What about legal concerns? If a system predicts a person as unlikely to payback a loan, this person might ask why the system denied him. In this scenario, it is not about the trust of who implemented / ordered the system, but the users.

Also, in some business, there might be changes in legislations that requires some changes to the models outputs. Interpretable models are easier to adjust without much cost (e.g., add another if) while complex models such as ANNs will likely need to be fully retrained (and might not be fitting to the new task, so more research on network architectures to solve it).