Skip to content

AI agents overview

What an AI agent is, how it reaches customers through a flow, and how agents, knowledge bases and intents fit together.

An AI agent answers your customers' questions on WhatsApp from your own documents. It says when it is not sure, and hands the conversation to a person the moment it should. Agents live under Automation › AI agents.

Note: If you see "AI is turned off for this workspace", AI agents, knowledge bases, intents and the inbox copilot are not switched on for your organisation. Everything else works as usual. Ask whoever runs your IconEngage account to turn it on.

The three parts

Menu itemWhat it holds
AgentsYour agents: their instructions, which knowledge bases they read, tools they may call, how confident they must be, and where they hand over. See Build an AI agent.
KnowledgeKnowledge bases: the documents and web pages agents (and the inbox copilot) answer from. See Knowledge bases.
IntentsYour intent vocabulary: the kinds of question a customer message can be read as, such as order_status or refund. Flows can start on an intent, and agents can be limited to some. See Intents.

How an agent reaches customers

An agent never messages anyone by itself. It speaks to customers only through a flow:

  1. Build and publish the agent. Until it has a published version, it answers nobody.
  2. In a flow, add an AI agent step and choose the agent.
  3. Publish the flow.

When a run reaches the AI agent step, the agent reads the conversation and answers the customer's latest message once, then the flow carries on:

  • answered: the agent replied. Connect this exit to what should happen next, for example End.
  • handed over: the agent passed the conversation to a person (see below). If this exit is not connected, the run simply ends.
  • refused / failed: the reply couldn't be sent (for example, outside the 24-hour window) or the AI provider failed.

A common set-up is a flow that starts on Customer sends a message (or Intent detected), with an AI agent step followed by End: each new customer message starts a new run, and the agent answers it. Remember that a contact can be in only one run of a flow at a time.

The step always uses the agent's live version, so when you publish a new version of the agent, every flow that uses it changes from its next run on.

Confidence and handover

Every answer is scored against what the agent found in your knowledge bases:

  • Above the caveat line: the agent answers plainly.
  • Between the hand-over threshold and the caveat line: the agent answers with a caveat (a short sentence you can word yourself).
  • Below the threshold: the agent hands over.

An agent also hands over when the customer asks for a person, when the message is an intent it may not answer, when a tool fails, when your organisation's AI budget for the hour is used up, or when a policy rule says so. A handover assigns the conversation through your routing, optionally after telling the customer, and the flow's handed over exit decides what happens next.

What an agent will not do

  • It doesn't message customers outside a flow, and the flow's normal checks (consent, the 24-hour window, a person owning the conversation) still apply to its replies.
  • It doesn't see customers' phone numbers: they are masked before anything is sent to the AI.
  • It calls at most one tool per turn, and only tools you listed for it.

Who can do what

TaskRoles
See agents, knowledge bases and intentsOwner, Admin, Supervisor, Agent, Marketing Manager, Developer, Read Only
Edit agents, knowledge bases and intents; test agentsOwner, Admin, Supervisor, Marketing Manager
Publish and unpublish agentsOwner, Admin, Marketing Manager

To put an agent in a flow, you also need permission to edit flows (Owners, Admins and Marketing Managers).

Still need help?

Write to us at [email protected] and tell us what you were trying to do. Already a customer? Sign in to IconEngage — the same articles are under Help in the app.