Mapping AI into Production: A Field Experiment on Firm Performance
Citation. Kim, H., Kim, D., & Koning, R. (2026). Mapping AI into Production: A Field Experiment on Firm Performance. INSEAD Working Paper 2026/20/STR. SSRN abstract 6513481. Preregistered AEA RCT Registry AEARCTR-0016746.
Summary. A preregistered randomized field experiment across 515 high-growth startups in a 3-month INSEAD accelerator. All firms got identical AI tools, credits, and technical training; the only treatment was case-study workshops showing how AI-native firms reorganize production. The paper isolates what the authors call the mapping problem — discovering where and how AI creates value inside a firm’s production process — and shows that relieving it causally raises firm performance. The binding constraint is managerial/cognitive search, not access to the technology.
Key claims
- Treated firms discovered 44% more AI use cases (8.8 vs 6.1; p<0.01), spread across more functions (+0.84 functional categories), concentrated in product development and strategy.
- They completed ~12% more tasks (20.9 vs 18.5; p<0.05), driven almost entirely by internal tasks (+22%); external tasks unchanged.
- They were 18% more likely to acquire paying customers (+11pp; p<0.01) and generated 1.9× higher revenue (2.2× conditional on any revenue; p<0.05).
- Gains concentrate in the upper tail (90th–95th pct) — AI “lifts the ceiling,” not the floor. No significant heterogeneity by baseline performance or technical-founder background.
- Capital demand fell
39.5% ($220k; p<0.05) with labor demand unchanged — “do more with less.” - IV/2SLS (treatment instrumenting AI use, first-stage F>33): each additional use case → +0.85 tasks, +3.9pp customer probability, +0.23 log revenue, −0.17 log capital demand.
Evidence & limitations
- Design: preregistered RCT (515 firms, 255T/260C), stratified randomization, strata fixed effects, firm-clustered SEs; IV/2SLS for returns to AI use. Primary outcomes double-coded blind to condition. Low attrition (1.6%, balanced); sentiment checks rule out a Hawthorne/motivation story.
- Limitations: working paper, not peer-reviewed; short 3-month window (no evidence on whether gains compound or fade); several headline results are marginal (p<0.1); revenue effect rides on a few upper-tail “breakout” firms; early-stage, self-selected, AI-focused accelerator applicants — not representative of established firms. Authors flag 0→8.8 use-case extrapolations as “heroic.”