Domain Experience Is the Moat

Over the past few days, two pieces of news made the rounds among people who work in software. The first was about an American developer who used Claude Opus 5.5 to rebuild seven Adobe apps, Photoshop among them, and released them for free. He says himself that they're rough, still alpha versions, and he didn't copy any code. He watched how the apps behaved and built new ones from scratch. Still, it was enough to show how far the cost of making a convincing substitute has fallen. The second was an open-source project called REA that hands reverse engineering, taking a finished program apart to figure out how it works, to AI agents. Work that used to take a skilled expert weeks now gets done by an agent driving the existing analysis tools.

Watching all this, I figured that since software would all end up cloned through reverse engineering, the answer might lie in content. That idea lasted about a day. On Friday afternoon at a Bloom Mixer (Bloom is the AI community I run in Seoul), someone told me that if you paste in a video link from a channel that's doing well and ask the AI to change only the topic, you get back almost exactly the same format. When they let it handle the editing too, the AI skipped burning tokens and just built its own editing tool and used that. I had thought content would come after software. Content was already headed down the same road.

Now that anyone can make or copy anything with AI, what moat keeps competitors from simply walking in? Talking with the people I met that week, the answer kept coming back to the same place: problems only someone who has lived through them on the ground would know about, and data only that person has.

We just... had no ideas

An older friend who runs a startup told me he'd spent his whole lunch that day talking about exactly this with the developers on his team. His team had decided to try something besides their current product, and everyone was supposed to bring in one new idea. In the end nobody brought anything. His verdict: "We really have no ideas."

I've thought something similar for a long time. People say ideas have no moat in the age of AI. As I see it, the problem is that there are no ideas to begin with. If everyone had the same ideas, whoever could build cheapest would have taken the whole market. Silicon Valley keeps taking it anyway, because it has the ideas, the planning, and the instinct to notice first what's worth building. We're always saying, "Oh, I should have built that," and yet we're never the ones who build it first.

(I wrote more about ideas themselves being the moat in an earlier post, Flowers That Bloom in the Mud.)

My friend said his developers had been feeling pretty demoralized lately, watching people use AI to take someone else's program apart and rebuild it in no time. They felt their own product could be rebuilt just as quickly. Here's how he summed up where the team landed that day:

Honestly, building software with AI and selling it is a dead end. It's better for someone with the domain knowledge to build it themselves.

In other words, the person who knows the work best can now build what they need with AI and use it themselves, and that beats a company that makes software and sells it. There was a colleague on his team who used to run the market-size numbers with AI first and then say the market was too small. That person has now completely changed their mind on their own. A company with a 15% operating margin is a really good company, they said, and the old math, where software has no costs beyond salaries and servers so you should keep 70%, just doesn't hold up anymore.

We built it, took it out, and hit a different problem

My friend's current product didn't come from sitting at a desk either. By chance he ended up meeting people who work in that industry and heard about a hassle they were dealing with. It wasn't hard to fix, so he went home that day, built something, and showed it to them, and they paid for it on the spot. Since then, whenever he has time, he goes and knocks on the doors of people like them. About one in ten will sit down and tell him their story.

He came to the Friday Mixer too, and told me one more story there. He took something he'd built out to the people in the field and asked them to try it. What came back was a completely different request: that's not the real problem, so skip that and solve this instead. That's how far apart the problem a builder guesses at from a desk can be from the problem someone in the field runs into every day.

So these days he gets stuck on planning, not development. He handles the planning and the design and does the sales too, and in the meantime the development is done in a few hours, so planning can't keep up. He isn't from that industry, so he has to hear directly about the hassles people there face before he can decide what to build next. On days when that runs late, he becomes the bottleneck for the whole team. He said the final polish is still better left to a developer, though. At first it seemed fine for him to do it himself, but the finished result was clearly better when a developer refined it.

There was a gap on the other side too. When he looked into competing products, he found that within a month or two, several products built by people from that industry had appeared. They had every feature you'd need, but they looked way too obviously AI-made and were awkward to use. Knowing a problem precisely and building a good product that solves it are still separated by a clear gap.

The dentist canceled his subscriptions

That same week, a post by someone who crossed that gap alone went viral. It was "A Neighborhood Dental Clinic's AX Transition" (AX meaning AI transformation), posted on Arcalive (a big Korean online forum). A local dentist who majored in biology used Claude and Codex like a team of staff, each with its own role, and ended up running everything inside his clinic himself, from answering the phone to checking insurance paperwork, management reports, and an internal messenger. Along the way, he canceled one monthly subscription after another, for things like call recording, document scanning, and appointment texts, and he dropped his accounting service and his ad agency too.

Read the original post on Arcalive (in Korean)

Developers who read the post said he must be really good with AI. I saw it differently. An AI expert couldn't have built this. He could build it because he's a dentist. When I brought this up at the Friday Mixer, one person said that above all, the dentist is someone who can define the problem.

Someone I met at the lunch Mixer said the same thing. Only people who know where problems actually keep piling up can automate them, so there's a gap between AI experts and the people doing the work. This person also said the daily repetitive tasks drove them crazy and they wanted to automate them, and the goal was to make it possible to run the company's back office alone. We even talked about how a system built that way could be sold on its own.

The dentist's own conclusion in his post said the same. Someone who didn't study tech can build things like this, but deciding what to have it do takes the person who knows dental work best, the dentist. He wrote that AI doesn't so much replace people as take away the work people hated. Staff only need to look at the items the AI has flagged as needing a check, people still make the decisions, and the AI is trusted only with reading, organizing, and making suggestions.

There was also a passage about how spending a lot on ads creates a setup where you have to get more out of each patient to cover it, and then you end up recommending treatment people don't need. He said that thanks to AI he could cut his ad spending, and that brought him closer to running a clinic that can tell patients when they don't need a treatment. He wrote that AI will probably end up doing all the office work that deals with numbers, but it still can't do the hands-on work inside someone's mouth. He also said the biggest bottleneck isn't technology but people. Many of his patients are older, so there are still times he has to take paper charts.

He said that within two years anyone with an idea will be able to build things like this, and he plans to release most of what he built as open source for other dentists. He noted that he wrote that long post with AI too. It still got close to 300,000 views, and the reason wasn't smooth writing. It had stories only he could tell, like the server rack that went into the clinic's back room, or the staff member who watched the insurance paperwork reconcile itself and asked, "So am I getting fired now?"

The moat is data no one else can make

I asked the same question on a call with an old friend in the US. The way my friend sees it, the moat in the AI era is data nobody else can produce, or a thought nobody has had yet.

You need new data or a new idea.

My friend also said they had no idea what people are basing it on when they say AI will solve everything within a few years. What AI helps with is picking out candidates worth trying, faster and in bigger numbers. Checking whether those candidates actually hold up still has to happen in the real world, and it takes time.

It's not going to run the experiments for you. In the end, people have to do it.

That's also where my friend wants to go next. When someone has a problem they want to solve, my friend wants to be the person who looks into the data that person has and the way they work, and then makes it possible for AI to take over that process.

What I heard at the Friday Mixer pointed the same way. If AX (AI transformation) at a company that makes physical products goes all the way, you end up at a stage where you describe something and a product comes out. Your words become the blueprint, and the blueprint goes straight into production. But ahead of that sits a huge piece of homework: field data that someone has to deliberately collect and organize.

That's why people say Korea's physical AI still isn't competitive. For physical AI to really work in a factory, you first have to turn everything in the factory and every other site into data, and they say we're nowhere near ready for that. Physical AI will probably go down a similar road someday, the way software and content did. When it does, the ones left standing will be whoever built up the data from their own field first.

Originally published in Korean on Brunch, October 17, 2026.