I read an interesting breakdown of an experiment at a startup recently.

They put together an outstaffed team, gave them AI tools, and set them a task: build a new web module.

4 months. $60k.

Fast and cheap, or so it seemed.

The in-house team at the same startup averaged 9 to 11 bugs per quarter on a comparable codebase.

The vibe-coding team landed at 419.

Not a typo.

But there was one detail that got to me more than the numbers.

The subsidized illusion

The CEOs of Copilot and OpenAI have said publicly that current subscription prices are roughly 3x below actual cost.

Companies are deliberately subsidizing access, building habits, and capturing the market.

Not a secret. Just a strategy.

So that $60k in the experiment was $60k at subsidized prices.

At real cost, you are looking at $180k to $200k.

Factor in the bugs, rework, and QA load, and suddenly you are at $300k for six months.

My version of the same problem

I am not a developer.

But I have my own version of this, just without a dashboard to show it.

I ask AI to write a post, get five options, read all five, pick “this one but change that,” read it again, and then rewrite it myself anyway.

And there is this feeling of having worked, even though starting from scratch would have been faster.

Marketing has no “419 bugs” metric.

So it is easy to miss the point where AI starts costing more than it saves.

What to measure instead

The good news is that this is measurable.

Not perfectly, but measurable.

• How many hours a week go into reviewing AI output?
• How many iterations happen before a final version?
• How many of those do you rewrite anyway?

Multiply that by your hourly rate and compare it to what the subscription costs today.

Then multiply that subscription by three and see if the decision still looks the same.

AI might still come out ahead.

But at least then it is a real decision, not an illusion of cheapness.