Lately, I’ve been playing around with vibe coding, trying to build a few app ideas I’ve had in my head.

The appeal is pretty obvious. You don’t need to know that much about coding anymore to turn a business idea into something that actually works. You can describe what you want, iterate on it, and within a surprisingly short amount of time, have a prototype in front of you. A few years ago, I probably would have needed to find a developer, learn much more coding myself, or simply leave the idea as an idea.

So yes, the cost of building has gone down dramatically.

But after spending more time with these tools, I’ve started to wonder whether the cost of operating has really gone down in the same way.

For someone who doesn’t come from a strong coding background, there is still a lot of trial and error. Which vibe coding tool should I use? What is each one actually good at? How should I communicate with it? When it gives me something that doesn’t work, is the problem my prompt, the model, the platform, or something underneath that I don’t understand?

And sometimes, you have to pay before you can even answer those questions.

None of this changes the fact that building is cheaper than it used to be. If the total cost was once “building + operating,” and the building part becomes much smaller, then of course the total cost comes down too. But I think there is a difference between making something cheaper and removing the cost altogether. Some of the cost has simply moved somewhere less obvious.

I had been thinking about this even before I started vibe coding.

A while ago, I became curious about a question that sounds like it should have a very straightforward answer: How much has AI actually improved productivity?

I did quite a bit of research and went through reports from large consulting firms. There were plenty of numbers: productivity improved by X%, costs reduced by Y%, tasks completed Z% faster.

But I kept looking for something much more concrete.

Take the same project, with roughly the same scope and the same team. Before AI-assisted coding, how long did it take to launch? After AI-assisted coding, how long did it take? How many hours did people actually spend from beginning to end?

I had a surprisingly hard time finding comparisons at that level.

I recently talked about this with a friend who works in consulting, and her experience made me think about it even more.

AI has made slide-making dramatically faster for her. Something that might have taken two days before can now be done in half a day. That sounds like an enormous productivity gain. And because the initial work takes less time, she can work across more projects at once.

But there’s another side to it.

The slides still need to be reviewed. AI-generated content needs to be checked. Numbers, wording, logic, and sources need to be verified. And when you can produce more, there is simply more to review.

So how much more productive is she, really?

Probably more productive. But 20% more? 50%? Three times as productive? It becomes surprisingly difficult to say.

And I think this is where our perception of AI productivity gets interesting.

A lot of what AI removes is visible work.

Writing code is visible. Spending hours formatting slides is visible. Drafting a paragraph from scratch is visible. Searching for information manually is visible. When AI takes those things away, we feel the difference immediately. The workday feels easier. We spend less time staring at a blank page or figuring out how to make something from scratch.

But AI also creates a layer of invisible work that is much easier to overlook.

Choosing the right tool. Learning how to use it. Rewriting prompts. Checking outputs. Catching hallucinations. Fixing things that are 90% right but strangely difficult to make 100% right. Paying for tools that turn out not to solve the problem. Producing more because we can produce more, and then spending more time reviewing everything we produced.

Maybe this is why “AI makes me more productive” can feel obviously true on an individual level while being surprisingly difficult to measure precisely.

We tend to notice the work that disappeared.

We’re much worse at noticing the work that quietly replaced it.

And maybe that’s the part of the AI productivity conversation I find most interesting right now. The question isn’t simply whether AI saves time. It clearly can.

The harder question is: where did that time go?