The flow canvas
When to use a flow
Use a flow when the conversation has a shape you already know: an order-status check, a booking, a qualification questionnaire, a returns process. You get the same steps every time, in the same order, with the same wording. Use a Prompting Agent instead when customers arrive with open questions you can’t enumerate.The canvas
A Flow Agent opens straight onto its Flow tab. Drag nodes from the palette, connect the output of one to the input of the next, and click into a node to configure it. The canvas has its own Save and Publish controls — a flow doesn’t go live until you publish it.Because the canvas owns its own saving, the editor’s Test button is hidden while the Flow tab
is open. Leaving the tab with unsaved work raises Discard unsaved changes? first.
The node palette
Nodes are grouped by what they do.Start
Start
Every flow has exactly one Start node.
Send a message
Send a message
These say something and move straight on to the next node.
Ask a question
Ask a question
These wait for the customer. Buttons and lists each get one output per choice, so the flow
branches on what the customer picks. Ask Dynamic List builds its options at run time from
a system of yours — useful for available appointment slots or a live product list.The limits of 3 buttons and 10 list rows per section come from the messaging channels
themselves, and the canvas warns you when you exceed them.
AI
AI
These bring judgement into an otherwise fixed script. AI Collect keeps asking until it has
the variables you listed. AI Intent Branch reads free text and sends the conversation down
the branch that matches. AI Answer (KB) answers from your
knowledge base and nothing else, with a
No-answer reply for when the knowledge base has nothing.Route to Folder and Handoff to Human both end the flow — one passes the conversation to
whichever agent owns that folder, the other passes it to a person.
Logic & flow
Logic & flow
HTTP Request has two outputs — one for a successful call and one for an error — so you can
tell the customer something sensible when your system is down.
else is where conversations go to die.
Building a simple flow
Here is an order-status check, end to end.1
Greet the customer
Drag a Send Text node and connect it to Start. “Hi! I can check an order for you.”
2
Offer the choices
Add an Ask Buttons node with two buttons:
Track my order and Something else. The node
now has an output per button.3
Ask for the order number
From the
Track my order output, add an Ask Text node — “What’s your order number?” — and
save the answer into a variable.4
Look it up
Add an HTTP Request node that calls your order system with that variable, and keep the
response.
5
Answer, and handle failure
From the success output, add a Send Text node that reads the status back to the customer.
From the error output, add a Handoff to Human node so a person can pick it up.
6
Route the other branch
Send the
Something else output — and the else output — to Handoff to Human too.7
Save and publish
Nothing you draw affects live conversations until you publish.
When Hybrid is the right choice
A pure flow only knows the paths you drew. Customers don’t. Someone will answer “actually can I change the delivery address?” in the middle of your order-status branch, and a strict flow has nothing to say. The General tab of a Flow Agent has a switch for exactly this:Let AI handle off-script messages — When on, the agent follows your flow but falls back to the AI for anything the flow doesn’t cover (Hybrid). When off, it sticks strictly to the flow.
Leave it off when
The process is regulated, scripted or contractual, and an improvised answer would be a problem.
A strict flow never says anything you didn’t write.
Turn it on when
The flow covers the common path but customers wander. The agent stays on the script where the
script applies and answers freely where it doesn’t.
Next steps
Prompt, knowledge and tools
Needed for Hybrid fallbacks and AI nodes.
Folders and rules
Where Route to Folder sends a conversation.