While they can help us derive insights from the data, these insights need to be understood before handing over control to the algorithm. I believe that the best approach is to balance the AI insights with engineering principles (reliability engineering in an asset space), validating the analysis of the AI, and linking it back to something explained and understood. If you don’t then:
1) You don’t learn as much,
2) You are exposed to issues / faults in the data,
3) It's harder to correct and improve the algorithms.