90% of real grunt work is "stitch an extra appendage to this unholy abomination of God" or "*points at the screen* look at this shit, figure out why is it happening and fix it". LLMs are more or less useless for either of those things.
It's not really possible for an LLM to pick up on the hidden complexities of the app that real users and developers internalize through practice. Almost by definition, they're not documented! Users "just know" and thus there is no training data to ingest. I'd estimate though that the vast, vast majority of bugs I've been assigned originate from one of these vague reports by a client paying us enough to care.
or this about MS pushing for more internal AI usage, and the resulting hilarity (or tragedy, depending if you are the one having to read the resulting code)? https://news.ycombinator.com/item?id=44404067