For instance, running high speed rail without computer signalling would be a huge risk there's no sense in taking ( due to the speeds involved, by the time you see anything of issue like a train or damage in the tracks in front, it's too late to stop).
Scale creates problems that are partially solved via software systems, which only handle a subset of the existing variety and new variety introduced by scaling. This makes the systems leaky with regards to unhandled variety and incredibly rigid (i.e., not robust) against exceptions, exceptions that could often be readily be handled by humans. By software being a snapshot of some small group of humans' knowledge of a system, it is completely unable to respond to live operational exceptions. Stafford Beer's work is interesting background on this type of thing.
> A much more efficient automation.
There are many, many examples where that is both not true and true. For the not true cases, just see literally any customer service interaction with a large company in the past half decade. I suppose this could be arguable though if one asks the question "efficient for who?".
Another example I love, that is even less ambiguous, is restaurants (traditional, not fast food) that require waiters to take orders on touchscreen PoS systems, instead of taking the orders on a notepad and inputting them later. The touch screens are incredibly slower on a bigger party ordering appetizers, a main course, drinks and desert - it takes forever to navigate back and forth and find the right item, and usually has to be done again and again for each diner. By comparison, a notepad allows the waiter to write down exactly what they hear, often just making small extra notes to indicate duplicate items.
Of course, this particular one could be improved by arranging the menus more naturally for someone taking orders, and several other conveniences (such as always keeping the current order on screen, to easily tap to add "oh, and another beer for me", "I'll have the same desert as her" etc).