Building a Voice Agent for Customer Support

The difference between a voice agent and a "smart IVR" is that the agent can be interrupted, can ask follow-ups and can actually do things. This is a checklist you can follow from the first prompt to the first real call.

Hear one first

An agent that gets things done sounds like this — it confirms the booking, repeats the key details and leaves room to reschedule:

ボイス: 甜美治愈中国語17秒

ボイス: 甜美治愈中国語11秒

The prompt covers three things

1. Identity: who it is, which company it represents, what it may not promise.
2. Procedure: the order of operations for common requests — verify the caller before looking up an order.
3. Boundaries: when to hand off to a human, when to end the call.

Write the prompt in the customer's language; the agent mirrors whatever language the caller uses.

The voice is the first impression

Support voices should be warm, clear and slightly slow. The gentle female voice is the default; for technical support the confident tech voice reads as more expert. Audition in the voice library and bind one.

Knowledge base and tools

Turn FAQs, return policy and opening hours into text and attach them; order lookups and rescheduling go through tool calls. The agent answers only from the knowledge base and hands off when it has nothing.

Latency and interruptions

Real callers interrupt. Keep response latency under a second, and on interruption stop playback immediately and listen. Both are live in the demo on the voice agent page.

Cost

Billing is per second of conversation in 60-second units. A three-minute support call is about 45 credits — an order of magnitude below a human agent's per-minute cost. Try two free calls on the page first.