My boss asked me how much faster AI makes software development.

I said it depends.

He said:

“No, give me a percentage.”

So I gave him the most dangerous number a developer can give management:

“30%.”

Now apparently we can finish projects 30% faster.

The AI writes boilerplate quickly.

It generates tests.

It explains ancient code nobody wants to touch.

It writes SQL.

It converts one API format into another.

It even occasionally suggests a solution I wouldn't have thought of.

Great.

But then I spend 20 minutes reviewing a 30-second AI-generated function because the function is technically correct but completely wrong for our architecture.

Then another 15 minutes fixing an API call that uses a deprecated parameter.

Then another 10 minutes explaining to the AI that our database does not, in fact, contain a table called `users_archive_final_v2`.

Then I discover the AI confidently generated a library version that doesn't exist.

So yes, AI makes me faster.

But not in the way management thinks.

It doesn't remove the work.

It changes the work.

I write less code and review more code.

Which is probably fine.

Except now I have to review code written by someone who can produce 500 lines before I've finished my coffee.

Congratulations.

We automated typing.

We did not automate being responsible for the result.