In 2016, Tom Goodwin observed that Uber owned no vehicles, Airbnb owned no real estate, Facebook created no content. The point everyone took was about assets. The deeper point was about friction.
Not all friction is the same. Some of it is physics — the genuine complexity of matching a driver to a passenger, a room to a traveler. Some of it is psychology — the mistrust, the information asymmetry, the behavioral residue of transactions that had always favored one party over the other. The platforms that won attacked the psychology friction and left the physics friction alone. They didn’t make cars appear faster. They changed the social conditions under which the transaction happened.
Organizations looked at this and drew the wrong lesson. They decided friction reduction was the goal. Flatten the hierarchy. Eliminate approval steps. Accelerate decisions. They optimized for smooth without asking which friction was waste and which was load-bearing.
Now we’re doing it again, at a different layer, with higher confidence. Autonomous agents. Agentic AI. The self-driving business. The names are still settling but the promise is identical: handle the coordination overhead, automate the psychology friction away, and what remains is the work itself.
The self-driving car is not a metaphor here. It is a precedent.
A self-driving car doesn’t face the trolley problem occasionally. It faces versions of it constantly, at speed, with no time for deliberation. Someone encoded the value hierarchy in advance — engineers, ethicists, lawyers, liability actuaries, regulators, all arguing in a conference room about which competing good to sacrifice when sacrifice becomes unavoidable. The car executes that output. It doesn’t know the output was a frozen political compromise. It just acts.
When the situation is novel enough that the encoded hierarchy produces a wrong answer, the car doesn’t feel the wrongness. It just acts.
The organizational agent is the same structure. The rules it runs on came out of a process — someone’s prior judgment about what matters, made at a different time, under different conditions, by people who couldn’t anticipate this specific situation. The agent applies that judgment confidently, at scale, across thousands of decisions. When the hierarchy is wrong for this situation, it doesn’t feel the wrongness either.
This is what the automation promise delivers. Not the removal of human contradiction from the system. The consolidation of it.
The psychology doesn’t disappear. It moves upstream, into the rule-definition layer, where it becomes invisible and largely uncontestable. The contradiction that used to happen in real time — in the meeting, in the hallway, in the moment when someone said we can’t do this to those people and that stopped something bad from happening — now happened two years ago in a planning session and is running silently inside the logic of the system.
One kind of human presence you can automate — the coordination overhead, the middle management, the approval chains, the meetings. The other kind you cannot. The value hierarchy embedded in the rules, the frozen compromise, the political settlement, the judgment call made by people who are no longer in the room. Those don’t go away. They just become harder to find.
Uber didn’t automate taxi dispatch. It made taxi dispatch unnecessary — not because dispatch was bad work, but because dispatch existed to manage a coordination problem the platform dissolved. Once the psychology friction was removed, the entire layer of infrastructure built to contain it became overhead with nowhere to go.
But Uber still had humans setting the rules. Pricing algorithms. Driver rating systems. Surge thresholds. Each one a frozen judgment about whose interests to optimize for when interests conflict. Those judgments produced outcomes the original dispatch office, for all its inefficiency, never had the scale to impose on everyone simultaneously.
The coordination overhead went to zero. The accountability overhead went with it.
The question isn’t whether to remove psychology from coordination, which assumes the psychology was only ever overhead.
Some of it was. Some of it was load-bearing — the distributed, slow, expensive, contradictory human presence that occasionally caught what the hierarchy above it had wrong. The middle manager who compressed the signal slightly differently was also, sometimes, the person who felt the wrongness before they could name it.
Encoding it into rules doesn’t eliminate it. It just decides, in advance and in the abstract, which contradictions get resolved in whose favor. Then it runs that decision at scale, invisibly, without the friction that might have surfaced the error.
You can take the humans out of the business.
You cannot take the humans out of the business.
The question is where you want them — visible, slow, and variable. Or invisible, fast, and fixed.
Leave a Reply