Eating the application layer

Software implementation difficulty is a dying moat. What remains valuable is taste, distribution, and being right.

As AI improves, software is becoming cheaper to produce. Within knowledge work, it is one of the areas most exposed to this shift because it has several conditions that make AI improve fast: knowledge encoded in text, verifiable feedback loops, and expensive experts. This changes how software is built, and more importantly, the reasons that determine whether something is worth building.

For years, many applications captured value because building software required time, capital, and technical teams. Implementation difficulty acted as a barrier. A lot of mediocre software existed simply because it was expensive to make, and that cost protected it from defects like bad taste, poor usability, or weak performance. As that protection starts to fall, those weaknesses become exposed. The first impact lands on low-quality software because its main differentiator was that it existed, and there was no better alternative.

But implementation is getting cheaper across the whole curve. While bad software falls first because the bar for earning a place in the market gets higher, good software also becomes cheaper to produce because part of the value previously defended by development barriers gets compressed. AI amplifies the ability to build for everyone, including many people with taste, product judgment, visual sensitivity, deep understanding of workflows, and a clear vision. Those who previously lacked the technical skills to do it themselves no longer depend as much on others. That makes competition harder, because the barrier drops and brings stronger competitors into the market.

That said, building a good product remains extremely hard for non-technical reasons. Clones of successful products are a good example. They now appear faster and faster, and will continue to do so. Many of them will be technically robust too. But great products are not great because they are impossible to copy. Executing fast, while necessary, is not enough because it does not create users, reputation, or memorable products by itself. Today, for example, many users look at who uses a tool, who recommends it, or who is behind it before adopting it. Performance parity is no longer a winning strategy without those inherently human components.

In an era where production is cheap, taste, craftsmanship, and a clear view of what is and is not worth doing matter more. What to build, what not to build, which details to care about, how a digital product should feel, and what human behavior is actually tied to its use are factors that depend on the person building. The application layer will suffer from its own defects more than from AI making it unnecessary. First, bad software falls. Then, the price of good software drops. Finally, the products that endure change shape, because AI enables new forms of use. The most durable moat becomes a deep understanding of what a better state of things looks like in relation to the problem being solved.