BLUF — The alpha was never the model. It's the instructions your agents run on and the record of what they decided — your judgment, written down. Keep that layer on your side and you can build on any frontier model and still change your mind later. Sovereignty isn't which model you pick; it's where your operating logic lives.
The alpha isn't the model
When Palantir's Alex Karp warned that enterprises are handing over the “weights and alpha” of their business, the weights half is easy to picture. The alpha is the interesting half — and in an operation it's more concrete than it sounds.
Your alpha is three things. What your agents are told to do — the instructions, the rules, the checks, the phrasing that took four attempts to get right. How the work routes — which step runs when, what gets escalated, where a human has to put their name on it. And the record of what was decided and why.
The model supplies raw reasoning. Reasoning is becoming a commodity: several vendors sell it, quality converges, prices fall, and the leader changes every few months. Your operating logic is not a commodity. Nobody else has it, because nobody else ran your operation for the last decade.
The model supplies the reasoning. Your instructions carry the judgment. Only one of those is yours.
Where operations quietly give it away
Almost nobody decides to hand over their alpha. It leaks, in three ordinary ways.
• Buried. The instructions are hard-coded inside each agent by whoever built it. They work, so nobody touches them — and nobody can tell you what changed, when, or why the results drifted last month.
• Scattered. Every team writes its own version of the same rule. Nine agents encode the payer policy nine slightly different ways, and when the policy changes you find out which ones you missed from the denials.
• Borrowed. The logic lives inside a vendor's platform in a format you can't export. You can see it; you can't take it with you. That's the one that ends the conversation about switching before it starts.
Each of these is survivable on its own. Together they mean the most valuable thing your operation has built — its encoded judgment — isn't actually in your custody.
Keep the bottom layer, rent the top
Picture your agent stack as two layers, and be honest about which one you're protecting.
Read that way, the frontier-model question stops being existential. Use the best model available for the job. Just make sure that when the pricing changes, the terms change, or something better appears, the thing you'd have to move is configuration — not your operation.
Two reasons to pull your instructions out
It's your intellectual property. The escalation rule that says which supplier delay actually threatens the build schedule isn't a prompt — it's twenty years of a procurement officer's instinct, finally written down. The check that catches the clinical edge case a payer always denies is the same. That text is institutional knowledge in its most portable form, and it deserves to be treated like an asset rather than an implementation detail.
It's your exit — and this is the stronger argument. Be honest about the first point: a prompt on its own is copyable. Anyone can paste it. What isn't copyable is the accumulated, tested set of them, tuned against your real cases, plus the record of which versions actually worked. But even setting IP value aside, externalized instructions buy you something more practical: the ability to leave. If your logic sits in a versioned layer you control, changing models is a configuration change. If it's baked into agents someone built two quarters ago, changing models is a rebuild — and that's how enterprises end up staying with a vendor they'd rather leave.
Own your instructions and the model becomes a supplier. Leave them scattered and the model becomes a landlord.
Governance and sovereignty turn out to be the same layer
Here's the part that makes this easier than it sounds, and it's the reason I'd put it on this quarter's list rather than next year's.
The layer that keeps your alpha — one place where instructions live, versioned and reviewed, with a log of what every agent did — is the same layer that gives you governance. Change control, so untested logic never reaches a live workflow. An audit trail, so you can reconstruct a decision months later. Escalation, so the hard calls route to a named human. You don't run a sovereignty program and a governance program. You build one layer and collect both.
Which is worth saying plainly, because governance usually gets sold as a cost and sovereignty as a principle. In practice they're the same investment, and it pays twice.
What this looks like in a real operation
Take a prior-authorization team. Payer rules change constantly; the rules are the work. When they live as versioned instructions in one place, a change is made once, reviewed, dated and rolled out — and if it makes things worse, rolled back the same afternoon. When they live inside nine separate agents, a rule change is a project, and the first sign you missed one is a denial.
Or a defense procurement desk handling long-lead sourcing. Which clarification genuinely threatens the schedule, which supplier's slippage matters, what has to be escalated to a named officer before commitment — that encoded judgment is the operation's actual advantage. It should not be an artifact of who happened to build the agent, and it should not sit in a system you can't take with you.
What to do this week
None of this needs a platform decision. It needs to know where you stand.
Find out where your instructions live. Pick one agent running in your operation and ask a simple question: where is its prompt, and who approved the current version? If the answer is “in the code” or “ask the person who built it,” you've found the gap.
Version one instruction set. Take the rules for your highest-volume workflow and treat them like an operational policy — written down, reviewed, dated, with a way to roll back. Start with one; the habit matters more than the coverage.
Run the swap test. Ask your team what it would take to move that workflow to a different model next quarter. If the answer is anything more than a configuration change, you've measured your lock-in.
Keep the record on your side. Make sure the log of what your agents decided — and why — lands somewhere you own and can query, not only in a vendor console you rent access to.
A final thought
The enterprises that come out of this decade in a strong position won't be the ones that picked the right model. That bet resets every few months, and everyone gets roughly the same reasoning for roughly the same price.
They'll be the ones who kept writing their judgment down, in a place they control, while everyone else was busy choosing vendors. Frontier models are worth using. Just don't build your operation so deeply inside one that leaving would mean starting over.
So the question this week: if the instructions your agents run on had to move somewhere else tomorrow, could you find them all — and would you still recognize your operation on the other side? Tell me in the comments where yours currently live.
Recent blogs
Secure your agents
We’d love to chat with you about how your team can secure and govern Ai agents everywhere
