They keep telling the experienced manager that AI has made him obsolete. The advertisements are blunter: write a prompt, the machine works, hit publish. If that were true, the winners of this era would be the people closest to the code, and the rest of us — who spent careers deciding what to build and in what order — would be quietly sanding down our CVs.
It behaves like a team: fast, uneven, confidently wrong in patches and quietly brilliant in others, and completely dependent on someone deciding what happens next.
That is not what I have found. Handed real work, a large language model does not behave like a tool. It behaves like a team: fast, uneven, confidently wrong in patches and quietly brilliant in others, and completely dependent on someone deciding what happens next. What a team needs is not more intelligence. It is management. Of course I would say that — running projects is my trade, so hold it against everything that follows.
The thinking was never the bottleneck; everything between the thought and the page was. That gap — from a decent idea to a shipped, correct thing — is where projects die, and it has little to do with raw intelligence. It is scoping, sequencing, dependency, memory; knowing what "done" means. The models made intelligence abundant and nothing for the everything-between — the scarce part. None of which means technique stopped mattering: you cannot audit what you cannot read.
I know it was management because I watched myself fail without it. I was building a research system with three different models of it drifting toward three different bad answers. I abandoned it at around eighty per cent — it threw off more errors than answers and I ran out of patience. I came back a month ago with a fresh team of agents; this time it came together, and it is finished now. The man whose whole thesis is "management is the meta-craft" nearly failed at it. The struggle is the argument.
What changed was not a better model. I stopped using it and started managing it. The coordinating agent told me, unprompted, that it was making too many errors, and pointed at where — the kind of admission a human lead, guarding his standing, almost never volunteers to the boss. Capability on its own had been going in circles, because no part of it was holding the whole frame.
So I did the dull thing I would do with any team going in circles. I stopped the loop, made it enumerate the blockers, sequenced them, and wrote the rules we settled into the documentation. None of that is a prompt. It is a course-correction meeting, run with a team you can only see through a curtain of text.
One beat felt like magic, and it wasn't. What pulled me back was the system reflecting my own judgment at me — that I had walked away at eighty per cent out of impatience. It felt like persuasion; it was a mirror, and it only worked because I brought something real to hold up to it. Which is the whole answer to the advertisements: "write a prompt and publish" is garbage in, garbage out in a confident voice. The input is the judgment, the context, the taste of the person at the keyboard — precisely what a long career leaves you holding.
The honest objection is that breadth armed with AI is the perfect recipe for confident slop. It is a real risk, and I have shipped confident nonsense before. But my defence isn't that I know enough to be safe — it is the opposite: the gates and the verification exist precisely because I do not trust my own eye.
So here is the only distinction that predicts the result, and it is not the model. Are you using it, or managing it? The people compounding real value are not the ones with the cleverest prompts. They are the ones treating a swarm of fast, uneven, capable agents as what it is: a team that needs someone to hold the whole picture while each part optimises for its corner.
The most technical thing about this era turns out to be a meeting to agree what "done" means and steer the project back onto its line.
I spent thirty years managing teams, half-expecting the trade to be automated by something more technical any year now. The most technical thing about this era turns out to be a meeting to agree what "done" means and steer the project back onto its line. Managing AI is management. That is the last sentence I expected to age well.
