The best thing about personalized medicine is that it's obviously right. The worst thing is we mostly have no idea how to do it. We know that different people respond to different treatments. But outside a few special cases like cancer, we don't know how to predict which treatment will work for which person. Some psychiatric researchers claim they can do this at a high level; I think they're wrong. For most treatments and most conditions, there's no way to figure out whether a given sometimes-effective treatment will work on a given individual besides trying it and seeing.
This suggests that some chronic conditions might do best with a model centered around a controlled process of guess-and-check. When it's safe and possible, we should be maximizing throughput – finding out how to test as many medications as we can in the short time before we exhaust our patients' patience, and how to best assess the effects of each. The process of treating each individual should mirror the process of medicine in general, balancing the need to run controlled trials and gather more evidence with the need to move quickly.
I don't know how seriously to take this idea, but I would like to try it.