WisebotAI
← Back to blog

Voice agents with Vapi: answer the phone with your knowledge base

Add a voice agent that talks to callers, answers from your approved knowledge, calls tools, and hands off to a person — the same agent logic as your chat channels.

Written by

WisebotAI

Published

June 11, 2026

Text channels get the attention, but plenty of customers still call. With Vapi voice integration, the same grounded agent that handles your chat can answer the phone — using your knowledge, your tools, and the same handoff rules.

What you get

  • Live calls — a voice agent answers, understands, and responds in real time.
  • Phone numbers & assistants — connect a number and a Vapi assistant from the dashboard.
  • One brain across channels — the voice agent draws on the same knowledge base and tools as your web and messaging agents, so answers stay consistent.

Set it up

  1. Connect Vapi and pick (or provision) a phone number and assistant.
  2. Point the assistant at your agent so it inherits your knowledge and system prompt.
  3. Decide what voice can and can't do — keep sensitive actions behind approval or route them to a person.

Design voice prompts for the ear, not the page

Voice is different from chat. A few rules that help:

You answer Acme's support line. Speak in short, clear sentences.
Confirm what you heard before acting ("You'd like to reset your password — is that right?").
Answer only from the attached knowledge. If you're unsure, say so and offer a transfer.
For account changes or anything sensitive, transfer to a human.
  • Short turns. Long paragraphs are unlistenable. One idea per turn.
  • Confirm intent. Read back the request before taking an action.
  • Always offer a human. "I can connect you to someone" is the voice version of clean escalation.

A call flow

  1. Caller asks a question → agent answers from knowledge.
  2. Needs a lookup → calls a custom HTTP tool (order status, account info).
  3. Needs a sensitive change → transfers to a human with the captured context.
  4. Resolved → logs the outcome like any other conversation.

Start small

Put voice on one intent first — order status, hours, or simple FAQs — measure containment and caller sentiment, then widen scope. Voice rewards focus; a narrow, reliable voice agent beats a broad, shaky one.

Next: a support deflection playbook for the escalation design, and tool calling with MCP for the lookups.