The Best Thing About Building a Startup with AI Is Also the Most Dangerous Thing About It.

A few years ago, if you had an idea for a product, you needed months and a significant amount of money before you could show anyone something real. You had to hire engineers, write specifications, manage a development process, and wait. A lot of ideas died in that waiting period, which, looking back, was not always a bad thing.

Today, someone with no technical background can describe an idea to an AI tool and have a working product in a few weeks. The cost has dropped from six figures to something closer to what you might spend on a used car. A quarter of the startups that joined Y Combinator, one of the world's most competitive startup programs, in early 2025 had products that were almost entirely built by AI.

I have been investing in companies for more than twenty years. This shift is the most dramatic change to the early stages of company-building I have seen. And it comes with a problem that most people are not talking about clearly enough.

Making the product is no longer the hard part.

When starting a company required months of expensive engineering work, founders had no choice but to think carefully before building. You had to talk to potential customers. You had to be reasonably sure people wanted what you were making before you spent the money to make it. The difficulty of building was, in a strange way, a quality filter.

That filter is largely gone. You can now build something and put it in front of users in a matter of weeks. That sounds like progress. In many ways it is. But it also means you can spend weeks building something, show it to the world, and discover that nobody actually wanted it.

Building the product used to be the obstacle. Now the obstacle is figuring out what to build in the first place, and that question has gotten harder, not easier, because fewer people are being forced to answer it before they start.

The companies that are thriving figured out one thing early.

Cursor is a software tool that reached $2 billion in yearly revenue with a team of fewer than 200 people. The founders did not start with that product. They started with something else entirely, an AI tool for a different industry, tried it, looked honestly at whether it was working, and changed direction. The insight that led to Cursor, that people who write code needed AI built into their entire working environment rather than added as a small feature, came from that honest assessment. The building came after the thinking.

Lovable is another example. It reached $400 million in yearly revenue with 146 employees by focusing on one very specific type of person: someone who wants to build a digital product but has never learned to code. The founders decided who they were building for before they built anything. That clarity is what made the product work.

Both companies used AI to build faster. They both spent significant time before that on a much older question: who needs this, and why?

The numbers tell a consistent story.

MIT researchers studied 300 companies that tried to build AI-powered products. Their finding: about 95% of those efforts failed to produce any meaningful return. Only 5% worked in a way that actually changed the business.

Sequoia, one of the most respected investment firms in the world, looked at how people actually use AI products after they download them. The average AI product keeps about 14% of its users coming back daily. Compare that to the apps most people use every day, which typically keep 60 to 70% of users coming back. Most AI products, even ones that get a lot of initial attention, struggle to become genuinely useful habits.

Research tracking why startups fail has consistently found the same answer over many years: the number one reason is not running out of money or having a bad team. It is building something that nobody actually needed. AI makes it faster and cheaper to build the wrong thing.

What has stayed the same across every wave of new technology.

In twenty years of backing companies, I have met founders who were early to the internet, early to mobile, early to cloud computing, and now early to AI. The ones who succeeded shared something that had nothing to do with the technology they were using.

They understood a specific problem that real people had. They could describe those people clearly. They knew why existing solutions were not good enough. And they could explain, honestly, why they were the right person to work on it.

AI is a remarkable tool for executing on a clear answer to those questions. Finding the answer is still on you, and it still requires the same things it always has: talking to people, questioning your own assumptions, and sitting with a problem until you genuinely understand it.

The founders I work with who use AI well treat it as a way to move faster once they know where they are going. They still do the slower, less glamorous work first.

That part has not changed. It was never the part that was supposed to be easy.