The question guiding AI adoption/deployment often seems to be “what can we automate?” (which is, of course, not a new question, we just have a new tool to use in the answer.)

The better question is: what shouldn’t we automate?

Then work backwards from there.