Why did giving AI advice to real entrepreneurs widen the gap between strong and weak performers, instead of closing it?

This explores why GPT-4 narrowed the skill gap among BCG consultants but widened it among Kenyan small-business owners. The short answer from the corpus is that the AI's advice was not the deciding factor. What decided the outcome was who chose which advice to follow. In a randomized trial of 640 Kenyan entrepreneurs given a GPT-4 assistant on WhatsApp, revenue rose 18% for high performers and fell 9% for low performers. The researchers trace the split to the entrepreneurs' own choices about whether and how to act on the advice, not to worse advice going to the weaker group Does AI assistance help or hurt less experienced entrepreneurs?.

Compare the setup in the consulting study. There, 758 consultants were required to use the AI on tasks someone else had assigned and defined. Under those conditions the bottom half gained the most, and the gap narrowed Does AI help lower-skilled workers catch up to top performers?. That note already flags the risk that showed up in Kenya: once using AI becomes optional, the equalizing effect can flip. A consultant gets a clear task and a clear standard for a good deliverable. An entrepreneur has to work out what the problem is, decide which of the AI's suggestions fits their market, and live with the result. Weaker performers lose most in that step, because acting on advice well takes the same judgment they're short on.

Other notes support this from different angles. Aggregated LLM evaluations of business ideas only start to match expert rankings after many judgments are pooled; any single response is inconsistent and biased Can aggregated LLM rankings match human expert judgments?. An entrepreneur reading one WhatsApp reply gets one of those noisy single draws, and only someone with good business sense can tell a useful suggestion from a confident but wrong one. Evans makes a related point about tools at work: making a tool easier to use doesn't help people recognize which of their problems it can actually solve Does easier tool-building actually solve enterprise adoption problems?.

The surprise is that the consulting firm's structure was doing much of the equalizing. Microsoft's survey of 20,000 AI users attributes about two-thirds of AI's impact to organizational factors such as culture, manager support and incentives, and only about a third to individual effort Why do ready workers struggle to transform their work?. BCG supplied the consultants with structure: defined tasks, mandatory use and set standards. Kenyan sole proprietors had no such framework, so each had to supply the judgment themselves. The lesson isn't that AI favors the skilled in poorer economies. It's that when people use AI on their own, its benefits follow existing judgment, and narrowing the gap may depend on putting structure around the tool. The corpus has only these two head-to-head studies, so read this as a strong pattern rather than a settled rule.