You write the steps; the AI agent handles the conversation around them. It keeps to your order and rules while responding to what the customer actually says.
When a procedure starts
Each customer message is compared with the When to use and When not to use of your published procedures on that channel. The closest match starts. If none matches, the AI agent answers as usual.
If a customer asks for something another procedure handles while one is waiting for an answer, the new procedure takes over.
What the customer can do mid-procedure
The customer… | What happens |
Gives details before they're asked | Those questions are skipped |
Asks a side question ("do you accept Bupa?") | The AI agent answers it, and the procedure keeps waiting for its answer |
Changes an earlier answer | The procedure goes back to that question and continues from there |
Says "never mind" or "forget it" | The AI agent acknowledges it in one short sentence and the procedure stops |
Asks for several things in one message | They're handled one after another, up to 3, and the reply says what's next |
Writes in another language | Replies follow the customer's language |
Gives an answer that doesn't fit | The question is asked again (up to Max attempts, 2 by default), then If there is no answer applies |
If the customer changes an answer after the procedure already made a change, such as a cancelled booking or a created ticket, the conversation is handed to your team instead.
Questions with options
On WhatsApp, options and pick lists show as a list to tap, with up to 10 rows. On other channels they show as a numbered list. The customer can reply with the option's text or its number.
Confirming changes
Before any Write action, the AI agent shows the customer what will change and waits for a yes.
Yes: the change runs and the procedure continues.
No: nothing changes. The customer is told so in one short line, and the procedure stops, unless a later condition checks that action's result. Then it continues there.
Tickets, tasks, notes and topics don't need confirmation.
When the AI can't decide a condition
If a natural-language condition can't be decided from what the agent knows, it asks the customer one short question. If it still can't tell, it hands off.
What the customer sees
The AI agent writes every message in your business's voice and the customer's language. Your lines say what to tell the customer, not the exact words.
Hand-offs
The conversation goes to your team when:
the procedure reaches a Hand off to a team step,
a Read or Write action fails and no condition handles the failure,
a question gets no usable answer and If there is no answer is Hand off, or
the procedure hits an error it can't recover from, a condition still can't be decided after its one question, or the customer changes an answer after a change was already made.
Any other requests waiting in the same message are left for them.
Test, publish and track your procedures
Test every procedure on its draft before you publish it. After publishing, track how it performs from the procedures list.
Test a draft
In the editor, click Test. The Test this procedure panel opens: "Runs your latest draft. Nothing is sent to customers."
Under Select a contact to test, pick a contact. The test uses that contact's details, such as their name and phone number.
Write a message as the customer. The panel shows Preparing the draft… while it starts a test conversation.
Click Restart to begin a new test conversation.
Test is unavailable while a line is marked in red, or a line has no step yet. If the draft has a problem across lines, the panel shows This draft can't run yet with the reason.
What a test does and doesn't do
The procedure starts straight away. When to use isn't checked.
Read and Write actions run against your real connected systems. Say no at the confirmation if you don't want a real change.
Notes, topics, contact tags and contact attributes aren't saved. Tickets and tasks aren't created: in a test those steps fail, so a branch that handles a failed ticket runs.
Test conversations don't count in your procedure's results.
Sub-procedures have no Test button. Test them through a procedure that calls them.
Try every path: each branch, a customer who gives everything in their first message, one who changes an answer, and one who says no at a confirmation.
Publish
Click Publish. If the procedure is already live, the button reads Publish changes.
Publishing checks the whole draft again, then saves it as a new version. The status shows Published · vN.
You'll see: "Published. The editor reopened on the new version."
New conversations use the new version straight away. Conversations already inside the procedure finish on the version they started with.
Edit a live procedure
Your edits save as a draft while the live version keeps running. A dot marked Unpublished changes appears next to the status until you publish again.
Unpublish or delete
From the … menu in the editor, or More actions on the list:
Unpublish: "It stops starting new conversations right away. Conversations already inside it finish, and your edits stay as a draft."
Delete: "This can't be undone. Conversations already inside it finish."
A sub-procedure that published procedures still call can't be unpublished or deleted. See article 3.
Track results
The procedures list shows two counts for each procedure:
Triggered: times it started.
Resolved: conversations it resolved.
Click a count to see those conversations. Each triggered conversation shows its result: Resolved, Handed off, Abandoned or In progress.