What is escalation?
Escalation is the handoff from an AI step to a person when the situation is outside the script: anger, money, safety, or a simple 'I want a human.'
Also called: handoff, live agent
In plain English
The agent should have a boring list of when to stop. No list, and it will keep talking. The handoff needs a destination: a phone, a Slack channel, a shared inbox. 'Someone will follow up' is not a destination. Write the destination next to the trigger. Anger goes to the owner by text. A refund request goes to the bookkeeper's queue. 'A human will follow up' with no name is how leads die. Time the handoff once. If it sat all weekend, the rule is fiction. Also stop the bot from sending a second reply after the handoff, or the customer hears two voices. Practice the handoff on a slow afternoon. The first real angry caller should not be the first time the phone path is tried. Time it with a clock, not a feeling. Use a stopwatch once and write the seconds down.
How an owner would use it
Write five escalation triggers in the SOP. Name who receives them and how fast. Test one on purpose.
Common mistakes
An escalation rule that says 'if it is complicated, get a human' is not a rule. Name the triggers: refund, anger, safety, legal, 'I want a person.' A handoff with no destination is a dead end. Live agent means the person who receives it, not a second bot. Do not let the agent keep messaging after the handoff or the customer gets two voices. Test one trigger on purpose.
A worked example
A carpet cleaner
A stain complaint skips the cheerful draft
This is an illustration, not a customer result. If a message mentions a ruined rug, a refund, or asks for the owner, the agent stops drafting a 'glad we could help' reply. It posts the thread to the owner's phone with the customer's words intact.
Related terms
Prompts and guides
Questions owners ask
What is escalation?
Escalation is routing a conversation from the AI to a named person when the request is sensitive, unclear, or explicitly wants a human.
How do I set up escalation?
List the triggers, name the destination, and make sure the agent cannot keep acting after the handoff. Review those threads.
What if nobody is available to take the handoff?
Then the agent should say when a person will reply, using hours you wrote, and stop. Inventing a callback time or continuing to troubleshoot is worse than a plain 'we are closed, here is the morning window.'
Product guides are on AI apps for business.