Future systems could for example start asking questions more often
Current systems already do that. But they're expensive and it might be cheaper to have a human do it. Prompt engineering is very much a thing if you're working with high performance low memory consumption language models.
We're a long way from having smartphones with a couple terabytes of RAM and a few thousand GPU cores... but our phones can run basic models and they do. Some phones use a basic LLM for keyboard auto correct for example.