Voicemail is the quietest way to waste money on automated calling. The call connects, so it bills. The agent hears speech, so it responds. And it is talking to a recording that will never answer.
Why it costs more than you would think
Without detection, a voicemail plays out like this: the greeting is transcribed as if the person spoke it, the agent replies politely, waits, hears nothing, prompts again, waits again — and burns two to four billed minutes before a timeout ends it.
On a list where 15% of connects go to voicemail, that is a real line item. Worse, every one of those calls produces a transcript full of an answering-machine greeting and a confused agent, which then pollutes your summaries and your ratings.
What detection does
The agent watches for the shape of a machine rather than the words: a greeting that runs uninterrupted, a beep, phrasing that only ever appears in a recorded message. When it is confident, it ends the call immediately.
The important part is what does not happen: the turn is never persisted, never enters the conversation history, and never reaches the summary or the rating. A call that reached a machine is marked as such and otherwise leaves your data alone.
What should happen next
Do not leave a message by default. It feels productive and almost never works — people do not call back automated voicemails, and you have paid for the recording time. If your use case genuinely warrants one, an announcement agent does it far cheaper than a conversational one.
Retry at a different time of day. A number that hit voicemail at 11 AM is a good candidate for 6 PM. The same slot tomorrow usually produces the same result.
Cap the attempts. Two voicemails is information: this person is not answering unknown numbers. A third attempt is not going to change that, and it is the point at which persistence starts looking like harassment.
The related failure
Hold music and carrier announcements cause the same problem in a different costume — they reach the agent as speech and draw a polite reply each time. These are suppressed the same way: not persisted, not replied to, not allowed into history. Without that, a single call can run dozens of turns of the agent saying "sure, I'll hold" to a recording.