Every technology story arrives as a price falling. Compute costs less per token. Generation costs less per document. Code costs less per feature. The price is real, and it is the least interesting part.
Falling prices do not remove constraints. They relocate them. When one input becomes abundant, the scarce thing becomes whatever the abundant input still needs to be useful.
Authorization
An agent with a shell, credentials, and access to infrastructure can do far more than it is permitted to do. Capability means what the setup can reach. Permission is a separate record, granted by a person, and it does not follow from capability.
As execution gets cheap, the scarce thing is the decision that lets it run. A model that can take a thousand actions still needs someone to say which actions are allowed, under what conditions, and for how long.
That record is not generated by the model. It is granted by an institution, and institutions are slower than models.
Verification
A source changes; the claims citing it do not. A model produces a summary; nobody checks it against the record. Cheap generation increases the volume of claims that need checking, and checking does not get cheaper at the same rate.
The bottleneck moves to the review. That is where the cost lands, and it is the part of the workflow most deployments measure least.
There is a second-order effect. When generating a claim is cheap, the number of claims that could be wrong grows faster than the number anyone can read. The scarce resource becomes attention, and attention does not scale with compute.
Recovery
An agent can roll back an action and leave the consequences in place. Restoring the system is not repairing what it touched: the customer, the record, the payment, the patient.
Cheap execution increases the number of actions that might need repair. Repair is still human, slow, and local. The gap between how fast a system can act and how long it takes to undo the consequences is the number that matters, and almost nobody reports it.
A hypothesis, not a law
This is not a claim that technology always concentrates power. Cheap technology can dissolve a bottleneck: distributed generation weakened centralized utilities, and cheap publishing weakened gatekeepers.
The claim is narrower. When the cheap input is action, the constraints that remain are the ones that decide whether action was allowed, whether it worked, and whether it can be undone.
Where value moves
If that holds, the value in AI shifts from generation to the record around it. Who authorized this. What evidence was in front of them. What changed, and when. What is still wrong after the correction.
Those are the questions I build around, and none of them is answered by a better model. They are answered by records, checks, and the authority to stop.
The commercial version is blunt. A company that can generate the same output as its competitor has no advantage in generation. It has an advantage if it can prove what it did, show that it was allowed, and reverse it when it was not.
The question to ask
When a capability gets cheaper, ask what it makes scarce. The answer is usually not the capability. It is the permission, the verification, or the repair that the capability now requires more of.
The firms that win the next few years may not be the ones with the cheapest model. They may be the ones that can answer three questions faster than anyone else: who authorized this, did it work, and what is still wrong.