As software grows the problems go such that the effort to completely describe the necessary changes in plain language is often not all that much shorter than the code. Especially if you have a really really good autocomplete for the boilerplate parts of the code. So LLMs being really good at the tedious autocomplete makes LLMs have less marginal utility for the "write the whole thing" for certain types of work.

But the sneakier part of the problem is that as business rules get more complex it's usually harder to completely describe the problem in plain language than in a more formally specified language. For instance, plain-language "apply the user's promo code" doesn't capture the nasty if/else tree you might hit when you're deep in the codebase and see that there's already a bunch of restrictions on other types of promo codes and the product manager didn't think about which of those restrictions should apply to this new promo code. And at this point you're gonna need to use plain language to describe and refine the problem with the product manager - but if you instead were relying on an LLM to turn your short sentence into the right code, it might pick the wrong thing when it comes to modifying that existing code.