The Real Divide Isn't Human vs Machine. It's AI Fluency.

Give the exact same AI to two people and watch what happens. One of them asks it to write an email, skims the result, and shrugs. The other one, by the end of the week, has quietly rebuilt how a piece of their business runs. Same tool. Same models. Same monthly bill. The results aren't 10% apart. They aren't close.

I see this every week in my own work, and it's changed how I think about where all of this is heading. The whole public debate is framed as humans against the machines. Will the agents take your job? Can AI replace a marketer, a lawyer, a coder? I think that's the wrong fight to be watching. The line that actually matters isn't between people and AI. It's between people who are fluent with AI and people who aren't.

And "fluent" is carrying a lot of weight in that sentence, so let me be precise about it, and honest about what the evidence does and doesn't say.

Start with the honest part. The controlled studies everyone quotes are real, and they're smaller than the hype. When BCG ran a big experiment with its own consultants, the ones using GPT-4 finished about 12% more tasks and moved around 25% faster. And those tasks were deliberately chosen to sit inside what the model is already good at. A widely cited study of customer-support agents landed near 15% on average. Numbers like that are what you get when you measure ordinary people doing narrow, well-defined work. That's the floor rising, and the floor rising is good news. It just isn't the story.

Here's the part that gets missed. On that kind of bounded work, AI helps the beginner more than the expert. It hands the rookie the veteran's playbook. That same BCG study found its weakest performers gained the most, and the support study's biggest jumps went to the least experienced agents. So when the work has a known right answer, AI is an equalizer. It compresses the gap. If I stopped there, I'd be arguing against my own point.

But fluency was never about bounded work. It lives at the edge, where there's no known right answer, and that edge is stranger than people expect. The researchers who ran that experiment called it the "jagged frontier." Picture an invisible wall around what AI can do, except the wall zigzags. Two tasks that look equally hard sit on opposite sides of it. Ask for a sonnet and you get something lovely. Ask for a poem of exactly fifty words and it botches the count. From the outside you can't tell which is which. On a task the researchers had placed just past that wall, the consultants leaning on AI were about nineteen points more likely to get it wrong. It looked confident, and it wasn't.

Fluency is the tacit skill of feeling where that wall is. It's knowing which problems to hand over and which to guard. It's wiring a handful of tools into a workflow that checks its own output, and overriding a machine that's being confidently, fluently wrong. You don't get that from a subscription. You get it from driving the tools every day until the shape of the frontier lives in your hands.

Which is why the tool alone doesn't deliver. My favorite finding of the year comes from METR, which studied seasoned open-source developers doing real work. With AI tools, they were about 19% slower. And they had no idea. They'd predicted the AI would speed them up by a quarter, and even after finishing slower, they still believed it had helped. The tool was in their hands. The fluency wasn't. That gap, between having access and having skill, is the whole ballgame.

We've watched this movie before, on a smaller screen. In 1979, a program called VisiCalc, the first spreadsheet, shipped for the Apple II and quietly turned a hobbyist gadget into a serious business machine. It didn't replace accountants. It split them. The analysts who got fluent with it could model a whole business live and run a dozen what-if scenarios in an afternoon that used to take weeks. They answered questions in a meeting that nobody else in the room could even attempt. Same software, sitting on the same desks. The distance between the person who mastered it and the person still keeping the ledger by hand wasn't 10%. It was a different job.

That's where the real multiplier hides. Not a faster lap on the same track, but the ability to attempt work that isn't on the table for anyone else. Compound that across every step of how you operate, and you don't get someone 10% more productive. In some kinds of work you get someone fifty or a hundred times more effective, because most of what they're producing simply wouldn't exist otherwise.

People like that are going to be scarce, and they're going to be worth a fortune, because fluency can't be downloaded. The market is already saying so. PwC's 2025 jobs report found that postings asking for AI skills now carry a 56% wage premium, more than double what it was the year before. The industries most exposed to AI are growing revenue per worker about three times faster than the rest. As Karim Lakhani put it in Harvard Business Review, AI won't replace humans, but humans with AI will replace humans without AI.

This is the scarce thing I keep circling back to. When I wrote that I'd rather sell shovels than guess where the gold lands, fluency is the shovel. When I wrote that AI doesn't kill jobs, it just shifts the bottleneck onto judgment, this is the judgment. It's the same idea said three ways, because I think it's the most important thing happening in business right now.

So I'll tell you my actual goal, since an essay like this should cost the writer something. I'm not trying to pick the winning model or the winning startup. I'm trying to become one of these people. The kind who can look at a stuck business and spot a use for AI nobody else saw. The kind who can orchestrate the agents to go build it, and knows when the machine is lying to me with a straight face. That's the skill I'm compounding on purpose, every single day.

The technology showed up for everyone on the same morning. It's one browser tab away for you and for the person who's about to out-produce you fifty to one. The machine was never the divide. What you learn to do with it is the only variable left.