Send a text with a link to a ten-question survey after a call, and almost nobody finishes it. Ask the same core question out loud, in the last fifteen seconds of a call that's already happening, and completion rates look completely different.
That's the idea behind voice surveys: NPS, CSAT, and CES scored automatically, delivered as a natural part of a conversation your customer is already having, not a separate ask competing for their attention later.
The follow-up question is where it gets useful. A caller who answers 9 out of 10 doesn't need to be asked what went wrong, so they aren't. A caller who answers 4 gets one adaptive follow-up, what could we have done better, and that's the entire survey for most callers: one question, maybe two.
Under the hood it's not a separate system bolted onto your call flow. The question sequence is a graph agent like any other, the possible answers map to a scoring rubric per survey type, and a completed survey run is just a batch job, so the same concurrency and retry logic that handles a thousand-call campaign handles a thousand-respondent survey.
You can trigger it automatically after a call, push a respondent list through the API, upload a CSV, or set it on a recurring schedule, and every response lands in the same analytics view as the call it came from, no exporting to a separate survey tool to make sense of the numbers.