Letting Chloe answer customers is only safe if she knows when to stop. Escalation rules are how you tell her.
What an escalation rule is
A sentence, written in plain English, describing a situation in which Chloe should not be the one replying. When a conversation matches one, she stops and hands it to your team.
They live under Chloe, then Settings, in the section headed When to hand off.
They are not keyword filters. They describe a kind of situation, and Chloe judges whether the conversation in front of her is one.
The switch that turns them on
Above the list is Hand off automatically. With it on, a matching rule makes Chloe stop replying, leave a draft for your team, and flag the thread.
Leave it on. Turning it off does not make Chloe better at difficult conversations; it only removes the point at which she admits she is out of her depth.
The four rules you start with
Every workspace begins with four, each of which can be switched off individually:
- Billing disputes, refund requests, or payment problems.
- Legal threats or mentions of legal action.
- The customer explicitly asks for a human, or severe frustration continues after the assistant has failed to help.
- Account closure or data deletion requests.
They are a sensible default set: money, law, anger and account deletion are the four areas where an AI answering confidently and wrongly costs you the most.
How to add your own
- 1Go to Chloe, then Settings.
- 2Scroll to When to hand off.
- 3Click Add a rule in plain language.
- 4Describe the situation as you would to a new member of staff.
- 5Press Enter to add it.
It takes effect immediately, on every channel.
Writing a rule that fires when it should
Escalation rules describe situations, so write them as situations.
- Good. "The customer mentions a signed contract or a dispute with an employer". Recognisable from the conversation.
- Useless. "Anything complicated". Nothing distinguishes it.
- "The customer says the product has injured someone or damaged property". Exactly the kind of thing you want caught.
- Poor. "The customer is a big account". It is not something the conversation reveals.
What happens when one matches
Three things at once: Chloe stops replying, she leaves a draft for your team, and the thread is flagged. What happens when Chloe hands off a thread covers exactly what you see.
Rules are judged, not matched word by word
Two consequences of that worth knowing.
First, the rules work in any language the customer writes in. You write the rule in English; a customer writing in Armenian or Spanish is judged against it in the same way.
Second, matching is a judgement, so it will not be perfect. A rule occasionally fires when you would not have escalated, and occasionally misses. Err towards escalating: a conversation needlessly passed to a person costs a few minutes, and one wrongly kept from a person can cost a customer. ---