> ## Documentation Index
> Fetch the complete documentation index at: https://docs.firetone.com.au/llms.txt
> Use this file to discover all available pages before exploring further.

# Live assist

> An AI that listens to a call between two people and writes suggestions to the agent's screen.

export const Unverified = ({what, detail}) => <Warning>
    <strong>Not yet verified on a live call.</strong> {what} is built and covered
    by tests, but {detail} FireTone's engineering tracker records this as an open
    check rather than a passing one. Treat what follows as the design, and expect
    rough edges the first time you run it.
  </Warning>;

Live assist listens to a call an agent is handling and writes short suggestions
onto their screen. It never speaks.

<img src="https://mintlify.s3.us-west-1.amazonaws.com/firetone/images/oncall-assist.png" alt="Suggestions on the agent's screen" />

## Switching it on

Per **process** and per **campaign**, off by default — not per organisation.
Collections may want it; the switchboard does not, and a tenant-wide flag would
stream every internal call to a provider to answer a question nobody asked of
it.

You also choose **which assistant**: an agent profile, whose provider, model and
brief it thinks with. Its voice and identity are ignored, because it never
speaks. A generic listener produces generic advice, which is the fastest way to
teach an agent to stop reading the panel.

## Read this before you enable it

<Warning>
  **It sends the conversation to an AI provider.** Every answered call in an
  enabled process or campaign is streamed for its whole duration and billed per
  minute. Where you operate, that may be a recording you are obliged to tell
  people about — this switch does not ask them.
</Warning>

It also opens a second audio path per answered call on a node that allows a
limited number of new calls per second in total, shared with everything arriving
inbound.

## It cannot be heard

The guarantee is structural rather than configured. An assist session's audio is
discarded at the single point all audio leaves FireTone, so it cannot reach the
call even if a provider ignores the instruction to produce none.

<Note>
  Only OpenAI implements assist today. A tenant configured for another provider
  gets **no assist and a clear line in the log** — deliberately, rather than a
  session configured to speak.
</Note>

## What it costs

Assist opens an ordinary conversation record against the call, so its tokens
appear on the call record and in your AI usage exactly as a virtual agent's do.
See [what AI costs](/tenant/ai/what-ai-costs).

<Unverified what="Live assist" detail="no suggestion has yet been produced: it needs a call between two people, and none has run through this path on this installation." />