Tedious, repetitive stuff that you'd typically write a throwaway script for is one good use case for AI for me. This doesn't come up very often for me, however. But this morning I needed to update a bunch of import statements in TypeScript files to use relative paths instead of referencing a library name and Cursor did that for me quickly and easily without me needing to write a script.
I've also found that if you're unsure how to solve a problem within an existing codebase, or you're about to go down a rabbit hole of studying documentation to figure out how to wire something up, an LLM can get you up and running with code examples a bit quicker.
But if I already know, high level, how I'm going to solve something then the whole "needing to review the AI code" part of the workflow more than eats up any time savings. And the parallelism thing is just a ton of context shifting to me. Each time I need to go back to an agent to review what it did, it takes me away from the deep focus on a domain problem that I was focused on, and there's a good 20 - 30 minutes of productivity lost by switching back and forth between just two tasks.