It’s relatively straightforward to demonstrate a machine learning model on a static data set in the lab. Turning that model into an end-to-end system that increases availability, reduces non-productive time, and pays for itself many times over through avoided secondary damage for a fleet of ever-shifting assets can be quite a challenge! Assets in use tend to change in unanticipated ways, both slowly and abruptly. Closing the loop in multiple, previously independent legacy systems can cause all sorts of problems, e.g., data can fail to arrive, be misconfigured to arrive on the wrong channel, or arrive corrupted. Driving adoption and changing the culture of an organization can pose its own problems, like making sure maintenance gets reported accurately, completely, and reliably. This keynote will discuss some of the pitfalls of fielding asset health management systems for legacy fleets, and offer some insight on how to successfully and deploy and maintain related machine learning models.