You don't trust it yet, like a new human assistant you might hire - will they be able to handle all the variables? Eventually, they earn your trust and you start offloading everything to their inbox.

No, not like a human assistant. Competent humans will use logical reasoning, non-digital signals like body language and audible clues, and know the limits of their knowledge, so are more likely to ask for missing input. Humans will also be more predictable.

LLMs don’t learn. They’re static. You could try to fine tune, or continually add longer and longer context, but in the end you hit a wall.