Applying Clinical Trial Logic for AI: how Randomized Controlled Experiments build organizational knowledge.

Artificial intelligence is reshaping how organizations create, transfer, and apply knowledge, yet its effects remain highly context-dependent. This study examines whether randomized controlled trials (RCTs) can serve not only as evaluation tools, but also as mechanisms for organizational learning. We conducted a large-scale field experiment with 291 Accenture Italy consultants, randomly assigned to work individually with generative AI, in seniority-homogeneous pairs with AI, or in seniority-heterogeneous pairs with AI. Performance was assessed across quality, impact, novelty, feasibility, and technical sophistication, complemented by text analysis of final reports and AI interaction transcripts. Results show that homogeneous teams outperform individuals and heterogeneous teams on several performance dimensions, suggesting that shared professional frames reduce coordination costs under time pressure and task novelty. At the individual level, higher performance is associated with iterative-evaluative prompting, in which users repeatedly ask AI to assess, compare, and refine ideas. Linguistic analyses also reveal two distinct capabilities: expressive knowledge, required to communicate effectively with clients, and instrumental knowledge, required to instruct AI systems effectively. These findings extend absorptive capacity theory to AI-augmented work and show how field experimentation can generate granular, causal, organization-specific knowledge to guide AI adoption and team design.