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

# Transfer rules

> When the agent hands a call to a human, and who it reaches.

The agent handles most calls end to end. Transfer rules cover the ones it shouldn't — emergencies, escalations, and anything where a person needs to be on the line.

Getting these right is what makes the system safe to leave running unattended.

## What to transfer

<Columns cols={2}>
  <Card title="Always transfer" icon="triangle-exclamation">
    Genuine emergencies — gas smell, flooding, no heat in freezing weather, anything with a safety dimension. An angry customer. Anything legal or insurance-related.
  </Card>

  <Card title="Usually don't" icon="circle-check">
    Price questions, availability, booking, rescheduling, general enquiries. The agent handles these better than a transfer to a voicemail box would.
  </Card>
</Columns>

<Warning>
  Define your emergency triggers explicitly, in the words customers actually use — "I smell gas", "water everywhere", "no heat and there's a baby in the house". Don't rely on the agent inferring urgency from tone.
</Warning>

## Where transfers go

Configure the destination and the fallback chain:

<Steps>
  <Step title="Primary destination">
    Whoever should take the call — an on-call mobile, a dispatcher, a shared line.
  </Step>

  <Step title="Fallback">
    What happens when the primary doesn't pick up. A second number, or back to the agent to take a detailed message.
  </Step>

  <Step title="Last resort">
    Never a dead end. The caller should always finish the call having been told what happens next and when.
  </Step>
</Steps>

<Tip>
  Test the whole chain, including the fallback, by calling in and triggering an emergency yourself. An untested transfer rule is a promise you haven't checked you can keep — and it fails on the worst possible call.
</Tip>

## Hours matter

An emergency at 2pm and one at 2am should route differently. Set destinations per time period so the after-hours calls reach whoever is genuinely on call, rather than ringing an empty office.

## Cost

Each human escalation bills at \$3 on every plan — see [Plans](/plans). It's deliberately cheap relative to the value of catching an emergency; don't tune your rules around it. If your escalation count is high, the fix is usually a gap in the [knowledge base](/receptionist/knowledge-base), not a stricter rule.

## Reviewing transfers

Transferred calls are flagged in your [call logs](/receptionist/call-logs), and fire the `call.human_escalation` [webhook](/api-reference/webhooks). Review them regularly:

* **Transfers that shouldn't have happened** — the agent didn't know something it should. Add it to the knowledge base.
* **Calls that should have transferred and didn't** — a missing trigger. Add the phrasing the caller actually used.
