AI agents · Documentation

An AI agent is an LLM-powered agent. Once you add it to a queue, the system treats it like any other agent — assigning conversations to it and letting it reply to customers on its own. AI agents can also serve backend tasks such as tagging contacts when a conversation closes. This page covers how to create and configure an AI agent.

The Agent Directory

Open AI agentsto see the Agent Directory — every AI agent in your organization, one per row. Each row shows the agent's Name, the number of Queues it belongs to, and its Status. Use the ⋯ menu at the end of a row to Edit or Delete an agent. Click New AI Agent at the top to create one.

Status at a glance

Status	Meaning
Active	Enabled and running normally
Disabled	Turned off; does not take conversations
Restricted	The AI resource pack is exhausted; the agent cannot respond to requests.

When the AI resource pack runs low

Every AI reply spends tokens from your organization's AI resource pack. When the balance gets close to the minimum reserve, a banner warns you to top up. If it drops below the reserve, the system blocks new AI agents from being created and switches running agents to Restricted — they cannot work normally. Once AI resources are available again, agents automatically return to Active.

Creating an AI agent

Fields required to create an AI agent:

Basic information

Display Name — what this agent is called, up to 64 characters (for example, Support Assistant). Required.
Active — whether the agent is enabled.
Available for user-facing conversations — on by default, which lets the agent be added to queues and assigned conversations. Turn it off to make the agent backend-only — for example, one that only auto-tags a contact when a conversation closes.

A backend-only AI agent cannot remain in queues

If you switch Available for user-facing conversations off while the agent is still in one or more queues, saving is rejected. Remove it from every queue first, then save.

Model

Choose the large language model that powers the AI agent. More advanced models offer higher intelligence but also increase cost and extend response time. Pick a model based on the balance you want between answer quality and token cost.

AI agent core configuration

System Prompt— the instructions that define the agent's persona, tone, and rules (for example, "You are a professional support assistant; be friendly and concise."). Required. If you have saved prompt templates, click Load from template to fill this field in one click.
Temperature — how creative the replies are, from 0 to 2. Lower values give more focused, predictable answers; the default is 0.3.
Max History Turns — how many recent message turns the agent looks back on for context, from 1 to 100. The default is 20.
Max Concurrent Sessions — how many conversations this agent handles at the same time. Conversations beyond this cap are routed to the next available agent or wait in the pending pool. Your organization may set an upper limit on this number.
Welcome Message— an optional greeting sent at the start of a new conversation, before the AI's own reply. Leave it blank to skip the greeting.

Handoff settings

Handoff is how an AI agent transfers a conversation to a human agent. It applies to reception flows where the AI responds first and a human agent handles escalations. There are two complementary mechanisms, both configured in the AI agent's Handoff Settings section.

Keyword matching

Configure a keyword list (for example, human, complaint, manager). If a customer's message contains any of them, the conversation hands off to a human instantly — the AI does not run, so no tokens are spent. Matching is case-insensitive.

Smart Queue Routing

Turn on Smart Queue Routing to let the AI use conversation context to determine what the customer wants and transfer the conversation to the appropriate human agent queue. You configure two things:

Intent → Queue mapping — you can add multiple intents. Each intent requires two pieces of information: a short description that helps the AI recognize when the intent applies, and the target queue to route to when the intent matches.
Fallback Queue — where the conversation goes when the AI cannot tell which intent applies, or when the target queue has no human agents online. Required whenever Smart Queue Routing is on.

Keywords for speed, smart routing for accuracy

Keywords are best for unambiguous escalations a customer types directly — like "agent" or "complaint" — and cost nothing. Smart routing is best when you want the AI to interpret intent and send the conversation to the correct team. You can use both together.

Tool use

The system provides ready-to-use tools suited for common agent scenarios. You can also create more flexible custom tools on the Tool Calls page.

Built-in tools

System-provided capabilities you enable with a checkbox. Common ones include:

Tool	What it lets the AI do
Read conversation history	Recall earlier messages in the current conversation
Read user profile	Look up the customer's basic profile
Request handoff to human agent	Proactively ask to hand the conversation to a person
Add session note	Attach an internal note for later review (not sent to the customer)
Search knowledge base	Search content in knowledge bases — see the “Knowledge bases” section for details
Tag the user with an audience label	Attach a tag to the customer — see the “Contacts and tags” section for details
Add user to blacklist	When a user requests to unsubscribe, adds them to the blocklist — see the “Blocklist” section for details

Custom tools

External tools that communicate via MCP/HTTP, allowing AI agents to interact with external systems for information retrieval and action execution. For detailed descriptions, refer to the “Tool calls” chapter.

Some tools require matching settings

Authorizing Request handoff to human agent requires Smart Queue Routing to be on with a Fallback Queue set. Authorizing Search knowledge baserequires at least one knowledge base to be selected (see below). The form flags these in red and won't let you save until they're resolved.

Local knowledge bases

Using RAG (Retrieval-Augmented Generation), this section gives the AI agent access to local documents so it can answer based on enterprise-provided content. First create and index knowledge bases on the Knowledge Bases page; only active ones appear here. For details, refer to the “Knowledge bases” chapter.

Authorized Knowledge Bases — select every knowledge base this agent may read. An agent can only access knowledge bases you authorize here.
Auto-inject Top-K— each time a user sends a message, the system automatically retrieves the K most relevant snippets from the knowledge base and injects them into the AI's context. Range 0–5: 0 means no auto-injection — the AI can only read the knowledge base by calling the Search knowledge base tool. A higher number injects more snippets each turn, giving the AI more reference material, but also spending more tokens.

Auto-inject vs. the search tool

Auto-inject runs every turn and ensures the AI always has relevant context, but spends extra tokens whether or not the current question needs it. The Search knowledge base tool lets the AI search only when it judges necessary — leaner, but depends on the model's judgment. Enabling both gives the best coverage: a baseline context from auto-inject plus on-demand lookups from the tool.

Whichever you use, you must select at least one knowledge base. If Auto-inject is above 0 or the search tool is authorized but no knowledge base is selected, the form shows an error and blocks saving.

Adding the AI agent to a queue

Just like a human agent, once created it must be added to the appropriate queue before it can be assigned conversations. Refer to the “Routing groups and rules” and “Queues” chapters for details.

Editing and deleting

Editing an AI agent does not interrupt conversations already in progress; updates only affect conversations started after the change is saved. All conversations associated with an AI agent must be closed before the agent can be deleted.