I Started Building Things With AI. Here's What Nobody Told Me.

For most of 2024, my relationship with AI was enthusiastic but passive. I used it every day, had strong opinions about which models I preferred, gave recommendations freely, and wrote about AI strategy with some confidence. What I did not do was build anything. That changed at the start of 2025, and it changed more of what I thought I understood than I expected.

The catalyst was a specific, unglamorous problem: I needed a tool that could take a research brief and return a structured competitive analysis. Nothing off the shelf was quite right for what I needed. I decided to build it rather than keep looking, which at the time felt like a small decision. It was not.

The first thing nobody tells you is how fast the early stages go. I had a working prototype in two days. Genuinely working: accepting input, calling the API, returning structured output. For someone without a software background, that felt like a minor miracle. Claude wrote the code. I described what I wanted, reviewed what came back, asked for changes, and iterated. The feedback loop was fast enough to feel like a real conversation.

The second thing nobody tells you is how quickly you hit the ceiling of what AI can do without you thinking harder first. Once basic functionality existed, the questions that mattered were no longer technical. They were: what should the output structure actually be? What sources should it draw on? How should it handle topics it does not know well? When is the output reliable enough to trust, and when is it wrong in a way that matters? Those questions required my judgment, and the quality of the tool ended up being directly proportional to how clearly I had thought them through in advance.

I made a significant mistake early on that cost me about two weeks of unnecessary iteration. I let the prototype define the product. Because the first version worked, I kept building on top of it instead of stepping back to ask whether the structure was right for what I actually needed. This is not an AI-specific mistake. It is a builder mistake that AI makes faster, and therefore more expensive. The speed is real. Bad architectural decisions made quickly are still bad architectural decisions.

What I discovered by July is that building with AI had changed how I think about problems, not just how fast I can solve them. The cost of starting something has dropped so dramatically that I test ideas I would previously have dismissed as too complex to attempt. The risk of beginning is now almost zero. The risk of building the wrong thing remains exactly what it has always been.

"AI makes starting easy. It does not make building the right thing any easier. That part is still entirely on you."