This is a critical thing for AI proof points, they work best when the outcome can be objectively verified and failures feed back in for retry, putting into the context window guidance that disqualifies the statistical path it would have chosen.
A human is so much more likely to get it right the first try, if they can get it right at all.
This has some utility, closest to practical in some cases of programming, though people can be bad at judging when that can be practical.
The other useful scenario is when being wrong is still "good enough" a lot of the time, like creating specific instructions for a human out of broader documentation. Yeah it's wrong a fair amount, but close enough that a human can compensate.
I have been a fly on the wall for some computer supply chain discussions.
Thanks to the contracts and commitments, there is a huge backlog and supply chain driving to the official plan.
However, they also are emphasizing that clients that are flexible to take early delivery are a big priority, because they expect a significant chunk of those to fall through.
I saw it happen once a few weeks ago, a client being asked to take early delivery because they had a huge cancellation and they needed new takers willing and able to take that stuff.
So large purchasing commitments that even if they are formally "uncancellable" end up falling through.