The hard part of an AI build is no longer the building. When almost anyone can ship something impressive in a weekend, the advantage moves to the judgment that comes before it: the call on whether a given build is worth doing at all. That judgment is what separates the AI work that shows up on the P&L from the work that only looks good in a screen recording, and most of what gets celebrated is the second kind.
A broccoli farmer in Hokkaido is the cleaner example, and it is worth slowing down on why. Hiroki Tomiyasu manages around a hundred hectares of broccoli and other vegetables in northern Japan. He has no engineering background; by his own account he came to farming from outside the industry, arriving in Hokkaido through Japan's regional revitalization corps in 2020, and he started building software with everyday AI tools from zero coding experience. People would screenshot his self-steering tractor, but the part worth studying is what he did before he touched it.
The feed rewards the build, not the result it never reports
What gets celebrated online is the artefact and the speed of making it. The video shows the agent replying, the dashboard refreshing, the workflow firing without a human in the loop, and the implied conclusion is that the work is done and the win is banked. The question that would actually settle it goes unasked almost every time. Did the build move revenue, free real capacity, or take cost out of the business. Anything, in other words, that shows up where the company keeps score. A build can be technically clever and still touch none of those, and most of what gets glamorized lives in exactly that gap: impressive to watch, invisible on the P&L.
This matters because the gap is still widening. As building gets cheaper, the volume of cool-looking builds goes up, and the share of them that earn their keep does not move with it. The skill that used to be scarce, the ability to make the thing, has now been handed to almost everyone. What stays scarce is the judgment to decide which thing is worth making at all, and that judgment is the only part the tooling cannot do for you.
Why the farmer got it right: he ran the math before he touched the tooling
The auto-steer is the proof. A self-steering tractor that holds a precise line across a field is the kind of capability farm-equipment makers sell as a proprietary upgrade, and the proprietary route is expensive. Tomiyasu started by understanding what the capability actually required, using AI to learn how the underlying positioning worked, and then he weighed what that capability was worth on his fields against what reaching it would cost him in effort and money. The math came first. The code came second, and it was the easy part by far. Only after that did he build, using open positioning hardware instead of buying the packaged version, and by the accounts he has shared the self-built rig ran a fraction of what the proprietary upgrade would have cost. He has described simple programs becoming usable in well under an hour.
That sequence is the entire difference between his auto-steer rig and the agent someone shipped to applause last week. He has built more since, including a greenhouse he can monitor and adjust from his phone, opening and closing vents and managing temperature remotely, and the same discipline runs underneath all of it. He treats AI as a way to lower the barrier to automation for an operator who has no tech team, which means every build has to clear a bar before it gets made. He is the worked example here, the operator who shows what the discipline looks like when it runs.
The three questions that separate a real build from an expensive hobby
The judgment that looks like instinct from the outside is actually a short sequence anyone can run, and it sits entirely before the first line of code. Three questions decide whether a build is worth doing, and they have to be answered in order, because each one only matters once the one before it has cleared.
What can AI really do here. The honest answer is about this exact task, on this data, inside this messy reality, once the easy 80 percent is working and the awkward last stretch starts fighting back. It is rarely the version that runs in a launch video or the one that worked for someone whose situation looks like yours from a distance. That ceiling is usually narrower than the feed suggests, and getting it right means resisting the pull of the impressive demo long enough to ask what the capability actually tops out at in the hands you have.
What would it take to get there, in effort and money. A working demo and a build you can rely on every day are separated by a long, unglamorous middle, and that middle is where most of the real cost lives. Pulling the thing the last distance to dependable, maintaining it when the inputs shift, the attention it pulls off other work. The cost lives in the time and judgment to make the capability hold up under load rather than in the licence fee, and that is the number people skip when they are excited about the demo.
Do the two line up. This is the question that does the actual work, and it is the one the feed never reaches. Lay the honest capability next to the honest cost and see whether the value clears it by enough to be worth the attention. When the two line up, you build with conviction, because you already know what it is worth and what it will take. A build whose value never clears its cost is a hobby, the kind of thing that earns applause online and quietly costs more than it returns. Tomiyasu ran exactly this on the auto-steer. The capability was real, the self-built cost was a fraction of the alternative, and the two lined up cleanly, so he moved. Most of what gets posted would not survive the third question, which is precisely why so few people ask it out loud.
No business is too unique for AI: the work is finding leverage instead of chaos
The reason this judgment travels is that the technology has crossed a threshold. It is good enough now that it almost always fits somewhere useful, in operations no one would call a tech company, on tasks no software vendor ever bothered to package. A humanities-background farmer reading positioning hardware off an AI conversation makes the point better than any case study from a software firm: if it fits a hundred hectares of broccoli in Hokkaido, the "we're too specialized for this" reflex is mostly a story a business tells itself before anyone with the right eye has actually looked. The fit is rarely the constraint anymore.
What is genuinely scarce is the expertise and the time to find where AI brings leverage and where it just brings chaos, and those two outcomes sit closer together than they look. The same capability that compounds when it is pointed at the right work will multiply mess when it is pointed at the wrong work, automating a broken process faster, generating output nobody trusts, adding a layer of tooling the team quietly routes around. Telling the two apart is the work, and it is the same work as the three questions, applied across a whole operation instead of a single build. Almost no business is too unique for AI. Plenty of businesses are one honest look away from finding the few places it would actually pay, and a longer list of places it would only impress.
The uncomfortable question worth sitting with
Run the test backward across the last year. Of every AI build a company has shipped, admired, or forwarded to the leadership chat, how many would survive being asked whether they showed up anywhere the business keeps score. Whether they were clever, fast to make, or fun to demo barely matters next to whether revenue, capacity, or cost moved because they exist. For most of those builds the honest answer is that nobody ever checked, because the feed rewards the screenshot and the screenshot never reports the result. The building has become the easy part, which moves the advantage away from whoever builds the most and toward whoever can tell, before they build, which of the impressive things in front of them is actually worth doing. That judgment is the part you still have to bring, and no model arrives carrying it for you.