It can only be good when the interface already knows a massive amount of context. For example the LLM in a future use case knows all your email and you just ask it to draw a time line of this thread and etc.
If you are asking it to do stuff from scratch like most of us do on ChatGPT, it’s quite a pain