Picture someone who runs a small photography studio. They're not a developer. They're tired of scheduling shoots over chat, so they spend a weekend describing a booking app to an AI tool. By Sunday night they have something with a calendar, a login screen and a payment button.
They should be proud of it. A few years ago that would have cost months and a developer's salary.
Now picture three weeks later. Clients keep booking the wrong slots because the calendar shows every hour of the day with equal weight. Rescheduling takes four taps in a flow that makes sense to nobody but the model that generated it. The owner quietly goes back to chat.
Nothing in the app is broken. It does everything that was asked for. It just isn't pleasant to use, and it's hard to say exactly why.
Some version of that story is playing out everywhere right now, and it says something about where software is heading.
When making becomes cheap
There's a pattern that repeats every time a technology makes production cheap.
When desktop publishing arrived in the 1980s, anyone could lay out a newsletter. Suddenly there were millions of newsletters, and most of them were hard to read. Typography didn't become less valuable. It became more visible, because now you could compare the good ones with the bad ones side by side.
When smartphone cameras became good enough for most people, everyone became a photographer. There are more photos taken every day than ever before. The photographers who still get paid aren't the ones who own cameras. They're the ones who know what to point them at, and what to leave out of the frame.
Software is going through the same thing now. The term "vibe coding" spread in 2025 to describe exactly what our photographer did: describing an app in plain language and letting AI write it. Production, the part that used to take the most time and money, is getting close to free.
The studio owner in our example didn't fail at building. They built exactly what they described. What they couldn't describe was the part that makes a product feel right.
So the question isn't whether AI can build software. It can, and it will keep getting better. The question is what becomes valuable once building is no longer the hard part.
The short answer: making got cheap. Taste and structure got expensive.
Economists have a name for what happens next. In 1865, William Stanley Jevons noticed that as steam engines burned coal more efficiently, Britain used more coal, not less. Cheaper production doesn't shrink a market. It floods it. The same is already happening with software: when an app costs a weekend instead of six months, the number of apps explodes, and standing out among them gets harder, not easier.
What the research says about AI and speed
It's tempting to assume AI makes everyone dramatically faster. The evidence is more interesting than that.
In a 2023 experiment, researchers from Microsoft Research, GitHub and MIT gave developers a well-defined task, writing a small web server, and found those using GitHub Copilot finished about 55 percent faster than those without it. For clear, contained tasks, the gain is real.
In 2025, the research group METR ran a different kind of study. It asked experienced open-source developers to work on real issues in large codebases they knew well, with and without AI tools. The developers expected AI to speed them up, and afterwards believed it had. Measured, they were about 19 percent slower. Time went into prompting, reviewing and correcting output that looked plausible but didn't quite fit.
Both findings can be true at once. AI is excellent at producing code for a well-described problem. It's much less reliable at fitting into a complex, existing context where judgment matters. Most real products are the second kind.
Right in structure, wrong to use
The most common problem with AI-built interfaces isn't that they fail. They usually work. Every feature is there. Every button does something.
The problem is that they're unpleasant to follow. Everything on the screen has the same weight, so the eye doesn't know where to go. Every piece of information gets its own label and its own column, because the model has no reason to leave anything out. It's a bit like a report where every sentence is in bold. Technically, everything is emphasized. In practice, nothing is.
A good interface is mostly a series of decisions about attention:
What should the eye find first? On a customer list, that's usually the name, then a short summary of the customer's situation.
What can be simplified? A full sentence like "RFID card registered" can often become a small dot next to a photo, with the explanation on hover.
What can be removed? The information that's technically available but rarely needed, and only competes with what matters.
That last point surprises people. A summary line on a list doesn't only look tidier. It reduces mistakes, because staff can see someone's status at a glance instead of opening each record. Good hierarchy isn't decoration. It changes how people work.
Those decisions are exactly what AI tends to skip. Its default is to show everything, because showing everything is never technically wrong.
Why "feels right" is worth money
It's easy to dismiss look and feel as decoration. Research suggests it's closer to function.
In 1995, two researchers at Hitachi's design center, Masaaki Kurosu and Kaori Kashimura, tested people's reactions to different layouts of the same ATM interface. Participants rated the more attractive layouts as easier to use, even though the functions were identical. The finding became known as the aesthetic-usability effect, and it has been replicated in various forms since: people are more patient with, and more forgiving of, products that feel well made.
There's also the question of first impressions. A study led by Gitte Lindgaard in 2006 found people form a judgment about a web page's visual appeal in about 50 milliseconds, and that the judgment tends to stick. Before anyone reads a word, they've already decided whether the product feels trustworthy.
That's why a technically complete app can still fail. People don't experience software as a list of features. They experience it as a feeling, formed in a fraction of a second and confirmed or broken over the following weeks.
Structure is the flow, not the feature list
When people start building software, they usually think in features: a calendar, a payment page, a customer list, a report. Features are easy to describe, so they're easy to ask AI for.
What's harder to describe is the flow: the order in which someone moves through the product, and what they see at each step. How does a new customer go from signing up to paying for the first time without getting lost? Which two steps should really be one? Which step should come first because it prevents errors later?
A feature list can't tell you that. A model will happily build every item on the list and connect them in a way that is logical to a machine and confusing to a person. The four-tap rescheduling flow in our example isn't a bug. It's a flow nobody designed.
"Why not just build it myself?"
This is a fair question, and anyone who works in software should be able to answer it without getting defensive.
Building it yourself is a real option now. For a small internal tool, a prototype or an idea you want to test, it might be the best option. But it helps to know what you're taking on.
Professionals who use AI every day still spend much of their time correcting its output. They adjust layouts, reorder flows, remove what shouldn't be there and fix the parts that work but feel wrong. That judgment comes from years of watching real people use real products. If people who do this for a living still need to correct the machine, someone starting from zero has to learn all of it first.
In other words, you're not only paying for code. You're skipping a year or two of learning what it takes to build and run a product people actually want to use.
There's a good parallel in photography's own history. When Instagram launched in 2010, anyone could apply a filter and make a phone photo look like film. Professional photographers worried they'd be replaced. What happened instead was that the value of photography moved. Taking a technically decent picture became trivial. Knowing what to photograph, how to light it, which one to choose out of fifty and how to make a set of images feel consistent became the thing people paid for. Software is going through its own filter moment.
What AI can't do on its own
Could AI build an entire product from start to finish? It can get surprisingly far. But the further it goes without direction, the more visible its mistakes become, because there's nothing behind them to catch them.
AI can be told what to do. It can't decide by itself what should be done.
That's not a limitation that better models will simply remove. Direction depends on knowing who the product is for, what matters to them, what the business is trying to achieve and what to leave out. Those are choices, not calculations. A model can suggest options. Someone still has to pick one and take responsibility for it.
The counterargument
It's worth taking the other side seriously. Won't taste become automated too? Models are already trained on millions of well-designed screens.
Partly, yes. The average quality of generated interfaces will keep rising. But that's exactly the point: when the average rises for everyone, the average stops being an advantage. If every new product looks competently generic, the ones that stand out will be the ones with a clear point of view, a flow designed around a real job and a visual identity people recognize. Those are the parts you can't get by asking the same tool everyone else is asking.
What's still worth paying for
If you're deciding whether to hire someone or build with AI yourself, these are the parts that still carry the value:
Direction. What the product is, who it's for and what it should deliberately leave out.
Flow. The order people move through the product, decided on purpose rather than generated.
Hierarchy. Where the eye goes first on every screen, and what gets removed so that can happen.
Identity. A visual direction people recognize, in a market where much of what gets generated looks the same.
Experience. Knowing what breaks once real people use a product, before your customers find out for you.
The code is the part that got cheap. Everything on this list didn't.
Build it, then ask the right question
If you're building something with AI this weekend, go ahead. It's one of the best ways to learn what you actually need. When you're done, don't ask "does it work?" It almost certainly will.
Ask "would someone who isn't me enjoy using this every day?" If the honest answer is no, and you can't say why, that's the gap where design lives.
Layerice is a software design and development studio for businesses. We use AI every day, and we still decide where the eye goes first.