A developer sits on a train, watching a coding agent work through a task on her phone. She reads what it's about to do, taps approve, then changes her mind — no, not that file, this one — and taps again. She hasn't written a line of it. Her job is to decide whether it's right, and steer it when it isn't. When OpenAI put Codex inside the ChatGPT mobile app in May, the headline was that you could now run it from your pocket. The part worth noticing was what you were doing while you ran it. You were triaging.

It's tempting to file this under developer tooling and move on, because most of us don't ship code. But the shape of that afternoon is coming for the rest of us, and in a lot of jobs it has already arrived. The doing gets handed off. What's left for you is the approving, the redirecting, the auditing of something else's work. Your job becomes triage.

This is the biggest change to knowledge work in a generation, and we're describing it far too narrowly. We keep talking about AI as a productivity tool, something that helps you do your work faster. That framing is already out of date. AI isn't speeding up your work — it's changing what your work is. The person who used to do the task now decides whether the task was done well. That's a different job, with different skills, different satisfactions, and a different way of going wrong. And oddly, it's a job somebody described in detail forty years ago.

The oldest advice in small business

If you've read Michael Gerber, this will sound familiar, because it's the oldest advice in small business. Gerber's whole argument in The E-Myth was that owners get trapped doing the work of their business — baking the bread, doing the consulting, writing the copy — when their real job is to work on the business rather than in it. Build the system. Let the system do the doing. Spend your time deciding whether the system is still doing the right thing.

For forty years that was advice you could ignore. Plenty of owners did, and their businesses ran fine, if smaller than they might have been. The reframe was true but easy to defer. What's changed is that AI has taken the choice away, and pushed the same shift down from the owner to everyone. The founder used to be the only one who stepped back from the doing. Now it's the analyst, the marketer, the paralegal, the support lead as well. Everyone runs a small operation of things that do the doing, and everyone's job is quietly becoming Gerber's job.

I wrote a whole book chasing that idea to where it leads — The E-Myth, Revisited Again, free if you want the long version. This piece is about one part of it: what happens to the doing.

Good triage needs judgment you earned by doing

This is the uncomfortable part, and I'm not sure how it resolves. Triage sounds like the easy end of the deal — read the output, approve or reject, redirect if it's off. But approving well is hard, and it rests on something invisible in the moment: a sense of what good actually looks like. That sense almost always came from having done the work yourself. You know a weak contract when you see one because you've written a hundred of them. You spot the marketing copy that's competent but has no edge because you spent years learning where the edge is.

Which raises a question a VentureBeat piece framed well in May, calling it the enterprise risk nobody is modelling. If the whole workforce shifts to triage and nobody executes anymore, where does the next generation's judgment come from? The junior analyst used to build the model badly, get it torn apart, and slowly learn what right felt like. If the model now arrives finished and the junior's job is to wave it through, they never build the instinct that waving it through is supposed to require. We hollow out the very expertise the system depends on us having. The people best at triage today are the ones who spent years executing — and we're removing the path that produced them.

There's a related worry I've written about before, that heavy AI users lose the thinking habits that made them valuable in the first place. This is the same worry widened from one person's head to the whole shape of a job — less about losing your edge, more about a role being redrawn underneath you.

Holding the shift on purpose

None of this is an argument against the change. It's happening whether we like it or not, and most of it is good — the leverage is real, and freeing people from rote execution is worth celebrating. It's an argument for holding the shift deliberately instead of sliding into it. A few things seem to matter.

Keep your hand in. Not on everything, and not out of nostalgia, but enough to keep your judgment calibrated. There's a version of this that goes wrong in the other direction — the founder who reviews AI-written copy all day hasn't escaped the technician trap, they've just moved the bottleneck from doing to approving. The aim is to do enough of the real work that you can still tell good from good-enough.

Take triage seriously as a skill, because rubber-stamping isn't the same thing. Real triage means knowing what you're looking for, catching the plausible-but-wrong, and having the confidence to send something back. That's harder than doing the work, not easier, and we should stop assuming the operator's chair is the relaxing one.

And if you run a team, protect the apprenticeship. Handing all the junior work to AI and leaving the juniors to approve it will look efficient on the org chart. It also stops them ever becoming the seniors who can judge the work. Someone has to keep their hands dirty long enough to learn what the dirt teaches.

Gerber's promise was that if you built the machine well enough, you could walk away from it. The machine got better than he imagined, and closer, and the promise turned out to hide a catch he never had to worry about: the better it gets at doing the work, the more carefully we must guard the one thing it still can't do without us, which is know whether the work was worth doing at all.