AI agents

Overview

An AI agent is a configured assistant with a name, a face, instructions, a model, a set of knowledge sources and rules for when to stop. The AI agent area in the sidebar has the agent list, four knowledge tabs (Website, Files, Q&A, Text), a playground to talk to the agent with your settings live, and a usage page. Every reply is grounded in retrieval: the agent finds the passages of your knowledge that match the question and answers from those, citing them, and says so when it has nothing to go on.

Create an agent

1. 1
   AI agent, then Create agent. Give it a name; the avatar is an icon and colour you pick, or a photo up to 2 MB.
2. 2
   Open the playground. The settings panel beside the chat holds everything about the agent; changes apply to the test chat at once and are saved when you click Save.
3. 3
   Write the instructions (the system prompt). The default describes a support assistant; add your tone, what it must never promise, and how to sign off.
4. 4
   Pick a model and a temperature (precise to creative), and a maximum reply length in tokens. Models include GPT-6 Luna, Gemini 3.8 Flash, MiMo V2.6 Pro and Mistral Large 4, all on every plan; the default is Claude Haiku 5.5, a small fast model.
5. 5
   Set the welcome message the agent opens with, the handover rules and the limits described below.
6. 6
   Attach knowledge (next section), then Activate. One active agent answers the website chat; others can be kept for testing.

Knowledge

• Re-import knowledge rebuilds the index after a batch of changes; the playground says which sources were used for each answer, so you can see what it read.
• A website source can auto-refresh daily, so a changed pricing page reaches the agent without anyone re-importing. The scheduler runs the refresh.
• Retrieval works on the stored text only, never the original files, and every agent in the workspace can share the same sources or use its own.

Handing over

The agent stops and leaves the conversation for a person in four cases:

• Low confidence: when the retrieved knowledge does not answer the question well, the agent hands over rather than guessing. The threshold is on the agent.
• Trigger words: a visitor message containing one of your words (human, agent, talk to someone) passes the conversation at once.
• The model decides to: when the visitor is upset or asks for something the agent should not do, it can hand over on its own judgement.
• A teammate joins: Take control or a reply from the inbox stops the agent in that conversation.

The handover message is what the visitor reads at that moment. The conversation lands in Unassigned with the reason noted, and the thread keeps the agent’s replies. On the Plus plan the agent can offer to raise a ticket instead, in the chat, when it cannot resolve something; see Tickets.

Limits and cost

Two caps on the agent: Daily answers and Monthly tokens, both checked before a model call. Past either, the visitor gets the replacement message and the conversation waits for a person. On top of that the workspace has its AI credit: each plan includes a monthly amount, an answered conversation draws five cents, and prepaid top-ups cover the rest; see Plans and billing. When the credit runs out the same replacement message applies. Nothing is charged past what you have.

Voice

On the Plus plan the agent can talk. With audio allowed in Settings, then Behavior, the chatbox offers a voice mode: the visitor speaks, the agent answers aloud in the voice chosen on the agent, and listens again. Spoken replies are kept short on purpose. The playground has the same voice mode for testing. Voice usage is metered per workspace and can be capped.

Usage and quality

The Usage page shows, per agent or for all, the AI conversations in the period and this month against the allowance, how many were rated helpful or unhelpful, how many were handed to a person, and a history by day. Six hours after an AI-handled conversation goes quiet, the visitor is asked to rate it; those ratings feed CX score in Analytics, whose AI page adds involvement and resolution rates. Read the unhelpful ones and fix them with a Q&A entry.

Where it answers

For the chatbox side of the agent, the ticket offer and voice mode, see The chatbox; to hand a conversation to the agent from a rule, the Let the AI answer action in Automations.