[blaird]

Yep, agreed. If decisions can be made by a human often they'll stick to that, often arguing there is no need for data.

In my former space (credit card fraud detection and underwriting), you obviously need a data driven solution. Without even considering latency requirements, you aren't do 6-10B manual decisions/year. The rationale for a more complex ML approach is easier to prove the ROI for, given the need is already there, just with an inferior technical solution.