Using the Romance languages as an example, I believe its commonly assumed that their personal endings evolved from agglutinating endings. By that I mean that long ago some language had suffixes for things like mood, tense, number and person, and over eons these merged together into sometimes singular phonemes.
Of course, a problem with this process is that as phonemes are dropped, distinctions are bound to disappear. For example, let's say a language had CV suffixes that marked tense, aspect, person, and number. Marking all four of these would thus require 8 phonemes. But over time these 8 are fused and reduced to just two or even one phoneme. Obviously, there's far more CVCVCVCV sequences possible than (C)V sequences in a language, regardless of the size of its phoneme inventory. Thus this means that over time some endings are bound to end up sounding identical. But if you look at say, the verb endings for Spanish, you don't see much of this going on. For the most part, every possible combination of mood, tense, person, and number has a suffix unique to it, though some are distinguished purely by the placement of stress in the word (which I surmise is something that developed to reduce ambiguity).
I was thinking of doing this to create fusional endings for a conlang, but of course I don't really see how I can reduce a 6-8 phoneme sequence into just 1 or 2 without some distinctions being lost. What I think would help is perhaps to see an actual example in a real language of chains of agglutinating endings morphing into fusional endings. I'm guessing there's no such thing for something like Proto-Indo-European, but perhaps there's an example from somewhere else in the world?