AI Isn't Taking Your Job. It's Taking the Job of Whoever Can't Tell Good Advice From Bad.
A Kenyan field experiment and a small agency study explain why the same AI tool helps some workers and hurts others — and why closing that gap is a governance job, not a mood.
AI POLICY & GOVERNANCE
AI Isn’t Taking Your Job. It’s Taking the Job of Whoever Can’t Tell Good Advice From Bad.
In 2009, I helped deploy a Palm device at an upcountry hospital in Uganda. The device talked to a wireless access point and pushed patient data somewhere a paper register never could. It was a small piece of kit, barely bigger than a calculator, and it terrified the facility data officer more than any audit ever had.
He’d spent years turning messy handwriting into monthly reports. Now a machine did it in seconds. He asked me directly: does this mean I’m finished here? I gave him a technical answer — throughput, sync intervals, offline caching — when what he actually needed was a governance one. Nobody had planned for him. The device arrived with a rollout schedule. He didn’t get one.
He kept his job. Not because the technology went easy on him, but because someone (you know who) eventually taught him to supervise the tool instead of losing a race to it. Sixteen years and several technology cycles later, that gap — between the tool arriving and the person being equipped to sit above it — is still the whole argument. We’ve just dressed it up in a new fear and given it a new name.
We’ve had this argument before, and we lost the thread the same way
When the car showed up, the confident response wasn’t excitement. It was more horses. Faster horses, stronger horses, better-bred horses — anything to avoid admitting the category itself had changed. The car wasn’t competing with a horse. It was competing with an entire trade built around horses: the farrier, the ostler, the coachman who knew which road flooded in the rains. Most of that trade didn’t survive intact. But driving as a livelihood didn’t vanish — it changed shape, and the people who learned the new shape did fine. The people who kept sharpening horseshoes did not.
I’m not telling that story to be glib about job loss, which is real and often brutal for the people living through it. I’m telling it because “will this replace me” is close to the wrong question, and it was close to the wrong question in 2009, and it was close to the wrong question when the first car rattled past a stable full of people convinced the answer was more horses.
Today, I have a smartwatch and a health tracker bolted onto both my arms. Here’s the live version: do wearables replace the triage nurse? A device that reads your oxygen, your heart rate, your sleep, flags an anomaly before you feel it — does that make the person assessing you at the door redundant? I don’t think it’s far-fetched at all. I think it’s exactly the right question, aimed slightly off target. The wearable doesn’t replace the nurse. It replaces the version of the nurse who only ever reads numbers off a chart and never learns to ask the question the numbers can’t answer. The nurse who asks that question becomes harder to replace, not easier.
That distinction — the worker a tool replaces versus the worker it can’t touch — turned out to be measurable. Two pieces of research published this August put numbers on the mechanism.
The evidence: same tool, opposite outcomes
Columbia Business School’s David Holtz and colleagues ran a field experiment with 640 Kenyan entrepreneurs — farms, restaurants, clothing shops, internet cafés, businesses ranging from struggling to thriving. Half got a GPT-4-powered WhatsApp business assistant. Half got static ILO training guides. Everyone was tracked on revenue and profit for roughly two and a half months. Full write-up here: business.columbia.edu/insights/ai-widening-gap-between-high-and-low-performers.
On average, the AI assistant moved nothing. That headline number is the least interesting thing in the study. Entrepreneurs who were already performing well pulled further ahead. Entrepreneurs who were struggling did measurably worse — nearly 10% worse than the control group, against gains north of 15% for the top performers. Same tool. Opposite direction. That’s not a rounding error, that’s a fork in the road with a sign nobody read.
The gap had nothing to do with who asked cleverer questions. Holtz’s team found the AI gave everyone roughly the same advice. What separated the two groups was whether the entrepreneur could tell useful advice from generic advice, and whether they knew how to apply it to their actual circumstances. Struggling owners followed generic recommendations — cut prices, spend more on ads — that quietly bled their margins. Thriving owners adapted the same advice and pulled ahead. Holtz’s own conclusion, stated plainly: “You need expertise as a complement to AI access.”
That’s the economic half. The behavioural half comes from a smaller study, and I want to be upfront about its limits before I use it: a Neuravox Policy Lab survey of 76 experienced AI users — exploratory, self-selected, not representative of a population. What it isolates is still useful: not how much people use AI, but how.
The finding worth keeping: frequency of AI use had almost no relationship with whether people delegated their thinking to it, or whether their independent judgement eroded. What mattered was the pattern underneath the frequency — did they question the AI when it conflicted with their own judgement, could they explain the finished work in their own words, did they retain the skill to do the task without the tool. People who held onto those three things could delegate heavily and keep their agency intact. People who let them slide lost ground even with light use.
Put the two studies together and the picture sharpens fast. Columbia shows the outcome: identical tool, opposite trajectories. Neuravox shows the mechanism underneath it: the difference was never exposure to AI, it was whether the person kept verifying, explaining, and deciding for themselves while using it.
What the myth gets wrong
The fear says AI replaces the worker. The evidence says something sharper: AI replaces the worker who has no scaffold for judgement, while simultaneously handing the advantage to whoever already does. Same tool, same moment, two directions at once — deprecating one person and developing another off the identical input. That’s the part the myth flattens into a single, simpler, wronger story.
AI’s rising tide will not lift all boats. Some of those boats don’t have a hull built to catch it.
My data officer in 2009 wasn’t at risk because a Palm device could type faster than he could write. He was at risk because nobody built him the scaffold — technology enabled data services training, a new mandate, a coach — that would let him move from producing the data to supervising it. Once someone did, the risk mostly disappeared.
I’ve watched the same pattern at bigger scale from the deployment side. Guild Digital’s early work on Uganda’s rCHMIS system taught me this from the other direction: hand a health worker a new digital tool with no human layer of training and support around it, and you don’t get productivity — you get exactly the bifurcation Columbia measured - even if that tool is a leading open source digital public good. Your strongest staff pull ahead. Your weakest fall further behind. The average tells you nothing, and if you’re only reading the average, you’ll miss the story entirely.
This is a governance problem, not a mood
If the gap runs on complement rather than access, closing it is a design question. Design questions can have answers.
The frameworks that should be doing this work already exist in principle. UNESCO’s Recommendation on the Ethics of Artificial Intelligence insists AI cannot displace human responsibility for decisions. The OECD’s AI Principles put human agency and oversight at the centre of the human-rights pillar. The African Union’s Continental AI Strategy is explicitly people-centred, tying responsible AI to dignity and development rather than efficiency for its own sake. None of these documents are wrong. What they lack is an operating layer — a way to turn “human oversight matters” into something an employer, a regulator, or a national strategy can actually check.
Rwanda is worth pointing to here, not as a finished product but as evidence the operating layer is buildable. Its National AI Policy was cabinet-approved in 2023, with a costed five-year implementation plan of roughly $76.5 million — not a communiqué, a budget line. By the end of 2025, over 5,000 teachers across every district had been trained in foundational AI literacy through a national programme run with Rwanda’s Ministry of Education, the Rwanda Basic Education Board, and MIT RAISE. That’s the coaching layer Holtz’s study says closes the gap, already running at national scale, in the region. It’s not the only model worth studying, and it’s not finished. But it’s a country that decided the operating layer was worth the line item, and that’s the more useful example than pointing at a strategy still in draft.
That operating layer, wherever it’s built, tends to look like four things:
Pair access with coaching, not just distribution. Holtz’s own conclusion from the Kenya study is that AI paired with human coaching closes the gap that AI alone widens — and it’s cheap relative to traditional advisory programmes. A national AI strategy that measures success by device count or licence count alone is measuring the wrong thing.
Build verification and explanation into workplace AI governance. Not a compliance checkbox — a retained-skill requirement. Can a staff member explain AI-assisted work in their own terms. That single question, asked consistently, protects a workforce more than any usage cap ever will.
Run capability retention audits. Ask people to do the task without the tool, periodically. Not to slow them down — to catch a skill atrophying quietly, before a system failure or an edge case forces the discovery at the worst possible time.
Map decision authority explicitly. Name, in writing, which decisions AI can shape and which require a named human sign-off. Ambiguity here — not the tool itself — is where displacement actually creeps in.
Where this leaves the fear
I don’t think the fear is irrational. I think it’s aimed at the wrong target. AI isn’t taking jobs the way a flood takes a village, indiscriminately, all at once. It’s taking the jobs of whoever nobody built a bridge for — and it’s cementing the position of whoever already had one.
And the bridges being built right now aren’t only in San Francisco or London. A Kenyan smallholder pointing a phone camera at a cassava leaf gets an offline diagnosis of viral disease in seconds, from an app called PlantVillage Nuru that runs its deep-learning model without an internet connection, because the researchers who built it understood that rural connectivity was the constraint that mattered. A newborn’s cry, recorded on a smartphone at the University of Port Harcourt Teaching Hospital, gets analysed by Ubenwa for early signs of birth asphyxia — a diagnosis that otherwise needs a $20,000 blood gas analyzer most Nigerian birthing centres will never own. And in Kenya, Apollo Agriculture combines satellite imagery, machine learning, and mobile money to score the creditworthiness of smallholder farmers who have never held a bank-issued credit file, and it’s now served tens of thousands of them.
None of those are demo-day stories built for a stage in the global north. They’re a farmer, a newborn, and a smallholder’s loan application — ordinary problems, solved by AI built with the actual constraint in mind, not despite it. That’s what the operating layer produces when someone bothers to build it.
My data officer in 2009 didn’t need me to tell him the Palm device was harmless. He needed someone to build the bridge before the tool arrived, not after. That’s still the job, sixteen years and considerably more capable machines later. It’s just no longer mine alone to do quietly in one district hospital. It’s a governance job now — and it’s the one worth doing properly this time, not just for the sake of the workers standing on the wrong side of the gap, but because the boats already sailing prove it can be done.
← My Written Thoughts