Voice stealing is useful when a synth receives more notes than it can process. Voice stealing ensures you hear the most recent notes, while killing the older notes. This strategy works well musically.
Check out the granular example in the examples folder. The poly~ in rgrano.maxpat has 16 voices, but you could easily ask it to play too many grains if you shorten the metro time and/or extend the duration of the grains.
A more typical case would be if you made a keyboard-controlled synth and played big clusters of notes with your arm. By killing older notes and letting the new ones play, you hear a more accurate representation of musical intent.