Some of that is done, for example with tumour genotypes or expression profiles for oncology drugs.
Of course most of this work is speculative and experimental and in the context of clinical trials these are rarely statistically powered to be informative in terms of "these patients are more/less likely to respond to treatment".
Of course, faced with 10 potential treatments for whatever disease you'd like to know which one the patient is most likely to respond to. That's the holy grail if you like, not treating someone with a drug they won't respond to anyway but going straight to the one that will work for them.
To the extent that that is down to individual physiology rather than random, it's largely something for the future when you can genetically profile everyone cheaply, with the compounding problems being that you are then looking for probably small differences in response between potentially hundreds of thousands of subgroups, plus all genetic stuff is multifactorial anyway, so correlating genetic background with response in some set of people is unlikely to apply to some other set. Not to mention we are entirely neglecting phenotype here.
So tailor-made treatment is theoretically very nice, but is likely to remain as much art as science for a long time.