A quick war-story on this: We had a system which was taking web-user input for human names, and then some of it had to be sanitized for a crappy third-party system. However some of the names were getting mangled in unexpected ways.
One of the (multiple) issues was that we were sometimes entirely dropping accented characters even when a good alternative existed. This occurred when we were getting "é" (U+00E9) instead of "é" (U+0065 U+0301), a regular letter E plus a special accent modifier. By forcing the second form (D normalization) we were able to strip "just the accents" and avoid excessively-wrong names.
Going further with K+D normalization, weird stuff like "⑧" (letter 8 in a circle) becomes a regular number 8.