Why the Kasparov experiment, not the model leaderboard, is the right way to think about AI at work.
In 1997 Garry Kasparov, world chess champion for fifteen years, lost a six-game match to a machine. He was the first reigning champion to lose to a computer under tournament conditions. Newsweek ran the cover ‘The Brain’s Last Stand.’
The interesting part of this story isn’t the loss, it’s what Kasparov did next. Instead of treating the machine as the end of human chess, he ran a different experiment: let humans and computers play as teams, and see who wins. The result that stuck with me is that a weaker player paired with a machine and a better process beat a stronger player paired with a machine and a worse process. The decisive factor wasn’t the strength of the human or the power of the machine. It was the quality of the process running between them.
That’s the point I keep going back to when the conversation at work turns to AI, because the conversation almost always starts in the wrong place: which model, which tool, how many credits. Those are the parts that change every few weeks and the parts none of us control. The model improves for everyone at once. A new tool is a click away for your competitor too. If your advantage is “we use the good model,” you don’t have an advantage, you have a subscription.
The part you own is the context. The materials you feed in, the examples, the playbooks, the way your team actually does the work when it’s done well. We’ve been describing this internally as a recipe with a few ingredients, and the one that matters most isn’t the model at all, it’s the engineered context. A model is a capable generalist that knows everything in general and nothing about how you specifically work. The thing that teaches it your way is a skill: a scoped, version-controlled instruction for one job, in your style. You write it once, you improve it over time, and it compounds. The model is rented. The skill is yours.
I’ll be straight about the other half of this, because a notebook is for the parts that don’t fit on a slide. When I laid this framing out for the team, nobody cheered. The sharpest feedback was about load. The worry, said plainly, was that we’d pile on new expectations without taking anything off the plate, and that “process is the moat” would quietly become “do more, faster.” That’s a fair worry, and it’s the real failure mode of this whole reframe. A better process is supposed to make the work better and the person freer. If a new way of working only adds, it isn’t better. It’s heavier, and you’ve missed the point of the Kasparov result, which was never “work harder.” It was “the process is where the leverage lives.” Leverage should give time back, not just consume more of it.
The teachable part
Stop optimizing for the tool and start optimizing for the context you own. The model is the easy part, it’s rented and it’s the same one your competitor can rent tomorrow. The advantage and your competitive moat lives in the process you run and the skills you encode, because those are yours and they compound. And hold the line on what a better process is for: if a new way of working only adds to the plate, it isn’t a better process, it’s a heavier one.