11 June 2026
The self-improving company
Stuart Wan
There's a version of the AI-native company going around in talks and videos: a closed loop. Work produces output, output produces learning, and the learning feeds back in so the next cycle runs better. A company that improves itself. It's a genuinely good idea, which is rare for something this hyped. So a few months ago I started pushing our own company toward it. We're a small team of builders. We build for other companies. If anyone should find this easy, it's us.
It was humbling.
The method we settled on is brutally simple: give every task to AI first. Not the tasks AI is obviously good at. Every task. The point is not that AI will succeed. At the start it mostly won't. The point is that the attempt is a probe. Every failure tells you something specific and true about your company, and unlike consulting advice, it's never wrong. So the rule has a second half: when a task fails, write down why.
Do that for a few months and the reasons start repeating. Ours compress into four. The task was too big or too complex. The task needed context that exists only in someone's head. The task needed human judgment halfway through. Or the tool at the end only works through a screen made for humans, and AI driving a browser is still rough.
That's the whole list, and I've stopped reading AI strategy decks because this list is the strategy. Every line on it points straight at a fix. Too big: decompose, and find the parts AI can run alone. Context in heads: write the context down where AI can read it. Judgment in the middle: pull the judgment out into its own step, a human call between two machine runs. Human-only tools: choose tools with APIs, because a system that can only be operated by clicking is invisible to your most scalable worker.
What the failures really exposed was us. To make even one cycle work we needed things we had never written down: a shared memory of decisions and context, so everyone, human or AI, starts from the same page. Standards for what good output looks like, with the why attached. Half of what we tried to write down had never actually been decided, only done. We had been holding the company together with talent, which is another way of saying we had never been forced to define anything. That works, right up until you want the company to improve itself. Nothing held together by talent can compound. Talent goes home at night.
The real shift came when we flipped who starts. The default picture of AI at work is a human initiating and AI assisting. A copilot. We're moving to the opposite. The AI got our failure list as rules, so it can triage incoming work itself: take what it can run, prepare everything else as far as it can get, and queue the residue for us. Then we work behind it. We do the parts it left, the judgment calls and the hard bits, and hand the result back. The human stops being the bottleneck at the front of the queue and becomes the specialist at the back of it.
The aha moment isn't AI doing something impressive. It's the first time AI starts the task and you finish it, instead of the other way around.
And this is where the self-improving part stops being a metaphor. Every failure we write down widens the next cycle. The company is slowly growing a spec of itself: what we do, how we decide, what good looks like and why. AI reads the spec, runs more of it each month, and the failures keep pointing at whatever is still undefined. The learning doesn't evaporate when someone is busy or leaves. It accumulates. That's the asset.
Last week I argued you should automate last, and this probably sounds like the opposite. It isn't. Automate last is about commitments: don't build automation on top of a process you haven't questioned, deleted and simplified. This is about probes: handing AI a task costs almost nothing, and the failure is a free X-ray of what's undefined in your business. You probe first to learn where to transform. You automate last because transformation has an order. Same idea, different scale.
We're still in the middle of this. Cycles break every week, and I expect the failure list to grow a line or two before it shrinks. But the company is better defined this month than it was last month, and in twenty years of working the old way, that was never true. Maybe that's what AI-native actually means. Not AI everywhere. A company that learns about itself, and keeps what it learns.
Weekly notes from conversations with my business partner, where we share what we're seeing on the ground across clients and the market - not the hype, the real shifts.