When the agent hands over to a human
Escalating a conversation to your team.
Last updated
A conversation always carries one of three states: "Unresolved", "Needs human", "Resolved". The AI agent only answers in the first. Moving to the second is what escalating means: the conversation is waiting for a person, and the agent falls silent.
The five ways a conversation gets escalated
- The agent decides. It has a tool for it, and uses it when it cannot help or when the visitor asks for a person. It first writes the visitor a message explaining the handover, then changes the state.
- The visitor asks. Below the agent’s latest answer, the widget offers a "Talk to a person" button: the conversation goes to your team without the agent answering, and the widget tells the visitor, in their language, that a person will reply.
- You decide. The state button at the top of the conversation cycles through the three states on each click: "Unresolved" → "Needs human" → "Resolved" → "Unresolved".
- You answer. Your first message in an "Unresolved" conversation moves it to "Needs human", and assigns it to you if nobody had taken it yet. So there is nothing to click in order to take over: writing is enough.
- A cap is reached. After 15 visitor messages in one conversation, at your plan’s monthly cap (AI-handled conversations, or agent messages), or if an unusual number of new conversations call on the agent within an hour, it hands over to your team. The widget tells the visitor so, in their language, and the escalation email gives the reason.
What the agent’s escalation sets off
- A message to the visitor, written for their case, before the state changes.
- An internal note explaining the reason, if "Internal notes writing" is on in the AI configuration. Only your team sees it.
- The tags it suggests, if they already exist among yours: the ones it invents are ignored, never created.
- An escalation email to every team member, except those who turned it off.
That email goes to every member of the organization by default. Each one turns it off for themselves under Configuration → Notifications: "Conversation Escalation", or "Email Notifications" for every email. It also goes out when the visitor asks for a person and when a cap makes the agent hand over; it then says why. A single conversation sends at most one every 24 hours.
Why the agent did not answer
Four conditions have to hold for it to answer. The first one missing explains the silence:
- The conversation is "Unresolved". Escalated or resolved, the agent writes no more — and that is deliberate: it does not talk over the person who took the conversation back.
- The plan includes the AI agent. The free plan has none: every conversation there waits for a person. An unpaid invoice past its grace period returns the organization to the free plan, so the agent goes quiet then too.
- The "Enable AI" switch is on, under AI Configuration.
- No cap is reached: 15 visitor messages in the conversation, or your plan’s monthly cap. At a cap, the agent does not just fall silent: it hands over to your team, and the conversation becomes "Needs human". The visitor’s message is always recorded — it is never lost.
What bounds the agent
- It re-reads the last 12 messages of the conversation, not the whole of it.
- It works in steps: 5 at most per message received, each of which can group several actions. Each action therefore also has its own cap per message — 4 searches of your knowledge base, at most.
- Its answer is capped at 800 tokens: it answers short, and hands over rather than holding forth.
- The passages it finds in your knowledge base reach it summarized, in 600 tokens at most.
- It handles at most 15 visitor messages per conversation; beyond that, it hands over to your team.