Startup founders who received training on how other companies organize work around artificial intelligence identified 44% more potential uses for AI than a comparison group in a field experiment involving 515 ventures.
The researchers also report differences in tasks completed, customer acquisition and revenue. Their 2026 working paper studies a group of high-growth startups in an accelerator, so its results should not be presented as a forecast for established businesses or as a guaranteed return from an AI tool.
Hyunjin Kim, Dahyeon Kim and Rembrand Koning investigated what they call a mapping problem: finding the parts of a business where a technology can help and redesigning the work around it.
What the experiment compared
According to INSEAD’s summary of the working paper, firms assigned to the treatment received examples of how other companies reorganized production around AI. The researchers report that the additional uses were concentrated in product development and strategy.
The study was conducted in a startup program, with participating ventures assigned to different training. The intervention was a package of information and instruction, so the outcome should not be attributed to one example or one software product.
The authors report that the treatment group completed 12% more tasks and was 18% more likely to acquire paying customers. They also report higher revenue, with gains concentrated among firms toward the upper end of the distribution. The figures are research estimates for the study period, not promises of future performance.
Learning to redesign work
Buying access to a tool does not identify which task should use it. A startup may first need to identify where staff spend time, what information is needed and how an output will be checked.
For example, a team could test whether a tool helps organize customer feedback before considering whether it belongs in a product-planning workflow. Human review, data protection and the cost of correcting errors remain part of the decision.
Training may help people notice possibilities, but it can also consume time and resources. A company should compare the full cost of a workflow with an appropriate alternative, including software fees, supervision and any changes to roles.
Limits of the findings
The paper is preliminary and the firms were selected for a high-growth startup program. Their teams, incentives and access to resources may differ from those of an established manufacturer, retailer or public agency. A short experiment also cannot determine whether gains persist for years.
The research did not show a proportional increase in labor demand among the participating firms, according to the authors. That result describes this experiment and should not be used to predict the effect of AI on employment across the economy.
For companies exploring the technology, a sensible starting point is to identify a specific task and measure whether a revised process delivers reliable results. Our earlier coverage of how business practices can spread through board connections looked at a different way organizations encounter ideas from other firms.
