Most AI help comes from one of two kinds of vendor. One writes the strategy and hands you a roadmap. The other builds whatever you put in front of them. Projects tend to stall in the gap between the two, because the people who understand the problem aren't the ones writing the code, and what ships ends up solving a slightly different problem than the one you had.
I do both halves myself. The early part of an engagement is spent working out what's genuinely wrong, which is often not what people expected going in. Then I build the fix and stay until it's actually running in production. I leave a person in charge of the parts of a system that need judgment. The automation handles volume and consistency, but someone real still owns the decision.
I've spent about eight years on operational problems that are messy and expensive at the same time: compliance, trust and safety, localization, workforce planning. My background is in math and economics, which is most of why I care as much about how something gets measured as about what gets built.