Let's use the linux kernel as an example, since it does very few, well-defined things, and still doesn't work. This comes from an inability to imagine better, an unwillingness to use proper tools, and an attitude that kernel panics be acceptable.

It's fine to solve a vague problem by simply having the machine ask for human direction in a few cases. It's not fine to have the machine do something inappropriate or crash because a valid case wasn't handled in any way.

Everything below these vague areas can, and should, be perfect. People who claim this be an unobtainable goal are liars.