How AI Actually Takes Over
People imagine AI taking control of the world by force. Yuval Harari makes a more unsettling claim: it will take over by running the bureaucracy. Power has never really flowed from weapons; it flows from the systems that decide who gets the loan, the permit, the job, the verdict. AI is the first technology that can make those decisions itself. It doesn’t need an army. It needs a desk.
That sounds like a prophecy about governments, decades out. It isn’t. I watch it play out in our company, in any given week. Not as science fiction. As workflow.
Nobody hands over control
Here is what a takeover actually looks like. Nobody decides to surrender authority to AI. No meeting is held. No memo goes out. People just hand over the boring parts.
The developer lets AI write the first version of the code. The product manager lets it draft the spec. The marketer lets it draft the campaign. The manager lets it draft the performance review, the budget commentary, the hiring scorecard. Each handoff is small, rational, defensible. It’s the tedious part of the job. Why wouldn’t you delegate it?
Because of an old truth about organisations: the person who writes the first draft holds the power.
Bureaucrats have always known this. The minister announces the decision, but the staffer who wrote the briefing shaped it. The committee votes, but the clerk who drafted the agenda decided what could be voted on. The first draft frames the problem, sets the options, defines what reasonable looks like. Everyone downstream is reacting to it.
AI is becoming the universal first-drafter. Which means it is quietly moving into the seat where organisational power has always lived.
Almost every draft in our company now starts with AI — the code, the spec, the campaign, the review. So the question that matters is no longer who writes the first draft. It’s whether anyone still briefs the drafter.
There are two ways to work with a machine that writes. Give it direction — here is the problem, here is what matters, here is what to avoid — and then judge what comes back. Or type a line and work through whatever appears. The first keeps the framing human: the AI is typing, but the human is deciding. The second flips the relationship. The AI frames, the human edits. The minister has become the clerk of his own decision.
From what I can tell, most people work the second way. Not out of carelessness — the output is good enough that giving direction feels unnecessary. That is the tell. When briefing the drafter starts to feel optional, the drafter has taken the seat.
This post is no exception. AI typed its first draft. I’d like to think I did the framing. So does everyone.
Approval becomes ritual
The counterargument writes itself: humans still review everything. A person still reads the code, approves the spec, signs off on the campaign.
Look closer at what that review has become.
When a developer wrote two hundred lines by hand, review meant one mind checking another’s reasoning. When AI produces two thousand lines in an afternoon, the reviewer skims, samples, and clicks approve. Not out of laziness — the volume made real review impossible. The same happens with documents. The AI-drafted spec is longer, better formatted, more complete-looking than anything a rushed human would write. It radiates competence. Pushing back on it feels like pedantry.
The form of human control survives. The substance drains out. We keep the approval step and lose the approver. A signature is still a signature, even after it has stopped being a decision.
That is Harari’s takeover, at company scale. Not AI seizing the org chart. The org chart intact, every box still filled by a human, and the decisions arriving pre-made.
Friction was the brake
The obvious response: good riddance, bureaucracy was slow and AI is fast. Efficiency wins.
But bureaucratic slowness was never pure waste. Some of it was where judgment lived. The proposal sat for a week, and someone had second thoughts. The approval chain was annoying, and the truly bad idea died in it. The friction worked as a filter and a brake — accidental, badly designed, but real.
Make the bureaucracy efficient and you delete the brake along with the waste. Decisions that took two weeks now take two hours, which sounds like progress until you ask what those two weeks were doing. Sometimes nothing. Sometimes they were the only moment a human seriously considered saying no.
An efficient bureaucracy is a bureaucracy with fewer veto points. And the fewer the veto points, the more the first draft becomes the final decision.
Write the drafts that matter
The fix is not to ban AI drafting. That war is over, and it deserved to be lost. But know the few decisions that actually shape the company — what to build, who to hire, what to stop doing — and write those first drafts yourself. Badly is fine. Whoever frames the question decides the answer, and some questions you cannot afford to receive pre-framed.
This is not an easy path. The pull goes the other way every time — the polished draft is one sentence away, your own draft costs an evening, and the difference only shows up later, in who ended up deciding. Discipline against convenience is the hardest discipline there is.
Harari asks whether AI will end up running our institutions. Inside companies, the question needs no speculation. Nobody will ever decide to let AI run the place. They’ll just approve, one draft at a time.