Human-driven systems, embedded in the real world, are still inefficient enough[0] that they are much more resilient than the kind of systems we work with. At the very least, each part of the system has the ability to independently evaluate the process they're a part of, and stop it in its tracks if they think a fatal mistake is about to happen. See e.g. Stanislav Petrov and be thankful his job wasn't automated, or we'd not be here to talk about it.

--

[0] - Which, in this context, is a good thing.