How do I set up the AI Assistant step?

What is it for?

The “AI Assistant” step answers visitors' free-form questions with AI. The conversation stays on this step; the agent uses the knowledge base, your tools and integration actions on its own when needed, and moves the flow to another step at the right moment (for example a Request step).

The conversation can also complete on this step: when the visitor says they're done, or 30 minutes after their last message if they wrote at least once.

The step is in the “AI” category of the palette. Its panel has the “Content”, “Tools”, “Integrations”, “Knowledge” and “Tracking” tabs.

System prompt and model

1. Write the agent's first line in “On-screen text” (for example “Ask me anything you'd like to know.”).
2. In “System prompt”, describe who the agent is, which topics it knows and how it should answer. If it's empty, the agent can't reply. See How do I write an effective system prompt?
3. Optionally pick another model from the “Model” list. The list has six model families; “Which one should I pick?” asks a few questions and suggests a model that fits. Credit use varies a lot from model to model. “Default provider” uses the default Claude model. The choice applies to this step only.

If the chosen model returns an error, the step switches once to the same provider's backup model. Timeouts and rate limits aren't retried.

Knowledge, tools and integrations

The agent treats these three sources equally and picks which one to check for each question. It doesn't make up what none of them has; it says it doesn't know.

• “Knowledge” tab: “Off”, “All entries” (shared entries and this agent's entries, including ones added later), “This agent's entries only”, or “Entries I choose” (only the ones you turn on; new entries start off). See Knowledge base.
• “Tools” tab: turn on the tools the agent may call one by one, or choose “All tools” (tools added later are turned on too). Each tool's description tells the agent which questions it is for.
• “Can take tickets” (at the top of the Tools tab, on by default): if a visitor wants to leave a request or pass something on to your team, the agent asks for their contact details and opens the ticket itself; it appears on the Tickets screen. If the same person adds information within an hour, no new ticket is opened; the information is added to their open ticket, and if the ticket was forwarded to your help desk, it is added there as an internal note (the customer doesn't receive it). If the step routes to a ticket step, tickets are taken there. At most 5 tickets per hour are taken from one visitor. If you turn it off, the agent says it can't take requests and makes no promises.
• “Integrations” tab: turning an integration on opens a window where you pick its actions; the agent uses only the ones you pick. The window's “All actions” switch turns on every action, including ones that delete or send. Use “Configure” to fix fields: a field you fill stays fixed and the visitor isn't asked for it, while a field you leave empty (“Let the AI fill”) is filled from the conversation. A {variable} inside a fixed value isn't filled in; the value goes as you wrote it. Choose places like the board, list or channel here.

Integrations that aren't connected are listed under “Not connected” with a “Connect” link; the agent can't use them until they are connected.

Handoffs

In the “Handoffs” section, use “Add handoff” to add the steps the agent may move the flow to, and under each one answer “When should the AI move to this step?”.

For a handoff to a Request step, an empty box uses the default rule: the agent moves to the Request step when the visitor wants to leave a request or complaint, wants to be contacted back, or agrees to have their question passed on when no answer is found. For other steps the editor warns you if no description is written.

“Details to collect”: under a handoff, use “Collect details first” to add the details the AI must get from the visitor before moving to that step (for example a date, the number of people or an email address). They are set up like form fields: “Field name”, “Variable name”, “Field type”, “Required” and “Validation rule”. The AI doesn't move to the step until these details are complete; if one is missing or doesn't fit its rule, it asks the visitor again. The details can be used in later steps as {variable}, and a later “Summary and confirmation” step shows them and lets the visitor correct them.

Safety and limits

The actions you turn on are called by an AI that talks to your visitors: it can tell the visitor what it reads, and what it deletes or sends can't be undone. Such actions are marked with “Deletes” and “Sends messages” badges in the action list. In the agent these actions need no extra confirmation; they only carry the badge.

The agent declines general tasks that have nothing to do with your business (such as writing code or doing homework).

To prevent abuse, a workspace can run at most 20 deletions and 100 message sends (such as email) per hour; Integration Action steps in your agents share the same limit. Turning on many actions can slow down replies, so turn on only the ones you really need.