Teams spend their design effort on the agent and treat escalation as an afterthought. The escalation path is what determines whether the channel is trusted.
The five triggers
1. The customer asks for a person. Immediately, without negotiation and without a retention attempt. Arguing with someone who asked for a human is how a neutral interaction becomes a complaint.
2. The agent is uncertain. Low confidence should route out rather than guess. A model returning a well formed answer that happens to be wrong is the failure mode that damages trust most, because it is confident.
3. The action is irreversible. Money, cancellations, anything you would have to apologise for. Automate up to the action, put a human checkpoint there, automate everything after.
4. The emotion rises. Repeated frustration, raised language, a second complaint in the same conversation. These are handover signals regardless of whether the agent could technically continue.
5. The same issue repeats. A second contact about the same thing means the first answer did not work. Sending them back through the same path is how a small problem becomes a public one.
What has to travel with the escalation
The same discipline as a sales handoff:
- Who they are and what the system already knows
- What they asked for, in their words
- What the agent already said or did
- Anything that was promised
- Why it escalated
Without this, the human starts cold and the customer repeats themselves, which undoes the benefit of the channel entirely.
The fallback decides everything
What happens when nobody is available is the question most deployments answer badly.
The honest options, in order of preference: transfer to an available person, offer a specific callback window and create a task someone owns, or state plainly that the team is unavailable and when they will be. What does not work is a promised callback that generates no task, because that is a broken promise with a paper trail.
Before switching an agent on, confirm somebody answers during the calling window. An agent that escalates into an empty queue is worse than no agent at all.
What to measure
Escalation rate. Suspiciously low is a warning, not a success. It usually means the agent refuses to hand over.
Time to human after escalation. The number the customer experiences.
Outcome after escalation. Did the human resolve it. If not, the escalation is arriving without context.
Repeat contacts. The clearest signal that the first interaction did not work, whoever handled it.
The design rule worth keeping
Build the escalation path first, then the agent. Teams that do it in that order ship agents people trust. Teams that do it the other way spend their first month after launch discovering the paths they did not design.