AI can write the feature. But who decides when the feature is actually finished?

I've noticed a strange shift in my own workflow since using AI more heavily for development.

Before AI coding assistants, "done" was usually tied to implementation:

I wrote the code → tested it → reviewed it → shipped it.

With AI, implementation has become much less of a bottleneck.

The agent can generate the feature.

• It can fix the error.

• It can write the tests.

• It can refactor the code.

• It can even explain what it changed.

But I've started noticing that there's another question hiding underneath all of this:

When is the feature actually finished?

A feature can be technically complete and still be wrong.

For example:

• It solves the wrong user problem.

• It adds complexity nobody needed.

• It works in the happy path but creates awkward edge cases.

• It technically matches the specification but not the product experience.

• It introduces another workflow that users now have to understand.

• Or it simply shouldn't have been built in the first place.

That's made me think that AI coding is shifting part of the developer's job from "How do I implement this?" toward "Should this exist, and what does good actually look like?"

The interesting part is that an AI agent can probably help with the first question much more reliably than the second.

So I'm curious about people building with AI:

What is your definition of "done" for an AI-generated feature?

Is it passing tests?

Your own code review?

A real user successfully using it?

Or something else?