Part 7 - Your Intelligence Should Stay Yours
Alex Karp, the CEO of Palantir, recently described the way frontier AI does business with enterprises using one word: stealing.
Not hacking. Not breaching. Stealing, as the structure of the deal itself. His argument, compressed: companies pay per token to run their most sensitive work through frontier models, the vendor collects the fee and the compounding insight into how those businesses operate, and the customers stay publicly silent because no executive wants to announce which parts of their operation just became someone else's training signal.
You can discount Karp as a man selling the alternative. Fine. The structure he named does not go away when you discount the messenger. This essay is about that structure, and about the version of AI where the intelligence accumulates on your side of the table.
The arrangement
Look at the deal from one step back.
You pay per token. Token pricing is pricing for compute, not for outcomes. It quietly tells you the vendor is not confident enough in the output to charge for its value, so it charges for the electricity of producing it.
You supply the context. For the model to be useful on your work, your work goes through it: strategy memos, client threads, deal terms, the operational texture of your business. Consumer AI products may use exactly this to improve their models. The terms vary and change; the direction of flow does not.
And the improvement compounds on the vendor's side. The next model release is better partly because of what flowed through the last one. You paid a fee and contributed the raw material, and what you keep is a transcript. What they keep is capability.
Run that loop for a decade and ask where the accumulated intelligence of your working life is supposed to live. Not the documents. The distilled part: the patterns, the decision history, the operational knowledge. In the current arrangement, its permanent home is a datacentre belonging to whoever sold you the tokens.
Put it in the terms this series has been building: someone else ends up holding the map of your territory.
What "yours" actually means
"Your data stays yours" has been marketed into meaninglessness, so here is the concrete version, as built.
MrAgentˣ routes reasoning through the API tier, which under confirmed policy is not used to train the models. Not the consumer chat product. The API. The distinction sounds bureaucratic and is everything: one is a product that may learn from you, the other is infrastructure that processes and forgets.
And everything extracted, every commitment, decision, risk, pattern, and relationship, is stored in your own database. Your instance, your credentials, your control. Both layers live there: the raw record of what happened, and the structured intelligence built from it.
The vendor, me, provides the pipeline and the skills that run over the graph. The graph itself, the compounding asset this whole series has been about, sits in your infrastructure. Cancel everything tomorrow and the intelligence stays where it always was. With you.
That is the architectural inversion of the arrangement Karp described. Not a privacy checkbox. A different answer to where the value accumulates.
One more consequence worth naming. When a vendor's product is a pipeline rather than your data, pricing can finally follow the value delivered: what was caught, flagged, and documented, rather than tokens burned. Token pricing is a confession. It should not be the permanent shape of this industry.
Why this matters now, and for whom
Be honest about who should care.
If you are early, a solo founder with a young inbox, sovereignty is not your urgent problem, and I will not pretend otherwise. Compounding is your problem; the previous essays are yours.
But if you run an agency, a consultancy, a legal or financial practice, a business whose edge is literally how you operate, then your operational record is the most valuable thing you own that has never appeared on a balance sheet. Every client negotiation, every pricing decision, every hard-won pattern of what works. That is the material the current arrangement quietly exports.
The irony is that these firms are the heaviest adopters. The more operational intelligence you have, the more AI helps you, and the more the default arrangement costs you. The customers with the most to gain are the ones with the most to leak.
They are also, not coincidentally, the executives Karp says are furious in private.
The line
The whole essay compresses to one sentence, so here it is.
Your workflows should make you smarter, not your vendor.
The problem with most AI is not that it does not work. It works remarkably. The problem is where the intelligence accumulates when it does.
If your operational intelligence matters enough to protect, it matters enough to compound. Next essay: the name of the layer that does both.
MrAgentˣ is in private beta. You will never start from zero again.
Context Windows Close. AI Forgets Everything. Your Work Should Never Start From Zero.
MrAgentˣ is in private beta. Limited to the first 1,000 until launch.
Join the waitlist