These are usually internal tools, workflow improvements, and one off features. Anything really central to the game’s code gets human coded.
I think the further you are from the idea part, the less fun AI coding will be for you. Because now you need to not just translate some spec to code, you have to translate it to a prompt, which ups the chances of playing the telephone game. At least when you write the code yourself you are getting real with it and facing all the ambiguities as a matter of course. If you just pass it to an LLM you never personally encounter the conflicts, and it might make assumptions you would not… but you don’t even realize it because they are assumptions!