The AI conversation is back to extinction — the systems get too capable, escape our control, and that's the end of the story for us. It's a serious question. It's also, for anyone responsible for how an organisation actually works, the wrong altitude. If the ending is already written, leadership has nothing to do but worry.
There's a nearer question, and it's the one every functioning society has had to answer: what's the deal? A social contract is the set of terms people agree to — what they hand over, what they keep, what they'll defend. Organisations are writing one with AI right now, clause by clause, mostly by default. Every time work is initiated, decided, or signed off with a machine in the loop, a term gets set. The risk isn't that the terms are hostile. It's that no one is reading them.
Four terms decide whether it's a deal worth making.
The failure that should worry a board isn't a system cleverer than its staff. It's a system that sounds certain and is wrong, and an organisation that has stopped checking because the confident output is faster to accept than to challenge. Authority without judgment isn't intelligence; it's a very efficient way to be wrong at scale. The term to hold: the machine informs the decision, a person makes it.
Work is trusted because someone did the thinking — that's what a customer, a regulator, an auditor is actually relying on. AI is good at producing the shape of effort without the effort, and output you can't trace to a person is output you can't defend. The term: keep human effort where it counts, and make it provable. Not a claim of human-in-the-loop — a record of it.
Technology is a multiplier: it makes capable people far more effective and confused ones far more efficiently confused. The moment the tool sets direction and the person services it, the driver and the engine have swapped. Keep people in front. The system is the force behind the work, not the hand on the wheel.
You can delegate the work; you can't delegate the responsibility. "The model did it" isn't an answer a regulator accepts or a customer forgives. Someone chose to deploy it, to trust it, to not check. If no one will stand behind a decision, it shouldn't be made that way.
Here's the part most AI strategies miss: a contract you only write down gets ignored. Terms that live in a policy PDF don't survive the first quarter of delivery pressure. The terms that hold are the ones built into how work actually runs — how it's initiated, governed, evidenced, and owned. That isn't a values statement. It's an operating model.
Done properly, "people-directed" stops being a reassurance you offer and becomes a fact you can show — as-measured, in the record, not on faith. That's the difference between a company that has opinions about the human-AI contract and one that's actually keeping it.
Which is the whole point of building an operating model deliberately. An AI-native one is the mechanism that encodes the deal into the work: people direct it, the machine multiplies, every action is recorded, and accountability has a name on it. It's how the four terms survive contact with reality instead of eroding under the first deadline.
The extinction story asks nothing of leadership but attention. The operating question asks something harder: decide what stays human, and build it so it holds under pressure. That's the work — and unlike the end of the world, it's entirely ours to do.