I am a fan of automation and assembly line , so I guess it starts there , unless ...

Designer of the AI system must ensure that 'recovery' from erroneous path ought to be faster and easier than 'accumulative' progressive path? ... Or is _that_ too much to ask for any unknown unknowns?

A favorite paraphrased quote by Brillouin: “A thousand lines in a newspaper, many lines of Shakespeare, many theorums of Einstein have NO value until thought about”, to think about something artificially requires a framework and deep understanding of complexity theory and related solutions. ML and math are each a subset of the comprehensive framework required as well as prioritized emotional like importance based modeling. A good start might be a shameful plug on a colleagues and my book on Artificial Cognitive Architectures which discusses some of the key components to lay the foundation for Self Evolving Life Forms (SELF) for deep space and deep ocean exploration fo example. Maybe it helps some.