Prosody is what makes the same sentence mean different things depending on how it is said. Rising intonation turns a statement into a question. A long pause before a name conveys reluctance. Falling pitch at the end of a list signals completion. None of it is in the words.
For a voice agent, prosody matters in both directions. On the input side, the STT layer can lose the prosodic signal entirely (text transcripts are flat), which makes the agent miss sarcasm, frustration, or hesitation. Good systems either preserve prosody features or run a parallel paralinguistic model that tags the transcript.
On the output side, the TTS layer has to produce prosody that matches what the agent is actually saying. A confirmation should sound confident. A clarifying question should sound curious, not impatient. A bad-news reply should not sound cheerful. The control surface for this is usually SSML plus per-vertical voice tuning, and the difference between a vendor that has invested here and one that has not is audible within two turns.
Vorel tunes per-vertical prosody for the TTS output. The same agent sounds calmer on a clinic call and brisker on a parts inquiry, by design.