The oldest rule of business is a rule about ownership: what a company does not own, it does not control, and what it does not control can be repriced, withdrawn, degraded or turned against it by whoever does. Merchants have organised themselves around this rule for centuries, owning the recipe and the ledger while hiring everything peripheral. The rule survives intact; what changed is the thing it applies to. Judgement itself can now be bought as a service, and companies are deciding, mostly without noticing the decision, whether the thinking layer of their business is something they own or something they borrow. The question deserves a deliberate answer, because both answers are being given by default, one procurement decision at a time, in choices nobody frames as constitutional.
Someone else’s judgement
A model is judgement, frozen into a machine. Its makers decided what it read and what it never saw, what it favours and what it refuses. Those decisions reflect the makers’ data, which carries the makers’ biases, assumptions, commercial interests and blind spots. None of this is visible from outside, none of it is announced when it changes, and all of it is inherited, wholesale, by every company that builds on the model.
The other cost of inherited judgement is sameness: it is the same judgement everyone else inherited. A company reasoning with the same borrowed mind as its competitors will price like them, write like them, evaluate like them, and miss what they miss. Sameness at the thinking layer is a strategic condition, and it cannot be fixed at that layer, because the layer is shared. A firm’s edge has always lived in judgement that differs from the market’s. Judgement bought from the market cannot, by construction, differ from it.
None of this requires anyone upstream to behave badly. A maker’s judgement enters its model the way an author’s judgement enters a book, through ten thousand small choices nobody itemised. The difference is that a reader can see the book. A company cannot read the model it is reasoning with, and it cannot compare yesterday’s judgement with today’s after a silent update.
Inside someone else’s walls
Borrowed thinking has a second property, older than any technology: work done inside someone else’s walls is visible to the owner of the walls. Everything a company sends through a supplier’s model passes through the supplier’s machines: the contracts under review, the strategy being drafted, the numbers being questioned, the weaknesses being probed. The questions alone are revealing. A company’s questions are a map of what it does not know.
The scale is what is new. Showing one letter to one hired translator was a bounded exposure, weighed case by case. Routing the reasoning of an entire company through outside machinery is a standing arrangement, renewed silently with every use, covering whatever the day’s work happens to be.
A supplier that promises not to look has adopted a policy, and policies change with ownership and with pressure. The ability to look is part of the architecture, and no change of management revisits that. Confidentiality that depends on a counterparty’s continued restraint amounts to trust with better marketing. And the liability does not travel with the data: a company answers to its clients for what happens to their information wherever the exposure occurs, so the duty remains at home even after the visibility has moved out.
Terms set elsewhere
Whoever owns the tool sets its terms. A supplier can reprice a model, restrict it, retire it or change its behaviour, on its own schedule, for its own reasons, and a company built on that model absorbs the change with no vote and often with little notice. A notice period in the contract softens the timing of a change without touching the power to make it. That is the permanent structure of depending on what someone else owns rather than a defect of any particular supplier, and it is as old as the lease. The depth of the dependency has no precedent: not one process but the reasoning inside every process, moving on someone else’s roadmap.
Prudent companies keep alternatives for critical supplies, and the habit applies here with a twist. The only real alternative to a supplier of thinking is a capacity for thinking held in-house, because a second supplier of the same borrowed judgement offers only the same dependency wearing a different logo.
What the assets are
Owning the AI begins with naming the assets rather than with training an enormous model, because a company cannot own what it has not named:
- The data. The records of the business, prepared to the point where a model can use them. This asset exists before any AI does and outlives every model that reads it.
- The standard of correct. The company’s own written judgement of what a right answer looks like, tested against real cases. It measures every model, and it transfers intact when models change.
- The models it can hold. Weights the company controls, or its own adaptations of open ones, trained on what the company knows. Scale matters less than title: a small model the company holds moves on the company’s schedule.
- The record of use. What was asked, what worked and what failed, captured as property rather than surrendered as a supplier’s telemetry. It is the raw material of every future improvement.
The spending test follows directly: money spent on AI counts as investment when it leaves behind an asset on this list, and as a permanent expense, paid again forever at prices set elsewhere, when it leaves nothing behind but output.
Balance sheets exist because ownership is how a business stores work. Effort spent on owned things accumulates somewhere it can be counted and defended; effort spent on borrowed things accumulates in the borrowed thing, for the benefit of its owner. The same ledger logic now applies to intelligence, and most companies have not yet opened the account.
Where owning is a mistake
Ownership is a default, not a religion. Work at the frontier of capability belongs on frontier models for as long as the frontier matters to the task. Exploratory work, where the company is still learning what AI is for, deserves cheap borrowed capability rather than premature commitments. Processes that touch nothing confidential and produce no differentiation can run on anyone’s model indefinitely, and owning them would be vanity. The line sits where the two arguments above begin to bite: where the work is core, and where the information is nobody else’s business.
Drawn honestly, the line still leaves most of the interesting work inside it, because the work that defines a company is precisely the work that is core and confidential at once. The mistake to avoid is borrowing the middle of the business and believing it was rented furniture.
The oldest rule, the newest property
A company is, increasingly, its thinking layer. Products can be copied and buildings are leased; what compounds over decades is the quality of the decisions and the knowledge those decisions rest on. That layer is now partly mechanical, which means that for the first time it can be owned outright, like a ship or a recipe. To own the AI assets is simply the oldest rule of business, applied to the newest thing a company can hold title to.