Skip to Content
Platform guideThe AI agent

The AI agent

The AI agent answers customers in the widget on your behalf. It’s designed to be grounded and accountable — it answers from your knowledge base, is honest when it can’t help, and hands off cleanly to your team.

How it answers

  1. When a visitor sends a message, the agent searches your knowledge base for relevant content.
  2. It composes an answer grounded in that content, in your configured tone and agent name.
  3. If it can’t find an answer, it says so honestly rather than guessing.
⚠️

The agent answers only from your knowledge base. If it doesn’t know something, add the missing content in Knowledge — that is how you “train” it.

What the agent can do

  • Remember the conversation. It has the full transcript, so visitors can ask follow-up questions and it references what was already discussed.
  • Check human availability. Before promising a handoff — or when a visitor asks for a person — it checks whether an agent is actually online and sets expectations honestly (“someone can help now” vs. “they’ll reply shortly”).
  • Ask if it resolved the issue. After answering directly, it asks whether that solved the problem, and hands off if it didn’t.
  • Escalate to a human. When it can’t help or the visitor asks, it moves the conversation to Awaiting agent, auto-assigns it, and leaves a short summary so the agent never starts cold.
  • Create tickets for concrete problems (a bug, refund, or account issue).
  • Run your workflows and integrations when a request matches one.

Model routing

Simple questions are answered by a fast model; harder reasoning is routed to a more capable one automatically. A semantic cache reuses answers to near-identical questions to keep responses fast and costs low.

Configuring the agent

From Settings, you can set:

  • The agent name shown to visitors.
  • The tone / brand voice.
  • The welcome message and expected response time in the widget.

The best lever on quality, though, is your knowledge base — more complete, well-structured content means better answers and fewer handoffs.

When a human takes over

Once an agent takes over (or the AI escalates), the agent steps back — it won’t reply on top of a human. When the conversation is handed back, the AI resumes.

InboxKnowledge base