Workflows
Build your agent's behavior using visual, step-by-step graphs
Workflows are the building blocks of your agent. Each workflow is a visual graph that defines how your agent processes input and responds to users — step by step, node by node. An agent can contain many workflows, each responsible for a different part of its functionality. Think of them as modular pieces of logic you compose together.
Navigate to Designer > AI Workflows to build, configure, and manage your workflows and everything connected to them.
What Is a Workflow
A workflow is a directed graph made up of two elements:
Your agent executes the graph step by step, starting from the Start node and chaining through each connected node. A workflow ends when it reaches a node with no outgoing edge or a Redirect node that transfers execution to another workflow.

Types of Workflows
When you create a new workflow, you select one of four types. Each type comes with a preconfigured template and may enable specialized nodes.
The standard workflow type. Build custom conversation logic using any combination of message, prompt, action, and logic nodes.
Designed to collect structured information from users step by step. Use for surveys, applications, or any scenario requiring sequential data collection.
Collects user feedback about the conversation. Keep the default tags so results appear in Observatory analytics.
Facilitates handoff to a human agent for real-time communication. Activate LiveChat from Designer > Plugins before use.
Creating a Workflow
Managing a Workflow
You can manage a workflow using the Action column buttons.
Edit
Open the workflow on the canvas
Delete
Permanently remove the workflow
Clone
Duplicate the workflow into the same or a different agent
Preview
Run the workflow in isolation
Favourite
Set this workflow as the default for the flow preview
Deleting a workflow is permanent. Make sure you no longer need it before you select Delete.
Editing a Workflow
Use the Canvas to design your workflow visually. Simply drag and drop nodes, then connect them to define the execution path. To edit a workflow, either select it in Designer > AI Workflows > Flows or click the respective edit action button.
The canvas enforces various rules to keep your workflows valid. For example:
Every node must be connected; you cannot leave detached nodes in the workflow
Only valid connections between compatible node types are allowed
A node cannot have multiple outgoing edges to different children (unless using a control flow node)
If you try to exit the canvas with errors, you get a warning highlighting the issues. You need to fix all errors before saving a workflow.
The Workflow Lifecycle
Each workflow executes according to the visual path established in the graph.
Step-by-Step Execution
The workflow chains through nodes one at a time, following the edge direction. At each step, the next node is determined by the outcome of the current one. If a node branches into multiple outgoing paths, the workflow follows a single path based on the result. The result is an execution flow that adapts to each conversation rather than following a fixed sequence.
A single agent response can span multiple workflows chained through Redirect nodes. The response completes when the last workflow in the chain reaches a terminal node.
How the Agent Picks a Workflow
Every agent comes with built-in triggers that decide which workflow runs at a given moment:
Each one targets a workflow you have built, but they differ in what triggers them, where they apply, and where you configure them:
Purpose
Greets the user when a conversation begins
Responds to any new user message
Redirects the conversation when a keyword matches
Trigger
Conversation opens (before any user input)
User input
User message matches an Interrupt keyword
Scope
Per deployment
Global (applies to all deployments)
Agent-wide
Where to set
Deployment settings
Designer > AI Workflows > Default Flow
Designer > AI Workflows > Interrupts
Availability
Depends on the channel
All channels
Modern agents only
Start Workflow
The start workflow is the initial workflow triggered the moment a user begins a new conversation. It defines the agent's first impression, allowing you to establish context or gather information before the main logic takes over.
Understanding the Start Workflow
This workflow is triggered when the user first accesses the agent, such as when they open a webchat interface in their browser. Because this workflow can be any flow you have built, it offers a high degree of flexibility for personalizing the user experience.
Common uses for a start workflow include:
Send a personalized welcome message
Set specific variables to track user context
Ask for initial feedback or user preferences
Introduces the agent's capabilities
The start workflow is set per deployment, not globally. Different deployments can trigger different workflows, even on the same channel.
Channel Availability
The availability of the start workflow varies depending on the specific channel you choose for deployment. Some channels do not support the concept of a 'new conversation' session, while others restrict introductory messages to a one-time initial interaction.
Webchat
✅ Yes
API
❌ No
Microsoft Teams
❌ No
Slack
❌ No
❌ No
Viber
✅ Yes
Messenger
✅ Yes
❌ No
Unity
❌ No
Configure a Start Workflow
You choose the start workflow when you create a deployment in Designer. If you need to update the initial experience, you can change the assigned workflow at any time from the deployment configuration settings.
Default Workflow
The default workflow is the main starting point of your agent. It is the central sequence that triggers automatically whenever a new message is sent to the agent, ensuring a responsive interaction for every user query.
To configure the default workflow, go to Designer > AI Workflows > Default Flow.
Understanding the Default Workflow
When a user sends a message, the default workflow is responsible for handling the agent's response. It is defined as a chain of responses that are executed sequentially and you can configure each response separately by selecting one of the four types.
Display customized messages to your users with full support for rich formatting, including emojis and lists. You can personalize the response by including agent-specific variables such as user metadata or the current date and time.
For granular control, you can enable or disable thumb ratings for specific messages, allowing you to override your global agent settings.
Redirect the control to the selected workflow. This is typically your main workflow that orchestrates your agent behavior. You can only redirect to Generic, LiveChat or Survey type workflows.
This is only available as the last response item in your default workflow sequence
Redirect the control to the selected User Feedback workflow. For any other type of workflow, use the Redirect response. You can set the probability to ask for feedback (0-100%).
Define variables to manage your agent’s behavior. Select an existing variable or create a new one by typing the name and then assign a new value for the variable.
You can also choose how the data is stored: as a session variable for the current conversation or as a contact variable for long-term storage.
There is only one default workflow allowed for each agent to maintain a clear and consistent logic path.
Build a Default Workflow
Interrupts
An Interrupt redirects the conversation to a target workflow when a user message matches a trigger keyword. Use them for global escape phrases like cancel, start over, or talk to a human that should work no matter where the user is in a conversation.
To configure interrupts, go to Designer > AI Workflows > Interrupts.
Interrupts are available on modern agents only.
How Interrupts Work
You define an Interrupt by pairing one target workflow with a list of trigger keywords. When a user message arrives, the agent compares it against every keyword in every Interrupt before any other LLM step runs. If a keyword matches, the conversation jumps to that Interrupt's target workflow.
Matching is controlled by two agent-level settings that apply to every Interrupt: a similarity algorithm and a threshold.
The user message must equal a keyword exactly. The threshold is ignored. Best for short, unambiguous commands like cancel or stop.
A fuzzy similarity algorithm tuned for short strings and common prefixes. Use the threshold to tolerate typos and minor variations (1 is exact, 0 accepts anything).
A fuzzy similarity algorithm that counts character edits and transpositions. Use the threshold for tolerant matching of phrases or longer keywords (1 is exact, 0 accepts anything).
Toggle Enable Interrupts off to suspend matching across the whole agent without losing the keyword lists.

A few rules to keep in mind:
Each workflow can be the target of only one Interrupt
A keyword can appear in only one Interrupt; duplicates are rejected on save
Deleting a workflow automatically removes any Interrupt that targeted it
Create an Interrupt
Add Keywords
Type a keyword and select Add Keyword, or use Import from CSV to load a batch from a file. Imported lists are deduplicated automatically against the keywords already in use by other Interrupts.
Manage an Interrupt
Each row in the Interrupts table has edit and delete actions. Edit lets you change the target workflow or update the keyword list. Delete removes the Interrupt entirely. Inside the editor, Delete All clears the keyword list without removing the Interrupt itself.
Workflows as Nodes
The Redirect node allows you to trigger a workflow from within another workflow, effectively turning complex sequences into modular "building blocks." This architectural approach is essential for building scalable and maintainable agents.
For example, imagine you have a main routing workflow that acts as a switchboard. Instead of building every single response in one place, you use Redirect nodes to send the user to specialized flows:
Main Workflow → User asks about a package → Redirect to Order Tracking Flow
Main Workflow → User asks to return an item → Redirect to Returns & Refunds Flow
Main Workflow → User requests a human → Redirect to Live Support Flow
Best Practices
Keep workflows focused: each workflow should handle one task or conversation path, not the entire agent logic
Name descriptively: use names like
Order TrackingorPassword Resetinstead ofFlow 1orTestSet a default workflow: always configure a default workflow to handle unrecognized input gracefully
Match the channel: a Webchat starting workflow can be more visual, while a Viber starting workflow should stay text-focused
Document your workflow: Sticky notes are available in the canvas sidebar. They are useful for documenting and organizing a workflow and facilitating collaboration between team members.
You now understand how to create, manage, and connect workflows to build your agent's conversational logic.
Last updated
Was this helpful?


