Designing the AI-to-human handoff

Three triggers worth escalating on

We've watched thousands of handoffs across Callable deployments. The triggers that consistently produce a positive outcome are simpler than you'd expect:

• Emotional escalation: voice stress, raised volume, profanity, or three negative-sentiment turns in a row.
• Out-of-scope requests: legal questions, custom pricing, account access, or anything the agent's knowledge base doesn't cover with high confidence.
• Direct request for a human: any phrasing that explicitly asks to speak to a person, no matter how politely worded.

Hand off, don't disappear

The worst handoff pattern is the silent transfer: the agent stops talking, the line hisses, and the caller hears nothing for 30 seconds while the system pings a human rep. Always acknowledge the handoff in one short sentence — 'Of course, let me get a teammate on the line, this will take about thirty seconds' — and play a soft hold tone, not silence.

Pass the context, not the recording

When the human picks up, they should see a one-paragraph summary of the call so far, the trigger reason, the caller's stated goal, and a link to the live transcript. Not the full audio. Forcing a human to scrub through a recording before they can help is the single biggest source of cold-handoff failures we've seen.

Never hand back

Once a human is on the call, the AI is done. Trying to return control to the agent mid-call breaks customer trust every time. If the conversation needs to continue afterwards (a follow-up SMS, a calendar invite), let the AI handle that as a separate, asynchronous step.