Escalation
What happens when the agent should not decide alone. Handing off well is a feature, not an admission of failure — the alternative is a confident wrong answer.
What triggers a handoff
- The customer asks for a person. The gate stage recognises this as its own intent, before retrieval runs. Nobody should have to argue with a bot to reach a human.
- Repeated abstention. An agent that cannot answer twice in a row is not going to answer on the third attempt.
- A guard refused something the customer needs. A blocked write with no alternative path is a handoff, not a dead end.
- Sentiment. The gate classifies tone; a customer who has become angry is escalated rather than processed.
Modes
| Mode | Behaviour |
|---|---|
INBOX | The conversation lands in the dashboard Inbox and is marked escalated. Your team replies from there, in the same thread the customer is already in. |
EMAIL | The transcript is sent to the address you configure. Use this when your team lives in a helpdesk rather than here. |
OFF | No handoff. The agent says it cannot help and stops. Only sensible when there is genuinely nobody to hand off to. |
What the human receives
The full conversation, the customer’s verified identity if there is one, every action proposed or executed, and the turn traces. Your team sees why the agent gave up, not just that it did.
Replying as a human
Replies sent from the Inbox appear in the customer’s widget in the same thread, attributed to your team rather than to the agent. The conversation stays open until someone closes it.
Resolved means two different things
The distinction matters for your metrics, so the product keeps them separate:
- Autonomously resolved — the agent handled it alone, with no human reply and no escalation. This is what the resolution rate on the dashboard measures.
- Manually resolved — a person closed it from the Inbox. This is recorded, but it never counts toward the autonomous rate.
Collapsing the two would let you improve your headline number by having humans close more tickets, which is exactly backwards.
How resolution is computed
Autonomous resolution is calculated on a schedule, not at write time. A conversation counts as resolved when it goes quiet for the configured window without escalation and without a human reply. Computing it live would mean every conversation is unresolved right up until it is, which produces a metric that only ever looks bad.
The window and the cadence are server configuration. The Analytics page shows the rate over time alongside cost per resolution, so you can see what autonomy is actually worth. See The turn pipeline for what each turn records.