Voice Outbound Action

Every other node in this module answers. This one starts the conversation: the workflow decides somebody should be called, and their phone rings. A lead filled in a form, a payment failed, a delivery is late — all good reasons, and all of them reach a person who did not choose that moment.

Which is why this node is the most heavily guarded in the product. It can dial one number or a thousand, with an AI agent or a member of your team on the line. Nearly everything in its configuration exists to bound that: who may be called, when, how often, and how quickly.

call_input is the field that matters and it takes three shapes: one object with a phone number and the variables the agent should know, an array of them, or a reference to an array an upstream node produced. The last is how a campaign is built — a search returns contacts, and this node calls them.

You will also see a phone_input field in older workflows. It is deprecated and call_input replaces it. Do not start anything new on it.

Three of these are on before you touch anything, and knowing which is the difference between a compliant campaign and an expensive one.

Turn answering-machine detection on for any AI campaign. With it off, voicemail is indistinguishable from a person to the agent — it talks, records a call that reached nobody, and bills for the minutes. With it on you also choose what to do: continue leaves a message, hangup does not. Either is a decision; the default is not.

Business hours are an enum, not free text. The timezone list is fixed — nine zones — so if the one you need is not there, that is the answer rather than a typo to fix. The windows are a list, so a lunch break is two windows rather than one with a hole in it.

This is where people lose an afternoon. The node has three time limits and they measure different things.

The automatic default for the middle one is worth internalising, because three minutes is short. On Exotel, a workflow that waits for a call and does something with the transcript will stop waiting partway through an ordinary conversation unless you set the value yourself. For batch campaigns it is the setting to get right first.

In batch mode the node paces itself: three simultaneous calls by default, five seconds between them, and a failure does not stop the run. Raise the concurrency and you raise the chance of your provider throttling you and of several people picking up to silence because the agent is busy. Three is a sensible number.

Retries are off by default and worth leaving off until you have thought about them. When enabled they back off exponentially, and the base delay cannot go below two minutes — a limit of the scheduler rather than a policy. Consider that a floor and set something much larger; retrying a missed call two minutes later reads as harassment, not diligence.

A dropped number is not a failure. Numbers removed by the Do Not Call check simply do not appear in the call list. The run succeeds, the batch is smaller than your audience, and nothing says why. If a campaign reached fewer people than you expected, compare the count going in against the calls placed before you go looking for a bug.

The workflow pauses per call, and the pause is where batches get expensive. Like the inbound node, this one parks the run until the call completes. In a batch that is fine; in a loop calling one number at a time it means the workflow is held open for the sum of every conversation. Use batch mode rather than a Loop around a single call.

Caller ID is what the customer sees, and it decides whether they answer. Fixed mode picks one number; dynamic lets a variable choose per call, which is how a regional campaign shows a local number. Neither is the number the call comes from technically — it is the number displayed — so a caller ID nobody can ring back is worse than none.