Satya Nadella recently published a short piece we keep coming back to. His argument is that the AI era will not be won by whoever rents the best model. It will be won by companies that build a learning loop, where their people and their AI compound each other over time, and where the firm owns the institutional knowledge that loop encodes. As he put it, you can offload a task, or even a job, but you can never offload your learning.
We think he is right. We also think the takeaway for most companies is simpler and more urgent than the framing makes it sound, so we want to make it concrete.
The best thing your company can do right now is document how it actually works, and start building systems with AI that put that knowledge to use and get better every time. That is the on-ramp to everything Satya is describing. And almost no one is doing it.
The model is not the moat
Most leaders are still asking which model to bet on. It is the wrong question. Everyone has access to the same models. They improve every few months, and each step makes generic capability cheaper for everyone at once. If your whole AI plan is "use the best model," you share that plan with every competitor — and with the model maker who can do it cheaper than all of you. Generic capability is being commoditized in real time. Building your advantage on it is like building your advantage on having electricity.
What cannot be commoditized
What cannot be copied is what your company actually knows. How a job really gets estimated. Why you walked away from a certain kind of work last year. The judgment your best person carries after twenty years. The reason a step exists in a process that looks redundant until the one time it saves you. That knowledge is specific, hard-won, and almost always undocumented. It lives in people and habits, not in systems, and it walks out the door when they do.
That is the real asset. Most companies are sitting on it without capturing a word of it.
Write it down, then wire it up
So the move is not complicated to describe: write down how you actually work, and build systems with AI that use that knowledge and improve with every use. It comes in two parts.
Write it down
Capture the real process, not the version on the org chart — the decisions and the reasons behind them, the mappings, the exceptions, the hard-won "don't do it that way." Put it somewhere both your people and your AI tools can reach. This is the part people skip.
Wire it up
Once that knowledge exists in a usable form, build systems on top of it that handle the repeatable work and get sharper each time. Each documented process makes the next AI output better. Each correction teaches the system something it keeps.
That second part is where it gets interesting. It is the learning loop Satya describes, and it is available to you now with tools you already have.
"Models can do the work. Only you can own what you learn from it."
Why this compounds, and why it wins
Here is why this matters more than any model decision. Most business advantages are one-time. You make a hire, win a deal, cut a cost. They help once and they are spent. A knowledge system is different. It compounds. Every improved workflow makes the next one easier. Every captured decision makes the next decision better.
The company that starts now does not get a one-time bump. It pulls a little further ahead every week, on top of an advantage competitors cannot even see, because it is built from knowledge they do not have. You cannot buy that — you grow it. And the gap is hard to close, because by the time a competitor decides to start, you are years of compounding ahead. That is how you move faster and smarter than your competition, and stay there.
Own the knowledge, not the tool
One rule protects all of it: own the knowledge, not just the tool. Models will come and go. You should be able to swap this year's model for next year's without losing the company veteran you have been building inside your systems. If your accumulated knowledge lives only inside one vendor's product, you are renting your advantage, not owning it. Keep it in a form you control. This is the same point Satya makes about sovereignty, and it is the difference between building an asset and leasing one.
A practical test: if you switched AI vendors tomorrow, how much of what your systems know would walk out the door with the old tool? If the answer is "most of it," you are leasing your advantage. Keep the knowledge in a store you own and let the model plug into it.
The unglamorous part is the point
None of this is glamorous. It is writing down the procedure everyone assumed was obvious, and capturing the reason behind a decision before it is forgotten. It feels like overhead — the work you do instead of the real work. That is exactly why most companies will not do it, and exactly why it is an edge.
So start small. Pick one workflow that matters and runs often. Write down how it actually happens. Put that knowledge where your people and your AI can use it. Build one simple system that handles the repeatable part and feeds what it learns back into the same place. Then do the next workflow. The goal is not a giant documentation project that dies in a binder. It is a habit of turning what you know into something that compounds.
Start writing it down
The next few years will not be won by the company with the best model. Everyone will have a good model. They will be won by the companies that turned their own knowledge into systems that get smarter with every use, while everyone else was still debating which tool to buy.
Let the models keep getting better. Your job is to build the thing they cannot hand you: the compounding record of how your company thinks, works, and wins. The best day to start writing it down was a year ago. The second best day is today.
Key Takeaways
- Everyone rents the same models, and they get cheaper and better every few months. Generic capability is being commoditized — it cannot be the basis of a durable advantage.
- What competitors cannot copy is what your company actually knows: how the work really gets done, why decisions were made, the judgment your best people carry. Most of it is undocumented and leaves when they do.
- The highest-leverage move is unglamorous: write down how you actually work, then build AI systems that use that knowledge and get sharper with every use. That learning loop compounds week over week.
- Own the knowledge, not the tool. Keep it in a form you control so you can swap models without losing the institutional expertise you have built — an asset, not a lease.