The CEO should have been the first role AI replaced. Not the writer, not the call center agent, not the junior analyst churning out decks at 2am. The CEO. And history will remember our failure to do so as one of the most consequential misreads of this era.
This sounds like a provocation. It isn’t. It’s a description of where the math works.
Where the math works
Most CEO labor is pattern recognition under uncertainty, capital allocation across imperfect options, and synthesis of information arriving from too many channels at once. That is precisely the work these systems are best at. There is no defensible argument that a frontier model in 2026 cannot read a board pack, weigh three acquisition targets against a strategic plan, and produce a recommendation that is at least the median of what a Fortune 500 chief executive would have produced. The honest answer is that it can — and probably better, faster, and at a price that rounds to zero against the comp packages it would replace.
Compensation at the top is already wildly decoupled from verifiable contribution. The “great man” theory of corporate leadership has always been empirically thin; outcomes ascribed to executive genius are mostly attributable to industry tailwinds, capital structure, timing, and the work of the thousands of people below. Strip the mythology away and what remains is a coordination function — one we now know how to automate.
If you were optimizing purely for decision quality per dollar, the C-suite is where the math works first. That isn’t controversial. It’s arithmetic.
So why didn’t it happen?
Because deployment doesn’t follow value. Deployment follows authority.
AI gets pointed at whoever the deployer has authority over. That’s the only rule that consistently predicts where the technology lands. CEOs have authority over labor. Labor does not have authority over CEOs. Boards theoretically do, but boards are largely a CEO-class social network whose members rotate through each other’s compensation committees with the regularity of a turnstile.
So the technology that was supposed to be the great equalizer was instead aimed exactly where you’d expect a tool controlled by capital to be aimed… down. At the writers, the analysts, the customer service reps, the paralegals, the radiologists. At the coordinated, not the coordinators. At the people who couldn’t say no.
The “misread” wasn’t really a misread. It was the predictable outcome of who held the lever.
This is the part worth sitting with. The most-discussed deployment pattern of the most consequential technology of our lifetime tells you something cleaner than any business school case study ever will… AI was deployed downward to compress labor costs, not upward to discipline capital allocation.
That asymmetry is the entire story. It tells you who the technology was actually built to serve in this period, and it tells you why the productivity gains have flowed where they have.
“But AI CEOs are coming”
The standard rebuttal is that this is just timing. AI is replacing easy roles first; the executive layer is next; the asymmetry will resolve itself. In the limit, every role is automated and the question of who deployed it on whom dissolves.
This is wrong, and it’s wrong in an instructive way.
If every CEO is an AI, the question doesn’t disappear — it relocates. Who owns the AI CEO? Because an AI in the corner office is still answerable to something: shareholders, a founding owner, whoever holds the weights and sets the objective function.
The power doesn’t evaporate. It concentrates one layer up, into a smaller, less visible, less accountable group than the one you started with.
The historical analogue is mechanization in agriculture and manufacturing. Replacing the foreman with a machine didn’t flatten the hierarchy. It moved the rents from foremen to whoever owned the machines. The asymmetry didn’t disappear; it became harder to see and harder to organize against, because there was no longer a human face attached to the decision.
So in a world of AI CEOs, “executive function is automated” is not the same as “executive power is distributed.” If ownership of the executive AI layer is concentrated, which is the default trajectory given current capital structures and compute economics, then automating the C-suite is the more extractive outcome, not the more egalitarian one. The visible class disappears and the invisible one consolidates.
That’s the trap. The narrative of “AI is coming for the CEOs eventually” is doing the same work that “trickle-down” did in an earlier era… it’s a story we tell ourselves to make the asymmetry feel temporary while the structures that lock it in get poured in concrete.
The proposal
Ownership goes back to the workers.
Specifically… any role whose primary function is the coordination of other humans’ work is a candidate for automation, and the productivity gains from that automation flow to the coordinated, not to capital. As a working threshold, any role with four or more direct reports is doing more coordination than craft, and is replaceable. The savings, the difference between what the coordinator was paid and what the coordinating system costs, belong to the people whose work was being coordinated.
This is not a technology proposal. The technology is ready. This is a labor politics proposal, and it has to be argued on those terms.
The obstacles, named honestly
There are three, and pretending they don’t exist is how this kind of argument gets dismissed.
The first is legal. Workers don’t currently own the means of production in any equity sense; they rent their labor for wages. For automation savings to flow to workers rather than shareholders, you need a statutory change — codetermination, mandatory profit-sharing on automation events, worker equity vesting when roles are eliminated by software the firm owns — or a structural one, like cooperatives and ESOPs as the default rather than the exception. Neither is the trajectory we’re on. The current trajectory is the opposite: automation gains accrue to whoever owns the automation, and that’s capital.
You need an actual legal regime change. A normative claim won’t get there.
The second is coordination. The manager’s job doesn’t vanish when the manager does. Someone has to set the objective function the system optimizes against. If the workers collectively set it, you’ve reinvented the cooperative, which works at small scale and struggles at larger scale for well-documented reasons. If a smaller group sets it, you’re back to a hierarchy with fewer rungs and more concentrated power at the top, which is the trap I just described. There is no free lunch here. The cooperative form has to be made to work, or the concentration reasserts itself.
The third is the threshold itself. “Four or more direct reports” is a useful heuristic, but it’s not clean. It catches most middle management, which is correct. It also catches some senior individual contributors — principal engineers, tech leads, firm partners — whose reports are nominal and whose actual contribution is craft. And it misses the most extractive roles in the economy, which often have zero direct reports: PE associates, prop traders, pre-scale founders extracting from a labor pool of contractors. The org-chart shape is not a perfect proxy for automatable rent extraction. It’s a starting point, not an answer.
These are real obstacles. They are not reasons not to do it. They are the design problems anyone serious about doing it has to solve.
The conclusion
The technology is ready. The ownership structure isn’t, and it won’t be unless someone fights for it.
That is the actual answer to the question I started with. CEOs weren’t replaced first because the people who would have had to authorize the replacement were the CEOs. The deployment of AI in this period reflects the power structure of this period… not the value structure, not the productivity structure, not the moral structure. The power structure.
AI got pointed downward because the people holding the lever had authority over the people below them and not over the people above them.
History will note this. It will note that we built a technology capable of automating the most overpaid coordination function in the modern economy and chose, instead, to automate the most underpaid productive ones. It will note that we called this “efficiency.” It will note who got rich.
The interesting question is not whether the misread happened. It clearly did. The interesting question is whether the next decade is the one in which someone successfully argues — legally, politically, structurally — that the gains from automating coordination should flow to the coordinated.
Because if no one argues it, the default outcome is the one we’re already in… the visible hierarchy thins out, the invisible one consolidates, and “AI took our jobs” becomes the cover story for a much older and more familiar transfer of wealth.
The CEO should have been the first role replaced. The fact that it wasn’t is not a technological observation.
It’s a confession.
This is the first piece in a three-part sequence.
Piece two, "We Already Ran the Experiment", argues that the configuration that produced broadly shared prosperity from 1948 to 1973 is recoverable, and names the load-bearing constraints.
Piece three, "The Mechanism Has a Name", documents what stands between the recovery and its implementation.
Stick around.. it gets good.