In 2001 Marc Prensky gave us “digital native”. A generation raised with computers would think about them differently, he argued, and the rest of us would always be immigrants speaking with an accent. It was intuitive and quotable, so organisations built real decisions on it. Hiring assumptions and training budgets moved with it.
Bennett, Maton and Kervin reviewed the evidence in 2008 and the category didn’t survive contact with the data. Kirschner and De Bruyckere went back to it in 2017 and landed in the same place. Young people had more exposure to technology and no more capability with it. Fluency with an interface told you almost nothing about whether someone understood what sat behind it.
The same claim is being made again with AI in place of computers, and it’ll be wrong for the same reason.
Experimenting at Charlie Oscar
Earlier this year every person at Charlie Oscar was given an AI licence and training on how to use it. Access was complete, agency wide and had no exceptions. In any given week about seven of us used it. Forty people, and roughly one in six active.
The handful of jobs running autonomously in the business all belonged to one team, which happened to be mine. The people who did use it treated it as an answer engine: they asked it for things and didn’t give it work. That distinction turned out to be the whole problem, and no amount of additional access was going to close the gap.
What changed?
We stopped treating adoption as a distribution problem. We put AI inside Slack and the other tools people were using and began to think of our AI operating system as a thin interface connecting those tools so that, day to day, nobody had to learn a new product. We recruited one named champion per team by open call, and we built training around a single skill: notice what you repeat, then break it into steps something else can follow. All of it, infrastructure included, was two months of work by two people.
A month later 97% of the agency was active weekly. That sounds impressive, but it’s the easiest number to move. The one that matters is that 26% of the agency now owns a job that runs without them.
Ninety-seven per cent are using it. Twenty-six per cent are delegating to it. Exposure reached near-total inside a month because exposure is cheap to change: put AI in the places people already work. Capability reached a quarter because it’s a different thing and it moves at a different rate.
The number everyone reports
Weekly active users is the number nearly all AI programmes report. It tells you how many people opened it.
I would say that the reason it’s everywhere is that it’s the metric you can move without pressure. Put AI into performance reviews, run a leaderboard, hand out unlimited tokens, and usage goes up, because people are now being paid to make it go up. Whether anyone got better at anything is a different question and it isn’t the one being asked.
The survey data suggests the two come apart. ManpowerGroup found regular AI use rose 13 points last year while workers’ confidence in using the technology fell 18. More than a third of employees told WalkMe they skip AI on tasks where it would slow them down. And in a survey of 5,000 white-collar workers, over 40% of executives said AI saved them eight hours a week or more, against two-thirds of everyone else reporting under two. The people reporting the biggest gains are the ones furthest from the work.
One of the companies that went hardest at this, with AI usage written into performance reviews, dropped the tracking about a year later. The reason given was that it had become a policy about using a tool rather than about outcomes.
Which brings me back to native
“AI native” implied capability arrived by immersion: be around it long enough and it seeps in. But being fluent in a chat window teaches you nothing about how to separate your job from your role, or how to break that job into modular parts a system can run.
There will be AI natives eventually. They’re in primary school at the moment, so they’re no use as a hiring strategy for 2026.
AI first is a different proposition and you can act on it today. Pick an output the business already values, then work backwards to the strategy that gets you there. The question “what could AI do for us” reliably produces demos. The one that produces good work is “what do we already need done every week, and what would it take to hand that over”.
It only works as a loop, because the capability frontier moves faster than any strategy cycle. Work that wasn’t automatable in March is automatable now, so working backwards once leaves you with a plan with a short shelf life.
So, who does it?
None of this is technology work. Someone has to decide which output matters, break the job into steps, own the thing afterwards and answer for what happens the first time it gets something wrong. That’s operations and product work. I have an obvious stake in saying that, given my job title, so take it with the appropriate pinch of salt.
It also explains why AI capability keeps failing at the seams between functions. The work gets handed to IT or innovation, or tech, and it dies there, not because those teams don’t understand AI but because it’s not their job to reach into how every other team works. The problem is organisational and it keeps getting scoped as technical.
Charlie Oscar’s AI Ops team is two people. What it actually runs is an engineering loop: we define the output, the agents build the infrastructure, and we review the code and merge it so people can use the tools. That’s how two people put eight weeks of tooling in place. It’s also the argument applied to ourselves. Pick the output, work backwards, hand over whatever can be handed over.
Building the infrastructure and getting people using it is the fast part. Both are delegable, which is how two of us managed it in eight weeks.
Capability is the part you can’t delegate, and that’s the whole difficulty. It’s an awkward position for a team whose entire job is making work handoverable. I can build the tooling for forty people. I can’t be capable on their behalf, and there’s nobody I can hire to do the learning for them.
The 26% aren’t better at AI than everyone else. They’re the ones who handed over a real job once and found out what it takes. If capability comes from doing rather than from being told, the difference isn’t aptitude or the quality of the training, it’s whether someone has done it yet. You don’t teach your way to that. You engineer the first handover, one person at a time, and it’s slow work with no launch at the end of it.
It also means you can’t recruit your way out. Someone joining next year will arrive with plenty of exposure and no capability in your context, which is where this started. Prensky’s mistake was assuming the second follows from the first. It doesn’t, and it never has.
