AI-assisted execution: why four people is enough.
The cost of building collapsed in 2024. Most studios haven’t restructured around it — they hired around it. We did the opposite. The math, the stack, and the second-order effects on team design.
The number that changed
In 2021, building a B2C product to first revenue typically took twelve to twenty people across product, engineering, design, and ops. We know this number because we used to plan around it. The studios we benchmarked ourselves against had headcounts in that range and shipped on a timeline that matched.
By the end of 2024, that ratio was already broken. By 2026, it is unrecognizable. A small team with the right stack ships products that look indistinguishable from the output of those twenty-person shops. We have run the experiment in our own building.
The numbers, roughly, in our case:
The four people are not super-engineers. They are competent operators using a stack that did not exist three years ago: heavy AI tooling for code generation, design generation, copy, research, analysis, ops automation, and customer ops. Every role has a force-multiplier attached.
What this does to the model
The implication that matters: if four operators ship what twenty shipped, and you keep twenty operators, your cost structure is five times what it needs to be. Returns on every project drop accordingly. Most studios are still doing this and calling it “capacity for scale.”
A studio that did not shrink in the last twenty-four months either did not see the curve, or saw it and could not act on it. Either one is bad news for whoever is funding it.
What it does to the work
Less obvious, but more important. A four-person team ships differently from a twenty-person team. Not just faster — differently.
The decision graph is shorter. Nobody is briefing anybody. The person writing the code talks to the person running the marketing test. There is no roadmap document because the roadmap is the four of us in a room. Mistakes get caught earlier because the whole team sees the same telemetry.
The dark side is real. Four people cannot run twenty projects. Four people can run three to five well, if the projects are properly staged. We treat that as a feature: it forces the discipline to kill projects that aren’t earning their slot.
What about specialization?
The standard objection: at twenty people, you have a real designer, a real ML engineer, a real growth lead, a real legal person. At four, you don’t. True. But the AI stack covers most of the gap, and where it doesn’t, we contract for a week and move on. The fixed cost of specialization disappears.
The trade is acceptable because most specialization in 2021 was itself an artifact of cheap money. When capital was free, hiring a senior X was the right move. When capital is expensive, renting one for a week is.
The constraint that makes the rest work
We don’t grow into ten. We don’t hire an “AI person” and an “ops person” once we’re busy. We stay at four. The constraint is what keeps the cost structure honest, the decisions fast, and the off-switch real. Removing it breaks the rest of the model.
You don’t need permission for any of this. The stack is public. The math is public. We are showing it because in three years it will be obvious, and in the meantime there is a window.
← all essays and notes