I rolled my eyes at 'AI-powered everything' for a year. Then I found the arrangement that works.

For about a year I rolled my eyes at "AI-powered everything." Most of what I saw deserved it: demos that fell apart under real input, confident output nobody verified, tools sold as replacements for judgment they did not have.

What changed my mind was not a better model. It was finding the arrangement where the division of labor is honest. The model drafts. Guardrails check. I decide.

Each word in that sentence is load-bearing.

Drafts means the output is a starting point with no authority of its own. A draft can be excellent and still be wrong in the one place that matters. Treating generation as a draft, rather than an answer, removes the failure mode where confidence gets mistaken for correctness.

Checks means the verification is mechanical, not vibes. Type checks, tests, contract checks, lint gates, the same machinery that catches human mistakes. A guardrail that depends on someone remembering to look is not a guardrail. In my setup, generated work passes through the identical gates handwritten work does, and the gates do not know or care who wrote the diff.

Decide means one human owns the outcome and can reject the whole thing. Not review-as-rubber-stamp. Actual authority to say no, backed by enough understanding to know when to use it. The moment the human stops being able to evaluate the draft, the system has failed, no matter how good the output looks.

The skeptics are right about the version without the last two steps. Generation without verification and ownership is how you ship confident garbage at scale. But the full arrangement is just a faster version of how good engineering teams already worked: someone drafts, the system checks, someone accountable decides.

I did not become a believer in AI. I became a believer in that sentence.