There's no way it is *better* than a human at understanding human requests.
I wouldn't be so quick to underestimate the worth stored in the models. The tricky part is getting it out in a useful form.
If we allow for the loose use of "understand" in the context of LLM:s etc, I would say that one model understands and combines fragments of knowledge, behavior and expectations of millions, perhaps billions of humans. I couldn't possibly match that. My best bet is to have cultural, technical, organizational and domain knowledge that covers enough of the task at hand.