Turns out, humans are also pretty good at probably being almost right. Luckily though, they've figured out how to build systems with higher fault tolerance than their own.
It's easy for an ML model to do great in the average case but horribly in the worst case, so you're correct if "fault tolerance" is a simple loss function to be minimized, but wrong if "fault tolerance" actually means what we care about. That's much harder and not something we uniformly know how to do with ML. Like the costco example in the article.
To do that with ML, they'd need a loss function that captures what we care about and data it sounds like they don't have.
We do know how to build systems with high fault tolerance the old fashioned non "AI" way.