The best lawyer in the world cannot be called to the bar

Marc Andreessen spent part of last week making a point that is easy to hear and hard to absorb. The intelligence required to be the best doctor, lawyer, teacher or accountant on earth is already in everyone's pocket. And none of it can practise.

Marc Andreessen spent part of last week making a point that is easy to hear and hard to absorb. The intelligence required to be the best doctor, lawyer, teacher or accountant on earth, he argued, is already in everyone's pocket. And none of it can practise. It cannot be called to the bar. It cannot be licensed to prescribe, credentialed to stand in front of a classroom, or cleared for insurance reimbursement. He divides the economy into blue sectors, open to the technology, where prices fall, and red ones, healthcare, education, housing, where the gate stays shut and prices keep climbing.

You do not have to share his conclusions about the gate to accept his description of it. Somewhere between what these systems can do and what they are allowed to do sits a decision. That decision has a name in this framework. It is Permission, the first of the three Ps inside the Process layer, and it is the subject of this piece. Practice, what happens once the system runs, follows later today. Proof, how you demonstrate any of it to the people who carry the liability, comes on Friday.

What permission actually is

Permission is the deliberate act, taken before a system runs, of deciding what it may touch, on whose behalf it acts, and where it must hand the work back to a human. Three tests tell you whether an organisation has it.

It is written down. Not as a forty-page policy nobody reads, but as a scope: these inputs, these actions, these customers, these limits. If the scope cannot be stated on one page, it has not been decided.

It is owned by a name. Someone with the authority to grant the mandate signed it, and that person can be asked, later, why the boundary sat where it did. Permission granted by everyone is permission granted by no one, which is how committees launder the absence of a decision.

It is revocable. A mandate that cannot be withdrawn is not a mandate, it is an abdication. The organisations that move fastest are, counterintuitively, the ones whose systems can be switched off cleanly, because a reversible decision is an easy decision to make.

Note what permission is not. It is not compliance theatre, and it is not the risk register. It is the instrument that makes deployment possible, in exactly the way an operating licence makes a plant possible. Nobody calls the licence a brake on the plant.

The week permission became the product

If you want evidence that permission has become the scarce asset, follow the money the labs are spending on it.

Washington, weeks after holding the frontier models back from the world, has now agreed capability thresholds and disclosure timelines with the labs, to be policed by the NSA and a new standards body. The trusted players are back in. Fable 5 is live again. China remains out. Read quickly, it is a wall. Read again and it is a moat, one the labs helped dig. Sam Altman is proposing a global body to admit whoever follows rules he would help write. OpenAI has offered the US administration a five per cent stake in itself. Anthropic handed California's government Claude at half price. These are not the actions of firms selling capability. Capability is abundant. They are the actions of firms buying permission, at national scale, and the bar the new framework sets, safety teams, trusted-access programmes, a government liaison, is one only a trillion-dollar company can clear.

The same repricing is visible from below. Ireland's new rule on financial products, a statutory right to reach a person rather than a chatbot, is a permission boundary drawn in law: the machine may inform, it may not be the only door. Shenzhen drew its own line the other way, clearing robotaxis over the objections of 400,000 drivers. Two jurisdictions, two opposite answers, one identical mechanism. Someone with authority decided what the system may do alone, and wrote it down.

The strategic consequence for everyone else is uncomfortable. When permission is the product, the firms that can manufacture it internally, quickly and legibly, hold an advantage that no model subscription can buy.

The two failed regimes

Most organisations, asked to show their permission structure, produce one of two artefacts, and both are the absence of a decision.

The first is the blanket ban. Nothing is permitted, pending a policy. The predictable result is shadow AI: the work moves to personal phones and private accounts, the organisation keeps all of the risk and captures none of the learning, and the eventual policy is written by people who believe usage is zero. The ban does not stop the technology. It stops the organisation from seeing it.

The second is the eternal pilot. Everything is permitted, provided it stays in the sandbox. Proofs of concept multiply, demo days impress, and nothing ever receives the mandate to touch a live customer, a live batch record, a live claim. The pilot is not caution. It is a mechanism for deferring the one decision that matters, and its yearly cost is invisible because it is paid in absent compounding.

The way out of both is the same instrument: a scoped, named, revocable mandate. Start narrow. A system may draft, a human sends. A system may flag, a human releases. The narrowness is not timidity, it is what makes the signature possible, and the signature is what makes everything after it possible. Firms in regulated sectors keep discovering they are good at this, because the instrument already exists in their muscle memory. The MHRA has begun saying out loud what the best operators already knew, that a legible rulebook is a catalyst, not a barrier, because a firm that knows exactly what it must demonstrate can move at full speed inside the line.

The line Toyota drew first

Toyota in 1985 ran more automation at full speed than almost anyone, and it did so on the back of a permission boundary decided years in advance. Its machines held permission to stop the line, instantly and without appeal, the moment something deviated. That is jidoka, and it was a genuine mandate: the machine's judgement, inside its scope, was final. What the machines never held was permission to define what good looked like. The standard belonged to people, and every machine stoppage was referred back to them.

Notice the shape of that arrangement. It is not a compromise between trust and control. It is a precise allocation: full authority inside a written boundary, none outside it. That allocation is what made the speed safe, and the safety fast.

The AI version of the question now sits in front of every operating committee: what, exactly, does this system have permission to do alone? If the answer takes longer than a sentence, the real answer is that nobody has decided, and the capability in every pocket will keep waiting, as Andreessen says, for the only thing it cannot generate for itself.