The Learning Ledger: Your AI Vendor Is Banking Know-How You Never Booked
Every AI deployment produces two outputs. The first is the work: the drafted contract, the resolved ticket. The second is the learning: everything your team figured out about making the model produce that work. Which prompts held up under pressure. Which outputs needed correction, and how. What good looks like in your business, encoded in a thousand small judgments. Most companies capture the first output and hand the second one away. We call the account where that second output accrues the learning ledger. Right now, for most enterprises, the balance sits with the vendor.
What did Nadella actually say, and is he right?
He said AI buyers pay twice, once in money and once in proprietary knowledge, and on the diagnosis he's right.
On July 12, Microsoft CEO Satya Nadella published a post arguing that "you essentially pay for intelligence twice." The second payment is the know-how you feed a model to make it useful: the prompts your people write, the tools your agents call, the evals you run, the corrections your team makes when the output is wrong. His point is that this learning flows in one direction. The provider studies your usage and gets smarter about your business. You get a bill. The post pulled close to ten million views in two days, which tells you the anxiety was already out there waiting for a name. Palantir's Alex Karp has been circling the same idea for months: technical buyers want assurance that their means of production aren't being transferred to someone else while they pay for the privilege.
But the diagnosis stops one step short of the operator's problem. A leak implies you had the asset stored somewhere. Walk into most enterprises running AI today and ask to see the stored asset. There's nothing to show you.
What is the learning ledger?
It's the account where the learning generated by your AI usage gets booked, and it's always getting booked somewhere.
Treat every correction, every eval, every prompt revision as a ledger entry. If the entry lands in a system you control, it accrues to you. If it only lives inside a vendor's logs and traces, it accrues to them. There's no neutral outcome. Usage generates learning by default, and the learning compounds on whichever side captured it.
Picture a claims team that spent six months getting an intake agent to production grade. The prompt went through forty revisions. The team built a shared sense of which edge cases break it and which corrections stick. Now ask where all of that lives. If the answer is a chat history and two people's heads, the entry was made, and it landed on the vendor's side of the ledger.
Two distinctions matter here. We've written before about data you can't recall, which is an exposure problem: information leaving your control. The learning ledger is an accrual problem: value compounding on the wrong balance sheet. You can have airtight data protections and still bank nothing, because data protection covers what goes out, while the ledger covers what should have been kept. And if intelligence is capital, a position we've held since we coined return on intelligence, then the ledger is where the interest lands. Spend without one and you're paying compound interest to someone else.
Why can't most companies just start owning their learning loop?
Because there's nothing to own yet. The capture infrastructure was never built.
Here's what AI adoption actually looks like inside most enterprises. Teams run pilots. People correct outputs inside chat windows. The best prompts live in someone's personal notes. Quality gets judged by feel, by whoever happens to be reviewing that day. The learning is real and it happens constantly. It just evaporates by Friday, or it accrues in the vendor's telemetry, take your pick. No eval sets encoding what good looks like for your workflows. No versioned prompts. No captured corrections. No record of which change moved which outcome. When the champion who ran the pilot takes a new job, your opening balance is zero.
Nadella's prescription assumes a company that's ready to bank its learning. The prerequisite he skips is a company that bothered to open the account. That gap sits between his post and your Monday morning, and it's where the actual work lives.
Doesn't a model-agnostic stack already solve this?
No. Swap-ready architecture protects your choices; it doesn't capture your learning.
Part of Nadella's fix is decoupling the orchestration layer from any single model, so your capability survives if a given model goes away. We've been making that argument for a long time: build swap-ready architecture, treat the model as a socket, never let single-model risk set your roadmap. Watching the CEO of Microsoft arrive at the same place is validating. But portability and accrual are two different problems. Solving one leaves the other exactly where it was. You can swap models freely and still carry nothing forward, because the evals, corrections, and workflow judgments were never written down anywhere the next model could inherit them. The socket keeps you free to move. The ledger decides whether anything moves with you.
What should you do about it?
Install one rule: no correction happens off the books.
Every time a person corrects a model's output in a workflow that matters, the correction gets captured in something your company owns. A new eval case, or a one-line note on what failed and what fixed it. That's the whole rule. You don't need a platform purchase or a six-month data strategy to begin. Apply the rule to your two or three highest-value workflows and let it force the infrastructure question on its own schedule. Within weeks you'll need a place to put the entries, and that's when tooling decisions get easy, because you finally know what they're for.
A quarter of this and you have an opening balance: a body of evals and captured judgment that makes every future model better on day one and rides along when you switch providers. The vendors are already compounding your know-how. The only open question is whether you're compounding it too.
If your AI spend is producing learning you can't point to, that's a gap we can help closes: Learn about our AI Blueprint approach or reach us at contact@theyor.com.