength: strong causal design for short-run task performance among highly skilled consultants, with blind expert grading and quantitative productivity, quality, timing, and correctness outcomes.
- Limitations: conducted in one professional-services firm and one period of GPT-4 capability; frontier boundaries are task- and model-specific and may shift quickly; the results do not show long-run learning, deskilling, adoption durability, equilibrium labor demand, or organizational performance effects.