Tier-0 questions
Most support calls repeat the same handful of questions with answers that do not change call to call. “How do I reset my password” and “what are your support hours” are tier-0 questions; an AI voice agent for customer service can answer them consistently without adding another caller to the queue.
Account-specific checks do not fit a line that cannot look up live records. It should take a message rather than guess at an answer, then make clear that a person will follow up.
That lets people spend their time on calls that need a specialist or judgment, instead of repeating the same answer and recording the same details by hand. The caller also gets a direct answer without waiting behind routine calls or explaining the basics twice to another person on the team.
Callbacks instead of hold music
When a specialist is unavailable, the honest move is a callback: the line takes down the issue and the best number to reach the caller, then a person follows up instead of leaving the caller on hold or in a dead-end voicemail queue.
The person calling back has the issue in the caller’s own words instead of starting from zero, which keeps the follow-up focused on the problem rather than a retelling.
Where it should hand off
A support voice agent should hand off anything beyond tier-0: an account issue needing judgment, a complaint, or a technical problem outside the brief. It should say so, take a message, and not talk its way through an answer it does not have. The edge should be visible to the caller, not only in the system prompt.
A frustrated caller is also a hand-off signal. A person asking to be heard should not receive more scripted reassurance, even when the original question looked routine.
Write a short, specific list in advance: billing disputes, refunds above a set amount, or a caller who asks for a manager. A named list is more predictable than guessing when a conversation has gone too far, and makes the behavior easier to test before publishing.
Latency, measured honestly
How fast a line replies matters more in support calls than almost anywhere else — a caller who is already frustrated notices every extra second of silence. In one measured sample across four real phone calls, turn latency — the pause before the line replies — ran 602–985 ms, down from roughly five seconds before that specific fix.
That’s a real measurement from a specific test, not a guarantee for every call on every stack — latency depends on the provider chosen, the network the call travels over, and the complexity of what’s being asked. See /how-it-works for how it was measured.
Cost transparency
Support-line cost needs to be checkable. Every call is itemized by provider on the cost page, not blended into one number, so the carrier, transcription, model, and voice layers stay visible.
That matters as a queue grows: a single average can hide which provider is driving the bill. See the current, versioned price book at /cost.
Straight answers.
A phone line that answers the repetitive tier-0 questions — order status, a password reset, support hours — the same way every time, and hands off anything that needs a specialist or a judgment call to a person.