At a glance
- Source
- Otis, Clarke, Delecourt, Holtz and Koning (2026). The Uneven Impact of Generative AI on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya. Management Science (published 10 July 2026). doi:10.1287/mnsc.2024.06909
- How it was done
- Owners were assigned by chance to get a GPT-4 business adviser on WhatsApp or links to standard business training guides. The researchers compared revenue and profit, as reported by the owners, over about five months. Randomised experiment. People were assigned by chance to work with or without the tool, and the two groups were compared.
- Who took part
- 640 small-business owners from 44 of Kenya’s 47 counties. Most ran food, clothing or farming businesses.
- Where
- Kenya, from late May to early November 2023.
“This differential impact does not appear to result from differences in the questions posed to the AI or the advice it provided, but rather from the advice entrepreneurs chose to implement.”
What it found
- No average effect. Across all owners, the AI adviser made no detectable difference to revenue or profit.
- The split underneath. Owners whose businesses were below the middle on revenue and profit at the start did nearly 10% worse with the adviser. Owners above the middle may have done more than 15% better.
- Same questions, same advice. The two groups did not differ in how many questions they asked, what kind, or the advice the AI gave them.
- What they acted on. Stronger owners used the AI to find changes that fitted their own business. Weaker owners more often took generic advice, such as cutting prices or spending more on advertising, which may have lowered revenue and raised costs.
When the result holds
Running a small business is open-ended work. Nobody checked the advice against a standard before the owner acted on it. Studies that find less experienced staff gain most looked at well-defined tasks, where the output is easy to check. Here the weakest gained nothing or lost.
The participants ran very small firms. Their average monthly profit before the experiment was about $137.
Limits
- Revenue and profit were reported by the owners. The researchers gave the comparison group business guides so that both groups had a reason to report the same way.
- The adviser ran on GPT-4 in 2023. Later models may give better advice.
- One country. “Low performer” means low revenue and profit at the start, not low skill.
Practices that use this study
This study supports:
- For each kind of task, test whether a person with AI beats either one alone Do not assume that adding a person to AI, or AI to a person, gives the best result. Test it for each type of task before you settle the process. Do next · Effort: medium (weeks, a few people)
- Check AI advice on money, tax and law before you act on it When you use an AI tool as an adviser, check its advice on money, tax and law against a reliable source or a professional before you act. In a small firm, nobody else will catch a wrong answer. Do next · Effort: low (days, one person)
This study limits or contradicts: