The real reason AI favors generalists

The common wisdom is that AI favors generalists. I agree. But the reason usually given for it is wrong, and the wrong reason hides the interesting part.

The usual story goes: specialists are obsolete, because you can hand their work to an agent. I don’t buy it. Someone still has to guide the agent and check what it hands back, and that takes real expertise, usually built over years. The specialist isn’t going away. What’s changing is who has to have expert judgment.

I keep watching less-technical friends start a technical project with an agent and then stall. They get a working prototype, something they could never have built alone a year ago. Then making it real asks for skills they don’t have and can’t fake with AI: seeing that the architecture won’t survive its tenth user, or that the screen which makes sense to them confuses everyone else. What’s missing isn’t effort. It’s the judgment a colleague one desk over would have supplied for free if this were a job instead of a side project.

That gap used to be survivable.

Specialization was a bet on slow building

Specialization was always an economic bet, and the bet was that building was the slow part. When writing the code took weeks, handing off the judgment cost you almost nothing. A strong coder could get an hour of architecture advice from a senior and then go heads-down for a month. A frontend engineer leaned on a designer for the look and feel. A product engineer took the idea from a PM or the CEO and built it. You borrowed the taste you didn’t have, and against a month of implementation, the day you waited on that borrowed taste barely registered. And it didn’t even matter that you’d go back to that expert as the project evolved or after user feedback came in.

Whole companies were built around this. A cross-functional team is a machine for gathering every skill a project needs in one place, so no single person has to have all of them. Big companies got good at estimating the mix: how many engineers, how many designers per engineer, how much of a PM’s week, how much research. Staffing a project was mostly a question of ratios. And for a lot of people, the rote building was the job. A junior engineer shipped code a tech lead had already scoped and shaped. The taste came from elsewhere in the org, and that worked even in smaller companies once the giants had proven the model.

AI voided the deal

AI made building the fast part. And once building became fast, the speed of every other part of the process started to matter, and could even become the gating factor. The speed of code reviews, the turnaround from design, how quickly the PM triaged customer issues, the product owner’s decision on direction, all of it became a larger share of the total build time.

A day waiting on the designer, a couple of rounds of back-and-forth with the PM, used to disappear against weeks of implementation. Now that same day may be slower than the work it’s meant to direct. The agent produces the thing in an afternoon and then waits for someone to notice the interface is wrong. Coordinating taste across people, the thing companies spent decades getting efficient at, is now the bottleneck it was invented to remove. So the skill and the taste ideally come from the same person. Whoever guides the agent also has to be the one who can tell whether what came back is any good.

Take a small agency. The old way to run a project was to staff it: a strategist, a designer, a copywriter, someone to build it. The work was only as good as the team, and the team covered each other’s blind spots. Now one generalist can run the whole thing with AI doing the production. Faster and cheaper, until you notice the output is only as good as that one person’s weakest sense. If they can’t see that the design is generic or the copy falls flat, nothing else in the loop will catch it. The AI won’t. The team that used to catch it isn’t there. A blind spot that used to belong to one person now brings down the whole project.

The same force, pointed inward

I wrote a while back about how the first solo-founder unicorn gets built, and this is that same argument aimed one level down. That piece is about the boundary of the company: coordination costs falling until the edge of the firm moves toward a single person. This is the same observation applied to the team. When coordinating between people is cheap, you split the work across specialists. When it gets expensive relative to the work itself, the boundary moves and the roles fold together. The lines between them blur. Teams get smaller. It’s Coase run inside the org chart instead of around it.

What this isn’t

I want to be careful here, because it’s easy to read this as a layoff prediction, and it isn’t one.

Some rote roles change. The job that was mostly implementing someone else’s judgment is the one AI reaches first, and that’s real. But the frame isn’t fewer people. The building was never the valuable part. Choosing the architecture, designing the thing so people actually understand it, finding the feature worth building at all: that was always the work. When the rote part gets cheap, the time doesn’t vanish. It moves to the judgment, which is where the value sat the whole time. I’ve made the broader version of this argument elsewhere: fit-for-purpose AI doesn’t shrink the work, it moves people up into the part that was the point. Quality rises, along with people’s expectations. A web-only app isn’t enough anymore; mobile is a requirement. An internal tool has to be as fast and polished as the customer-facing one. The product has to be usable on day one, not after a week of training.

The part worth keeping

Writing code is slow and mostly rote. Choosing the architecture, designing the interface, finding the feature worth building, that’s the creative part, and it’s the part I’d keep if I could only keep one. Delegate the implementation and the day you get back goes straight to it. That’s what AI favoring generalists actually means. It doesn’t mean expertise stopped mattering. It means the same person finally gets to do both the aiming and the building, and the aiming was always the good part.

Key takeaways

• The usual case for generalists (specialists are obsolete because you can hand their work to AI) is wrong. Guiding an agent and judging its output still takes years of expertise.
• Specialization was an economic bet that building was slow, so splitting the skill to build from the taste to aim it across two people was nearly free.
• AI made building fast, which turns the handoff into the bottleneck. The same person now needs both the skill to guide the agent and the taste to aim it.
• It’s the coordination-cost argument behind the one-person company, applied to the inside of the team instead of the boundary of the firm.
• This isn’t a layoff prediction. The rote building was never the valuable part, so the freed time moves to the judgment work that always carried the value.