I find AI helpful when coding in unfamiliar languages or topics, as it generates initial code that roughly works, that I can then iterate on.
Even so, I have to constantly hound the AI to write concise code, to reuse code, to consolidate duplicate code blocks. When I ask it to remove the useless comments, it also removes the useful comments, so I have to goad it back into adding back individual helpful comments.
I had AI write some Python unit tests and it showed me how to mock, which I had never done in Python. That's great! But when I examined the tests, they were so "white box" and brittle that almost any change in implementation would break the tests.
When coding in a familiar language (C++), I tried turning on the auto-AI assistant (for our one project where security rules allowed it) and while I was impressed that it would auto-complete whole blocks of text based on my actual code base, not once was I able to accept those blocks as-is.
So for me at this point AI is at best a net +2% productivity improvement, though I surely have lots to learn about other ways in which it could be useful.