The tools got faster. Then the bottleneck moved. Why the moat was never the model, and what actually decides whether AI works in your business.
There is a version of the AI story that is mostly true and completely misleading. It goes like this: the tools got good, the work got faster, output went up. Anyone who has put Claude or Copilot in front of a capable team has watched it happen. More gets done, and it gets done sooner.
Spotify watched it happen at a scale most of us will never touch. Thousands of their engineers now run AI coding agents every day, across more than thirty-six thousand sessions. Code ships faster. More things get built. And then, in their own telling, something else happened.
The agents started making decisions that were fast, confident, and wrong. Not wrong the way a junior is wrong. Wrong in a more specific and more expensive way: technically correct, and operationally wrong. The work passed review on its own terms. It just did the wrong thing for that particular business, because the agent had no idea what it was working inside. It did not know who owned the thing upstream, what depended on it, or why it had been built that way in the first place.
Here is the part worth sitting with. The knowledge that would have prevented every one of those mistakes was not missing. It existed. It was sitting in chat threads nobody could find, and in the heads of the three people who built that system back in 2021. The company knew. The company just could not hand what it knew to the machine doing the work.
That is not a Spotify problem. That is your business, described precisely.
Think about the best graduate you have ever hired. Fast, sharp, works around the clock, never complains. And for the first six months, quietly dangerous, because they do not yet know how your business actually runs. They do not know which client you never push on price, which process looks broken but exists for a good reason, which supplier relationship is holding up half the operation. They are not short on intelligence. They are short on context. The value they eventually bring is not their raw ability. It is their raw ability plus everything they slowly learned about you.
An AI agent is that graduate, arriving new every single morning. It has the intelligence on day one. What it does not have, and what no amount of model upgrades will ever give it, is the thing your ten-year veteran has: a working knowledge of how this specific company operates. The labs can keep making the graduate cleverer. They cannot make the graduate know your business. Only you can do that.
This is why the race to buy the biggest, cleverest model is aimed at the wrong target. The model is the graduate's raw intelligence. It was never the real constraint, and it is getting cheaper every month regardless of what you do. The constraint is context: whether the intelligence you are renting knows enough about your operation to be right, and not merely clever.
Spotify's answer is worth studying, and it is not a tool. They stopped treating that knowledge as something that lives in people and started treating it as shared infrastructure. Every session an agent runs now starts already knowing the shape of the business, and every session it finishes leaves that knowledge behind for the next one, whether the next one is a person or a machine. Their own summary of the lesson is the sentence to underline: the organisations that win are the ones that compound their operating knowledge into shared, structured systems, rather than leaving it scattered across individuals.
Read that again with your own company in mind, and the code disappears. What is left is a claim about every business trying to use AI well. The winner is not whoever holds the smartest model. It is whoever has made their own operating knowledge legible enough to hand over.
The comforting mistake is to file this under "problems of the very large." It is not. The same failure shows up on day one and at the smallest scale: an AI that does the operationally wrong thing, confidently, because nobody ever made the right thing explicit. A fifty-person firm hits it the first week it lets a tool touch a real workflow. Spotify just hit it thirty-six thousand times, which made it impossible to ignore.
And the fix is the same at both ends. Before you scale the tools, make the business legible to them. Map how the work actually flows, not how the org chart says it does. Write down the reasons behind the processes that look strange to an outsider, because those reasons are exactly what an agent will steamroll. Move the knowledge out of three people's heads and into something the whole business, and its agents, can draw on.
We have said before that the order matters: map, then optimise, then apply AI. This is why. Mapping is not the boring prelude to the real work. Mapping is how you build the context that decides whether AI turns out to be an asset or a fast, confident liability.
Rent the model. It is a commodity, and a cheap one, getting cheaper. The moat was never the model. It is what your business knows, made legible enough to use.
Luke Lombe, Founder and CEO, Echelon One
