I've been dealing with similar this week. The reality is, very few data sources for these reports allow for real-time execution of R code from the back-end outside of embedding R code on the front-end. From a small business standpoint, this is totally fine. But, when you get into larger organizations, having to manage hundreds of R code implementations on the front-end is a pain. Having the ability to store, execute, train, and return predictions from something as simply calling a stored procedure is extremely powerful. No other data source really does that in many of the front-end tools you may use to visualize the data.

Thus, to force the machine learning down the path of writing all the code in the app versus storing it in the back-end is something you might want to think long and hard about. Simply storing the pre-trained data is not going to be enough if you get serious with ML.