Teams that launch on chat first often reuse the same playbooks for voice. The answers are right and the calls still go badly. Listening is not reading: callers can't scroll back, can't skim, and can't see a list.
Shorter sentences, one idea each
Keep spoken replies under about fifteen words a sentence and give one piece of information at a time. Put the answer first and the reason second: "Your order arrives tomorrow. It was held at the depot overnight."
Offer two choices, not five
A chat agent can show five buttons. A voice agent that reads five options will lose the caller by the third. Offer the two most likely choices and an open question: "Would you like a refund or a replacement, or is it something else?"
Confirm what you heard
Speech recognition is very good and still not perfect, especially with order numbers and postcodes. Read back anything the agent will act on, in small groups: "That's L S 1, 4 A P. Is that right?"
Listening is not reading: callers can't scroll back.
Leave room for interruptions
People talk over voice agents, and that is fine. Design replies so they still make sense if cut off after the first sentence, and let the caller barge in to change direction.
Test by listening
Read every new voice prompt aloud before it ships, then listen to a sample of real calls each week. Parley's simulations can play a scenario as audio, which catches the replies that look fine on screen and drag on the phone.



