Just because an LLM can better parse what you're saying it does not mean the business logic to act on it exists. The feature discoverability and churn problem will still exist.
This feature already exists for LLM's -- you can provide them with the list of actions that it can do. This is how you integrate with an API. You describe your "turn on the lights" endpoint and so when you say "I need a bit more light" and it knows it can do something about it.
The users of LLM-based products don't know that; its programmers do. And what happens when an integration is added or removed; how is the user supposed to know?
So, a programmer with LLM knowledge will have a startup idea to wrap all of this up in a nice little app. You say "turn on the lights", the app passes that to the LLM.
The LLM, having been pre trained on what to do, calls the API to turn on the lights.