Letting your AI agent run workflows

Your AI agent answers customers in live chat and on every other channel that it is assigned to. It can answer questions, ask for details, send links and transfer the conversation to a person. A workflow lets it do work outside the conversation: book a call-back, send the issue to your team as a ticket, add a hot lead to your pipeline or look up an order.

Use a workflow only for work that the agent cannot do on its own, or has not been set up to do. If the agent can already answer from its knowledge, it does not need a workflow.

• You are a property admin. Only admins can create and manage workflows.
• AI Assist is set up, with at least one AI agent.
• You know the basics of workflows. See Getting started with Workflows.

Tip: The fastest way to start is a template from the AI agent category: Escalate to a human, Qualified lead hand-off, Request a call-back or Order status lookup.

There are three ways:

• As a tool: you give the workflow an AI agent trigger. The agent decides when to use it during a conversation and fills in the details, for example the visitor's phone number.
• From a Scenario: the workflow starts in the background each time the scenario matches.
• When the agent cannot answer: the workflow starts in the background from the agent's Escalation settings.

The next sections explain each one.

The Tool description tells the agent when to use the workflow and what to ask first. A good description has three parts:

• The task: "Book a call-back…" or "Send the conversation to the support team…"
• When to use it: "…when the visitor asks for a call" or "…when you cannot solve the issue."
• What to ask first: "Ask for their email address first."

For example:

Look up the status of an order. Ask the customer for their order number and the email address they used at checkout first.

Inputs are the details that the agent collects from the conversation and passes to the workflow. If you add no inputs, the agent sends a short summary of the conversation instead.

1. Under Inputs the agent should provide, click Add input.
2. Type a Name, for example orderNumber.
3. Type a Description, for example The order number the customer mentioned.
4. If the workflow cannot work without this input, turn on Required.
5. Repeat steps 1 to 4 for each input.

Each name must be unique, start with a letter and use only letters, numbers and underscores. You can add up to 10 inputs. Write each description as an instruction to the agent: say exactly what goes in the input.

Each input becomes runtime data under trigger, with the name that you gave it. The run also knows which conversation started it: {{trigger.conversationId}} and {{trigger.channel}}.

For example, a call-back task can use:

• Activity title: Call back {{trigger.name}} on {{trigger.phone}}
• Description: Preferred time: {{trigger.preferredTime}} and Reason: {{trigger.reason}} on two lines

For all the details that a run can use, see Using runtime data in workflows.

Some tasks must happen only one time in a conversation, for example a new ticket. For these, turn on Run at most once per conversation. If the agent tries to start the workflow again in the same conversation, it gets the earlier run instead.

1. Add the steps that the workflow does.
2. In Settings, type a clear Name and Description.
3. Click Create.
4. Turn on Enabled.
5. Click Save.

For every step that you can add, see Building workflow steps: actions, conditions, waits and AI steps.

Important: If you selected All AI agents, every agent can use the workflow as soon as you enable it. To try it with one agent first, select Custom agents.

If you see one of these labels next to a workflow, the agent cannot use it yet:

• Not agent-enabled yet: the workflow has no AI agent trigger.
• Workflow is disabled: turn on Enabled in the workflow's Settings.

A background workflow starts without the agent choosing a tool. The agent fills in its inputs from the conversation.

• From a Scenario: in the scenario's settings, select a Background workflow (optional). The workflow starts each time the scenario matches. See Creating Scenarios for your AI agent.
• When the agent cannot answer: in the agent's Escalation settings, select a Background workflow (optional). The workflow starts when the agent cannot answer.

In both cases, the workflow must be enabled and available to the agent. If it is not, the agent settings show a warning and the workflow does not start.

1. Open the chat preview of your agent, or the bot tester.
2. Start a conversation that should start the workflow.
3. Open the workflow's Runs tab. Runs from your agent show Agent as the trigger. Runs from the preview show Agent test as the source.

Important: Workflows that start during agent tests are real runs. The preview shows This runs the workflow for real. If the workflow creates a ticket or sends a message, it really does it.

To read a run and fix errors, see Testing, monitoring and troubleshooting workflow runs.

A workflow uses AI credits in these cases:

• AI steps use credits each time they run.
• Background workflows use credits each time a Scenario or the Escalation setting starts them. One model call fills in the workflow's inputs from the chat, and that call uses credits.

A workflow with neither uses no credits.

Have a question we haven't covered or need support?

• Chat with us (live chat is available 24/7)
• Visit our Feedback page
• Schedule a call with us