Skip to content

Voice assistant

Kanbu can answer the phone too — the same agent, the same knowledge base, and most of the same tools as in chat, just spoken. Callers hear a natural, real-time reply, not a robotic menu reading "press 1 for...". The feature needs to be enabled for your organization by the Kanbu team first — without that, the Voice option in Configuration isn't selectable.

1. What the voice assistant can do

  • Holds a natural, real-time conversation with the caller (powered by Gemini Live).
  • Draws on the same knowledge base as the chat agent — web pages, documents, product feed, MCP connections, and custom API tools are all available over the phone too. A few tools and settings are chat-only — see 4. Differences from chat.
  • Can transfer the call to a human operator if it can't handle a question, or if the caller asks for one.
  • Can end the call itself — it says a natural goodbye and hangs up, just like a trained operator would.
  • Can optionally record calls — see 3. Call recording.
  • Every call is saved as a conversation transcript alongside your regular chat threads, with usage tracked the same way — nothing extra to review separately.

2. Getting started

The phone number for your voice assistant is usually set up by the Kanbu team — reach out to your contact. If you'd rather use your own number through your own Twilio account, that's configurable in an advanced Configuration section (Twilio account credentials) — ask your contact for help setting it up. Once a number is connected:

  1. In Configuration, switch Agent type to Voice (the same way you'd switch between AI assistant and E-commerce — see Agent types).
  2. Choose a voice for the assistant — 30 voices are available, each with a short character description (e.g. calm, upbeat) and a sample you can play. Optionally set Gender to narrow the list to male or female voices only.
  3. Set a Maximum call length — a safety limit in seconds after which the call ends automatically (default 600s, i.e. 10 minutes).
  4. Optionally turn on call recording — see below.
  5. Test a call and fine-tune the knowledge base and tools based on real caller questions, the same way you would for chat.

3. Call recording

In the Phone call recording section you can turn on the Record phone calls toggle — it's off by default. Recordings are kept for 30 days (a longer retention period can be set by the Kanbu team on request) and appear right next to the call's transcript in Messages, where you can play them back. Recording only applies to real phone calls (over Twilio) — test calls made from the browser are never recorded.

Note: Kanbu does not automatically tell the caller that the call may be recorded. If you turn recording on, make sure you meet your own legal obligations (typically informing callers that the call is recorded) under the laws of the countries you operate in.

4. Differences from chat

The voice agent shares its knowledge base and most tools with chat, but a phone call is a different medium, so a few things behave differently or are turned off:

  • Citations and sources aren't shown — there's no way to display a source link on a call, so citations and clickable follow-up questions are automatically disabled.
  • Product cards and forms don't work — the voice agent has nothing to render a card or form on. Recommended products are read out loud instead (typically two or three options) rather than shown as a card.
  • Operator transfer is "cold" only — the assistant hands the call straight off without first consulting the operator. You can optionally set a fallback phone number and a farewell message for when a transfer can't go through.

5. Transferring to an operator, and ending the call

The voice assistant uses the same operator configuration as chat (see Operator handoff). When it decides to transfer a call:

  • It tells the caller first ("transferring you to a colleague now...") and only hands off the call once it finishes speaking — the caller is never cut off mid-sentence.
  • If a live transfer isn't possible on that channel, the assistant reads out the contact details (phone, email) instead of pretending to hand off.

The same logic applies to ending a call — the assistant says goodbye first and only hangs up once that response has finished.

6. Example use case

An electronics store connects a phone number to its voice assistant. A customer calls asking about their order status — the assistant, connected to a custom API tool, checks the order and reads out the status and expected delivery date. When the customer then asks about a complex return, the assistant transfers the call to support.

Frequently asked questions

  • Do I need my own phone number?

    Not necessarily. The number for your voice assistant is usually set up by Kanbu together with you — reach out to your contact to discuss options (a new number, or connecting an existing one). If you'd rather use your own Twilio account, that's configurable too.

  • Do I need different content for voice than for chat?

    No. The voice assistant works from the same knowledge base, feed, and most tools as any other agent type. Responses are automatically adapted for speech, and a few chat-only elements (citations, product cards, forms) aren't shown on a call — see Differences from chat.

  • How do I pick a voice?

    In Configuration, the Voice field lists 30 options with an audio sample and a short character description. Use the Gender field to narrow the list to male or female voices.

  • Are calls recorded automatically?

    No, recording is optional and off by default — turn it on in the Phone call recording section. Remember your own legal obligation to inform callers if the law requires it.

  • Where do I find the call transcript?

    In Messages, the same place as chat conversations — the call appears as a thread with a transcript and usage stats. If recording is on, you'll also find a playable recording on the thread.

  • What happens if a call goes over the maximum length?

    The assistant warns the caller in advance that the time limit is approaching, and the call ends automatically once it's reached.