A team of researchers gave hundreds of small-business owners in Kenya free WhatsApp access to a custom GPT-4 adviser, prompted to act as a Kenyan business mentor, and tracked what happened over several months. The headline finding is deceptively boring: AI access raised profits and revenues for some entrepreneurs and lowered them for others. The study, How AI Helps the Best and Hurts the Rest, was published in MIT Sloan Management Review in April 2026 by Nicholas Otis, Rowan Clarke, Solène Delecourt, David Holtz and Rembrand Koning.

What makes it worth your time is the variable that split the two groups. It wasn’t business size, sector, education or how much data the owner fed the model. It was whether the owner already had the judgement to recognise bad advice when it arrived. The stronger operators got stronger. The weaker ones got worse — they acted on advice that sounded professional but didn’t fit their situation, and the cost compounded over months.

If you’ve been hearing that AI is going to be the great equaliser for small business — the cheap mentor at scale that finally levels the playing field — this is the study that complicates the story. For a meaningful number of the owners in the trial, the AI mentor made things worse, not better.

The study, on its own terms

The study deserves a fair hearing before we push past it. Hundreds of Kenyan small-business owners were randomly assigned either to a control group with no AI access or to a treatment group with WhatsApp access to a carefully prompted, locally framed GPT-4 adviser. The team then tracked profits and revenues over time. Standard randomised-trial design; the methods are solid but not the interesting part.

Earlier studies on narrow, well-defined tasks — drafting emails, writing marketing copy, summarising documents — have consistently shown that AI helps lower-skilled workers the most. The gap closes. That’s where the “AI is an equaliser” intuition comes from, and for that kind of work, it holds.

Managing a business is different. The decisions are vague, messy, full of context the model can’t see, and the consequences of a wrong call compound over months rather than showing up immediately. The Kenya study took the equaliser hypothesis out of the lab and into work where judgement matters more than information, and the result was uncomfortable. The same adviser, available to every owner on the same terms, produced opposite outcomes depending on who was using it.

Why management decisions are different from drafting emails

The "AI levels the playing field" argument has an assumption tucked inside it that most of us don't examine closely enough. It assumes the question being put to the AI is well-formed enough that a good answer is recognisable from a bad one. For a marketing email, that's broadly true — you can read a draft and feel whether it lands. For a pricing decision, a hiring choice, or how to handle a supplier who's started paying late, you can't. A plausible answer and a good answer look identical on the page. The difference shows up six weeks later in the numbers.

This is what AI advisers can’t help with. They’ve no view of the customer who’s been complaining for a month, no idea which member of staff is about to leave, and no relationship with the supplier whose contract is up for renewal. They produce a coherent recommendation from incomplete information, which is what they’re built to do, and we’ve got to decide whether the recommendation fits the situation they can’t see.

That decision is what I’m calling the judgement gap. It’s the difference between reading an AI pricing suggestion and thinking “yes, but my margin on raw materials moved last month, this won’t work” and reading the same suggestion and thinking “great, that sounds professional, I’ll do it.”

We’ve written about related ideas from different angles in AI doesn’t replace thinking — it replaces knowing and in The AI literacy gap: why most people use a Ferrari like a bicycle. The Kenya field experiment is what those arguments look like in profit-and-loss terms. When advice is cheap and fluent, the ability to evaluate it becomes the premium skill — and the people who already have that ability pull further ahead.

What good and bad advice look like in practice

A made-up example, but representative of the kind of thing the Otis team described in their interviews. A small retailer asks the AI adviser whether to extend credit to a long-standing customer who's recently been slower to pay. The AI looks at the data the owner provides, applies a generic credit-risk framework, and recommends a partial extension with a tightened payment term. Reasonable advice in the abstract. It would pass a textbook exercise.

An experienced owner reads it and thinks: this customer’s late payments started right when his main client lost a contract. The relationship’s twenty years old. He’s never defaulted in a downturn before. None of that is in the model’s view. She declines the AI’s suggestion, picks up the phone, and the relationship continues.

A less experienced owner acts on the advice as written. The customer takes offence at the tightened terms, takes their business elsewhere, and the retailer loses one of their most reliable accounts. Technically sound advice, applied without the context that would have told you not to follow it.

Multiply this across hundreds of small decisions a year — pricing, hiring, marketing, stock, lending, contract terms, supplier relationships — and you start to see how the same adviser produces double-digit profit gains for some owners and double-digit losses for others. The AI wasn’t wrong in either case. The owners just had different abilities to read around it.

What the study doesn't quite say, but should

The natural policy reading of this research is that the AI mentor needs to be smarter, better-prompted, more careful, more locally adapted. Maybe. But that's the technology answer, and it skips over the harder point.

What this study really shows is that judgement is a prerequisite for AI advice being useful in your business, and most of the conversation about AI mentorship glosses past that rather than confronting it. The dream of a cheap mentor at scale undersells how much we’ve still got to bring to the conversation ourselves. The AI provides options; we need to know which option fits, given everything the model can’t see.

That’s a much harder thing to scale than chatbots. You can build a million WhatsApp bots in an afternoon, but you can’t mass-produce business judgement. I’m not sure what would have to change for that to become possible, and I suspect it can’t, at least not through technology alone.

A practical kit for using AI advice well

If you're running a small business and using ChatGPT, Claude, Gemini or any other AI for business decisions, the question to ask yourself isn't "is this AI good enough yet." It's whether you've built the habit of checking AI advice against your own knowledge before acting on it. A few specific moves help, roughly in the order you should apply them.
  1. Form your own answer first. Before reading the AI's response, write down what you think the right move is and why. Half a paragraph is enough. If the AI agrees, you've gained confidence. If it disagrees, you've got a real comparison rather than an authority-shaped suggestion you'll defer to. Most of us skip this step because the AI is right there and answers are quick — which is exactly why it gets undue weight.
  2. Ask what the AI doesn't know. Every piece of advice depends on context the model can't see: which staff member is having a hard month, which supplier is wobbling, what your customer actually said in the last meeting. Make a quick mental list of what's missing from the conversation. If the missing context could change the answer, the advice isn't ready to act on.
  3. Run a small pre-mortem. Imagine it's three months from now and the advice has gone wrong. What broke? Did you misread the customer? Misjudge the cash flow? Underestimate the cultural fit of a hire? You're not trying to be paranoid. You're trying to surface the failure modes the AI's confident tone has papered over.
  4. Get a sanity-check from someone with skin in the game. A peer, a mentor, a partner, an experienced bookkeeper. Not to outsource the decision, but to triangulate. The AI gives you a plausible answer; the human who knows your business gives you a contextual one; you decide.
  5. Keep a small log of what you acted on and what happened. Not a corporate process. Just a few lines in a notebook or a Notes file. Over six months, this is the only honest signal you'll get about whether the AI's advice is actually good for your business or only sounds good. Without it, every successful month feels like the AI helping and every bad month feels like bad luck.
These are the habits that turn AI from a voice you defer to into a tool you use well. The owners in the Kenya study who came out ahead were doing some version of all five, even if they'd never have written them down as a list.

What this means for the next few years

The cheap-mentor-at-scale story has a strong emotional pull, and I understand why. It promises that the disadvantages of running a small business in a hard place (limited expertise, no consultant on retainer, no MBA network) can be patched with an OpenAI subscription. It would be lovely if that were true. The Kenya field experiment is the most rigorous evidence we've got so far that it isn't — at least not for owners who haven't already done the work of building their own ability to evaluate advice.

The question that matters most for anyone running an SME right now isn’t which AI tool to adopt next. It’s how quickly you can build the judgement to use the tools you already have, before the gap between you and the operators who already have that judgement widens further. The tools will keep getting better at producing answers. The work that matters, and the work nobody can do for you, is getting better at knowing which of those answers are yours to act on.

Quick answers

Does AI help small businesses or hurt them?

It depends on the business owner, not the tool. The 2026 MIT Sloan study found AI mentorship raised profits for entrepreneurs who already had strong business judgement and lowered them for those who didn’t. The aggregate claim that “AI helps SMEs” hides two very different stories.

Why does AI advice work for emails but not for management decisions, and should I stop using it?

Tasks like drafting emails have a recognisable right answer. Management decisions don’t: a plausible recommendation and a good one look identical until weeks later when the consequences arrive. But no, don’t stop using AI for business advice. Use it the way you’d use a smart consultant who’s never met your team. Useful for options and frameworks, not for the final call. The five-step kit in the article above takes about ten minutes to build into your workflow.

What does “judgement gap” mean in this context?

The gap between people who can read a confident-sounding piece of advice and tell whether it fits their specific situation, and people who can’t. AI has made advice cheap without making the ability to evaluate it any cheaper, which is why stronger operators benefit more and the distance between them and everyone else tends to grow rather than shrink.