Everyone assumed the young people would lead. Junior, fluent, fewer habits to unlearn, the digital natives who grew up with this stuff. That assumption is baked into how most companies think about AI: hire young, train fast, let them show everyone how it's done.
Then KPMG and researchers from UT Austin's McCombs School of Business actually measured it. They spent eight months watching about 2,500 employees work, across 1.4 million real AI interactions. Almost everyone was using AI, around 90% of staff. Only about 5% were using it well. One in twenty. And the best of them weren't the juniors. They skewed older, often above the manager level: the senior managers, directors, and partners. The researchers called it surprising, because it cut against the idea that being comfortable with a tool is the same as being good with it.
The easy read is "hire seniors." I think that's the wrong lesson, and the study's own data shows why. Nobody wins with AI because of a title or a birth year. They win because of how they show up to it: curious, patient, willing to get their hands dirty and keep adjusting. That's a way of working, and a sharp junior can have it while a senior misses it completely.
And I don't just mean the senior who has checked out. I mean the one who figures they already know enough. The one who keeps AI at arm's length because it feels like a threat, or a toy. Sitting at a senior desk and not being asleep at it are not the same as being good at this. The people who pull ahead are the ones who stay curious enough to keep learning, at any age.
Comfort and sophistication are different skills
Comfort is moving quickly through a tool. Sophistication is knowing what to ask it to do. From the outside they look the same, which is why people mix them up.
A junior who grew up with digital tools is genuinely fluent with AI. They'll try things, iterate fast, and get decent results on the easy stuff. That fluency is real. It's just the floor, not the ceiling. The KPMG researchers found the same gap and were blunt about it: there's "an important distinction between comfort and sophistication," and the people clearing that bar were doing something different.
What they did was treat AI as a thinking partner instead of an answer machine. They gave it a role to play, showed it an example of what a good result looked like, walked it through how to reason, made it explain itself, and pushed back when the first answer fell short. Instead of asking "write me a project update," they'd say something closer to: "You're a project manager briefing a nervous client. Here's last week's update as a model. Draft this week's, keep it under 200 words, and flag any risk that slipped." Same tool. Completely different ask. They brought it their hard, multi-step problems with real constraints and a clear finish line, not just their easy ones. And they picked the model to fit the task instead of always reaching for the same one.
None of that requires comfort with technology. It requires knowing what you're trying to get done well enough to steer someone else toward it, and enough know-how to spot when the answer is wrong. That's problem framing. You build it by solving hard problems and living with the results, and it carries straight over to AI.
What the usage numbers hide
KPMG had nearly 90% of its workforce on AI before the study even started, and only about 5% of them were sophisticated by any useful measure. How often someone used AI told you almost nothing about how well they used it.
When the researchers dug into why frequency was such a weak signal, they found something specific. Junior employees were more likely to point the company's AI tools at personal tasks, the kind of thing that never shows up in the work the company actually measures. The usage charts went up. The real output didn't follow. As the paper put it, "frequency of use may not be a reliable signal for productivity."
If your team's AI dashboard shows heavy use, that number is probably right. It's just not measuring what you think. The most active users at KPMG weren't the most skilled ones. The skilled ones showed up through a different set of habits entirely.
Staying in the work
Here's what that looks like in practice, with me as the guinea pig.
I lean on AI all day, in two main places: Cursor for my coding and project work, and Gemini for writing and thinking out loud. Cursor has an auto-approve setting that lets the AI make changes without stopping to check with me. A few people have given me a hard time for leaving it off. I leave it off on purpose, because I read what the AI does as it does it.
That's slower. It's also how I catch the wrong turn at step 3 before it snowballs into an hour of cleanup at step 40. Reading along the way isn't wasted time. It's the work. When the AI makes a move I don't follow, or a call I wouldn't have made, I stop and dig into why. Sometimes it's right and I learn something. Sometimes it's wrong and I redirect. Either way I stay in the loop, and what I hand off at the end is something I actually understand.
That's what the sophisticated users in the study were doing across 1.4 million interactions. They stayed present. They didn't fire off a request and come back at the end to see what showed up.
The opposite failure shows up in a question I get a lot, and not just from juniors. Plenty of senior people, some well above my pay grade, ask it too: "what's the prompt I use for this?" They want the magic words, the recipe to paste in. They're treating AI like a vending machine: put in a prompt, get out an answer. That lands them in the same place as fire-and-forget, because they're still not in the work. They're trying to skip it.
My average Gemini prompt runs 498 words. That's not a recipe. It's a brief. Context, constraints, what I don't want, and what a good answer looks like. If I want help with a hard client email, I don't type "write an email about the delay." I tell it who the client is, what went wrong, what we're doing about it, the tone I want, and what I'm scared of it sounding like. That habit came from years of handing people vague instructions, getting vague work back, and finally learning to stop. Experience taught me that. But experience isn't the only teacher. Paying attention to how things turn out, and being willing to pick them apart, teaches the same lesson a lot faster than waiting twenty years for it to arrive on its own.
The researchers landed in the same place. The difference, they found, wasn't really in the prompts themselves. It was in the pattern over time: sticking with a problem, guiding the reasoning, refining instead of just accepting. You can teach that. You can't fake it by typing more often.
So what's the actual variable?
The question everyone reaches for is seniority. Do we hire experienced people or hungry newcomers? Wrong frame. Yes, the senior folks at KPMG came out ahead. But look at why the researchers say they did: people below manager level were "less likely to use a deliberate strategy" and less likely to give the model clear requests.
Here's my read on the rest of it. The directors and partners who shine with AI are often the same people who already know how to run a team. They can brief someone clearly, hand off a task, check the work, redirect when it drifts, and coach the person who's stuck. Directing a model is that exact muscle, just pointed at software instead of a person. They've spent years learning how to get good work out of other people, and that skill transfers almost perfectly. None of which is locked to a title. There are juniors with that instinct already, and in a world without AI they'd have grown into strong managers anyway. The tool just lets them show it sooner.
That's the part I'd underline. Among the skilled 5%, the study found "wide variation in roles and seniority." And the behaviors driving all of it are, in the researchers' own words, "observable, teachable, and scalable." Seniority is a stand-in. The behavior is the real cause. And the behavior can be taught.
So the real variable has nothing to do with anyone's résumé. It's whether a person gets into the work or tries to skip past it. The fire-and-forget user gets weak results at every level. So does the person hunting for a recipe instead of learning the process. Both are betting that AI will do the work for them. The person who reads along, pushes back when the answer is wrong, and steers instead of shrugging will get better results, and will get better at getting them, fast. That holds no matter what the title says.
How to actually get better at this
The good news: these are habits, not talents. You don't need a new title or anyone's permission to start. A few that have moved the needle for me, and that line up with what the study found:
- Write the brief before the prompt. Get clear on what you actually want first. Who it's for, what the goal is, what good looks like, what to avoid. If you can't explain the task to a smart stranger, the AI doesn't stand a chance.
- Give it a role and an example. "Act as a skeptical CFO reviewing this" beats "review this." One real sample of the output you want is worth a paragraph of describing it.
- Never treat the first answer as the final one. Push back. Ask what's missing. Ask it to argue the other side. The second and third passes are where the quality lives.
- Read along, don't walk away. Stay in the loop while it works, so you catch the wrong turn early, when it's still cheap to fix.
- Match the tool to the job. Don't reach for the same model out of habit. Some are better at code, some at writing, some at thinking through a mess.
None of that takes ten years on the job. It takes a willingness to slow down and treat AI like a teammate you're managing, not a button you press.
The part everyone's scared of
There's a lot of fear going around that juniors are finished, that AI raises the bar so high a newcomer can't clear it. I don't buy it. What I'd tell a junior is this: the thing that sets you apart was never going to be your years on the job, because you don't have them yet. It's whether you stay curious, stay in the work, and keep getting better at directing it. That's available to you today.
And to the senior reading this: your experience is real, and it matters. Years in the seat build the judgment that tells you when an answer smells wrong. But that judgment only pays off if you bring it to the tool instead of holding it above the tool. Experience plus curiosity is where the gold is. Experience that refuses to engage just gets lapped.
Comfort with the tool is the floor. Sophistication is staying in the work long enough to direct it well. That's not something you read off a résumé. It's something you see in how a person shows up.