Does AI assistance help or hurt less experienced entrepreneurs?

A randomized controlled trial with 640 Kenyan small-business entrepreneurs found no average effect of generative AI access on firm revenues or profits, but a sharp split underneath that null result. Half the sample was randomly given access to a GPT-4-powered business assistant over WhatsApp; the rest received public business training guides. Overall the paper is "unable to reject the null hypothesis of no average treatment effect on firm revenues and profits." Splitting the sample by pre-treatment performance, entrepreneurs with below-median baseline performance ("low performers") saw the assistant lower revenue by about 0.08 standard deviations (p=0.01), a roughly 9.2% relative decrease worth about $33.34 a month. Entrepreneurs with above-median baseline performance ("high performers") saw a suggestive 0.16 s.d. gain (p=0.07), an 18% relative increase worth about $66.69 a month. The gap between the two groups, 0.23 s.d., is itself significant (p=0.01).

The paper's own account of the mechanism is explicit: "this differential impact does not appear to result from differences in the questions posed to or the advice received from the AI, but rather from which pieces of AI advice entrepreneurs selected and implemented." The AI's output was not uneven — what entrepreneurs did with it was. The authors frame their participants as facing considerable discretion: unlike the structured writing or consulting tasks in prior studies, Kenyan entrepreneurs "had considerable discretion in how and when they used AI assistance" and "had to exercise considerable judgment in deciding when to implement the AI's advice." The paper treats this discretion, not any deficiency in the model's answers, as the proximate cause of the uneven outcome, reached through an "abductive approach" to the rich text data from the AI interactions and surveys rather than a designed causal test of the selection mechanism itself.

This cuts against the leveling pattern reported elsewhere in the library. Can AI narrow the education performance gap? finds AI assistance narrows an education-based performance gap, with the largest gains going to lower-education participants. Does AI assistance help less experienced workers most? finds the same direction in a live customer-support deployment: the least experienced gained the most, the most experienced saw small declines. This Kenya entrepreneurship RCT points the opposite way — the initially stronger performers gained and the initially weaker performers lost ground. Setting may explain the reversal: those two studies involve AI giving in-context suggestions on a bounded task (writing a response, solving a set business problem), while here entrepreneurs get open-ended advice they must choose whether and how to apply across "myriad different tasks." The discretion over deployment that this paper treats as the key variable was largely absent in the other two settings, where the tool's advice was harder to misapply. Does AI-enriched analyst reports improve forecast accuracy? shares this paper's framing move: the uneven outcome is attributed to a human-side limit (processing load there, judgment in selection here) rather than to the quality of what the AI produced.

The excerpt does not show why low performers implemented advice poorly — the abductive analysis of the text data is asserted as the paper's approach but not shown here. The overall null effect means this is a story about heterogeneity within a non-finding, not about a working intervention: averaged across all 640 entrepreneurs, access to the assistant did nothing measurable. The high-performer gain is only "suggestive" at p=0.07 and would not clear significance on its own; only the low-performer loss and the gap between groups clear p=0.01. The sample is a single emerging-market context — mostly young, recently founded Kenyan businesses reached over WhatsApp — so the finding does not establish that AI advice generally helps strong performers and hurts weak ones elsewhere. The defensible claim is narrower: in this setting, who benefited from open-ended generative AI advice depended on entrepreneurs' baseline performance and on which advice they chose to act on, not on what the AI produced.