Kenneth Arrow won a Nobel Prize in part for describing a strange defect in the market for information: you cannot judge what a piece of knowledge is worth until you have seen it, and once you have seen it, you no longer need to buy it. The seller risks giving the product away just by selling it. AI has inverted that problem, and most firms have not noticed. Today it is the buyer who gives the product away, just by using it.
Paying for intelligence twice
When your firm subscribes to a cloud AI service, you pay with money. That part is on the invoice.
But there is a second payment that never appears on any invoice. To make the model useful, you have to feed it your work: your documents, your precedents, your internal standards, the way your senior people correct a bad draft into a good one. The better you want the model to perform on your firm's work, the more of your firm's knowledge you must reveal to it.
Arrow's seller risked disclosing knowledge in order to sell it. You, the buyer, now risk disclosing knowledge in order to use what you bought. Call it the Reverse Information Paradox.
And unlike a one-time disclosure, this one compounds. The provider learns more about your firm with every prompt, every uploaded document, every correction. You learn almost nothing about what they are learning in return. The information asymmetry gets more lopsided every month you remain a customer.
The leak is not your data. It is your learning.
Most firms hear this and reach for a data protection checklist. Encryption at rest, a zero-retention clause, a vendor questionnaire. That misses the point.
What leaks is not primarily files. It is what researchers call exhaust: the prompts your people write, the tools your agents invoke, and above all the corrections your experts make when the model gets something wrong. Every correction is a distilled drop of institutional know-how. It encodes what your firm considers good, what it considers risky, and how it actually makes decisions.
This is the knowledge Hayek called particular: the knowledge of time, place, and circumstance that no competitor could ever buy, because it exists nowhere except inside your organization. It is your real moat. And it leaks almost imperceptibly, trace by trace, correction by correction.
Notice the irony in the current arrangement. Model providers built their products by claiming broad rights to learn from the public's data. Then they turn around, impose restrictive terms on anyone who tries to distill their models, and reserve the right to learn from your usage. Learning flows in one direction only: toward them. When learning flows one way, economic value follows it, accumulating with the owners of the learning infrastructure rather than with the creators of the knowledge itself.
Alexander Karp of Palantir has made a version of this point about serious enterprise customers: they want control of their compute, their models, their data, and their edge, and they want certainty that their means of production are not quietly being transferred to someone else. The default cloud AI arrangement performs exactly the transfer he describes.
Patents solved Arrow's paradox. What solves this one?
Arrow's original paradox had an institutional fix. Patents let an inventor disclose an idea without giving it away. Disclosure and ownership were separated by law.
The Reverse Information Paradox needs its own equivalent, and it cannot be a clause in a vendor contract, because contracts govern data while the leak is in the learning. The fix has to be architectural: a hard trust boundary around the machinery your firm learns with, not just the files it stores.
Inside that boundary live your documents, your prompts, your traces, your evaluations, your model weights, and your institutional memory. They accumulate and improve together. Nothing crosses the boundary without your consent. Not even the exhaust.
This is precisely the boundary Koretex was built to draw. Everything runs on hardware you own, inside your own walls or your own cloud tenant. There is no provider on the other side of your usage, because you are the provider.
What a firm must own
In practice, owning your learning loop comes down to five things.
Control. Own your evals, because evals define what good looks like inside your firm. Own your memory, your traces, your feedback, and your right to use the outputs of models working on your own tasks. With Koretex, all of this lives on your machines by default. There is no one else's terms of service between you and your own work product.
Capability. Have a place where models can learn against your real workflows without exposing your knowledge to anyone. On Koretex, every draft your team accepts or corrects trains your private model, on your infrastructure. The improvement belongs to your firm alone. It never trains anyone else's model, and it never leaves.
Choice. Keep the orchestration layer decoupled from any single model. Ask yourself a simple question: if the model you rely on today were withdrawn tomorrow, would your firm's accumulated capability survive? When your evals, your fine-tuned weights, and your workflows live inside your own boundary, the answer is yes. Swap the base model and your firm's veteran knowledge stays with the veteran, not with the departed employer.
Cost. Decoupling also lets you route work sensibly. Small models on the laptops you already own handle everyday work at zero marginal cost. A single dedicated machine handles the heavy lifting. No per-seat subscription that grows with every hire, and no bill for the privilege of training someone else's product.
Compound. Put the four together and you get a continuous learning loop that belongs to you: a hill-climbing machine pointed at your firm's own definition of success. In the cloud era, firms accumulated data. In the AI era, they accumulate learning. The firms that own their loop will get measurably better at their own work every month, in a way no competitor can rent.
The right to learn privately
A firm should be able to use a model without surrendering the knowledge that makes it unique. That is the standard every risk committee, managing partner, and board should now hold vendors to. In consuming intelligence you are creating intelligence, and what you create should belong to you.
Arrow's paradox got patents. The reverse paradox gets a boundary: your models, your evals, your learning, on your hardware, behind your walls. That is what we build at Koretex.