The Channels column
A workflow belongs to one brand and runs for that brand on every channel the brand talks on. Open Brand → Workflows and you will see a Channels column with five icons on every row:
- A speech bubble — the chat widget. Always available.
- The Facebook logo — Facebook Messenger.
- The Instagram logo — Instagram messages.
- An envelope — email conversations, the ones your brand's AI answers on its inbound email address.
- A telephone — phone calls answered by one of the brand's voice agents.
A greyed-out icon means this script cannot run there. Hover it and it tells you why. Chat is never greyed out; the other four depend entirely on which steps you used.
A workflow a channel cannot run is never offered to the AI on that channel — the AI simply answers the person itself. (This used to be true only on the phone: on Facebook and Instagram the AI could still pick a form workflow, the script was then refused, and the person got no reply at all.)
Exactly what rules a channel out
One incompatible step is enough to take the whole workflow off a channel — it is not step-by-step, it is all or nothing.
| Channel | Ruled out by |
|---|---|
| Facebook and Instagram | A Start a form step |
| An Ask a question, Buttons, Start a form or Send to VEX AI step, or a Collect data step for anything other than the customer's email | |
| Phone calls | A Start a form, Open link, Buttons, Show expected reply time, Request agent, Collect data, Ask a question or Stop — don't reply step |
The reasons are practical rather than arbitrary. A form is a panel you fill in, and Facebook and Instagram have nowhere to put one. On a call there is no screen, so buttons and links have nothing to render into, there is no person to hand a caller to, and a caller cannot be answered with silence. Collect data and Ask a question are ruled out on calls for a different reason: the phone side cannot yet park a script mid-question and pick it up again when the caller answers.
Email has the same problem in a stronger form: a later email never continues a paused script, it starts a fresh turn with the AI. So any step that waits for the person's answer would wait forever. Collecting the customer's email is the one exception — the address is already known from the sender, and the step fills itself in without asking.
Compatibility is also inherited through the Start a workflow step. If the script you call cannot run on the phone or by email, neither can the one calling it.
A Call an API step never rules a channel out: it runs in chat, on Facebook and Instagram, by email and on phone calls. On every channel, nothing from its reply reaches the person unless a later step prints one of its values.
Which trigger can start where
The steps decide where a workflow can run; its trigger decides where it starts by itself. The Starts when card at the start of the workflow sums up both in one line: "Starts by itself in: …".
| Trigger | Starts by itself in |
|---|---|
| User says | Website chat, Facebook, Instagram, email and phone calls |
| Rating given | Website chat, Facebook and Instagram |
| Nobody picked up | Website chat, Facebook, Instagram and email |
| Chat open, Chat closed, Visit, Online since, Custom event | Website chat only |
| No automatic start | Nowhere — your team, another workflow, a button in your chat widget or your website starts it |
How a workflow reaches someone who emailed you
On an email conversation a workflow does not send a message per step. Everything the run produces — messages, the expected reply time, a link, and the confirmation and offline notice of a Request agent step — is collected and sent as one email at the end of the run, inside the thread the person already has with you. Their reply lands back in the same conversation.
Two things are deliberately left out of that email: the "written by AI" line, because a workflow is your script rather than the AI's words, and the "Talk to a human" button. In your inbox the sent rows carry the envelope label, like any other email you send. If the email cannot be sent — no verified sender address, no visitor email, a mail-provider error — a private event in the inbox says so, and the messages shown above it were not delivered.
A workflow whose only step is Stop — don't reply sends nothing at all, and the conversation stays with the AI instead of landing in your team's queue — the intended way to ignore sales pitches and spam that arrive by email.
What the product tells you, and where
You get told twice, which is deliberate:
- While you draw. In the editor, every step that rules a channel out carries a small warning icon; hovering it names the channels, for example "Not on: phone calls, email" on a Buttons step. The Starts when card sums it up for the whole workflow, together with its trigger: "Starts by itself in: …".
- After you save. The Channels column on the Workflows list shows the verdict for the finished script. Hovering a greyed-out phone icon says: "Cannot run on phone calls — it contains a form, link, buttons, reply-time, request-agent, collect-data, ask-question or stop step"; a greyed-out envelope says "Cannot run on email — it contains a question, buttons, form, Vex AI or a collect-data step other than the customer's email".
These are recalculated every time you save, so fixing a step and saving is enough to get the channel back.
Building a workflow that works on the phone
A phone-safe script is a conversational one. In practice that leaves you these steps:
- Write a message — the agent speaks the line.
- Conditions — branch on the caller's country, email, first visit, visit count, or whether your team is online.
- Call an API — look something up in your own system; a following Write a message step speaks the values back. On a call it waits at most 4 seconds and tries once.
- Send to VEX AI — hand the question to the brand's AI.
- Start a workflow — as long as the script it calls is phone-safe too.
The design shift is that you stop offering choices and start asking plainly. Instead of a Buttons step with four options, write a message that asks the caller what they need and let the AI or a Call an API step take it from there.
Two extra conditions on the phone
Even a perfectly phone-safe workflow needs two more things to be true before it is ever offered on a call:
- Its trigger must be "User says". Chat open, Visit, Chat closed and the other triggers describe things that never happen on a call (see the table above). On a call, the AI decides to run a workflow because the caller said something that matches its intent — so the intent description you write on that trigger is what actually gets it picked.
- The brand must also have a Vex AI agent. Workflows and your own API tools are only offered on a call when the brand has a chat AI as well. Knowledge lookup and ending the call always work; those two do not. Most people never notice, because a brand that has phone agents usually has a Vex too.
Facebook and Instagram: what changes
Everything except forms works, but a few steps look different because Meta's messaging has its own controls:
- Buttons arrive as tappable quick replies rather than buttons in a chat window.
- Collect data with a dropdown also arrives as quick replies, and two things can change that. With more than thirteen choices the quick replies are dropped altogether and the options are written out as a numbered list instead, so the visitor answers with a number. With a choice longer than twenty characters the quick replies still appear, but the full wording is also written out as a numbered list underneath so nothing is cut short on the chip.
- Answers are checked. A malformed email, a non-number, or a choice that is not on your list gets a short "please try again" reply rather than being accepted silently.
- Open link is sent as plain text, because the link card belongs to the chat widget.
- Start a form, if a script somehow reaches one there, tells the visitor the form is unavailable, requests a human agent, and ends the workflow.
Next step
For what each step actually does, see Workflow nodes explained. To understand what a voice agent can and cannot do on a call, see What your voice agent knows, and where that knowledge comes from.