Call flow

A script asks the agent to do things; the call flow guarantees them. In the agent editor’s Call flow tab you draw the steps of the call and how it moves from one to the next. Use it for what cannot go wrong: confirming who you are talking to before sharing any detail, reading a legal notice in full, or always transferring with the same sentence.

The steps

• AI stage: the agent talks with instructions, data and tools of its own for that stage. It can add instructions to the agent’s or replace them with a short, focused task, and you decide whether it can hang up from there.
• Fixed message: text the agent says word for word (a legal notice, a warning) or, from a guideline, in its own words.
• Tool: runs one of the agent’s HTTP requests every time the call goes through this step, so it must be safe to repeat. It continues through Success or Failed.
• Save data: stores a value to use later in a condition or a message.
• Transfer to a number: transfers with one of the agent’s Transfer call rules, with its introduction and its summary for the advisor. It continues through Bridged or Failed.
• Hand off to another agent: the call continues with another agent in your organization and this call flow ends.
• End: ends the agent’s part and counts it as Success or Failure, with an optional goodbye.

How the call moves between steps

Every connection says when it is taken:

• Always: as soon as the previous step ends.
• The AI decides: it is an exit of the stage. The AI takes it when the conversation matches what you write in “when”, for example “The person confirmed they are the account holder”. When it takes it, it can note data, such as the role: account holder or co-signer.
• If a value matches…: compares call data (equals, contains, is greater than, is a date before…). It does not depend on the AI.
• By result: of a tool or a transfer.

A global step can be reached from any AI stage without drawing a connection from each one: “the person asks for an advisor” or “the person becomes aggressive”. In Except from you pick the stages it cannot be reached from, such as identification.

Each stage sees only its own data

In a stage’s Data tab you choose which data it sees. A stage cannot say what it does not see, even if asked: the value is not in its instructions. That way, the stage that confirms who you are talking to only knows the name, and the amount or the date show up once the person has confirmed who they are.

In Settings and issues → Sensitive data you mark the values that must never be seen or said before that confirmation: amount, date, ID number. The editor won’t let you publish a call flow that shows them earlier, and it warns you if the agent’s greeting already says them, because the greeting plays before any stage.

Messages that are said in full

A Fixed message starts with “The customer can interrupt” turned off: it is said in full even if the person says “yes” or “uh-huh” halfway through. What they say in the meantime is not heard, so the message has to be short and, if it is set to wait for the customer’s reply, end with a short question so they know it is their turn. The editor shows how many seconds it lasts and warns you if it is long.

Start from a template

Templates → Collections with verification builds a collections call flow that confirms who it is talking to before sharing any detail:

1. 1
   Identification

   Confirms it is talking to the person, without seeing the sensitive data. If it is not them, or they are not available, it moves to a message stage that sees no data at all.

2. 2
   Ownership

   Confirms whether they are the account holder or a co-signer.

3. 3
   Notice

   Reads your notice word for word.

4. 4
   Negotiation

   Talks with the agent’s script and all the data.

5. 5
   Transfer

   Says a fixed sentence and transfers with your rule. If no one answers, the platform tells the customer with a fixed message and the conversation goes back to negotiation.

It asks for the value with the customer’s name, the sensitive data, the rule to transfer to the advisor, the notice text and the sentence before transferring. The stage texts are a starting point: review them and adjust them to your business.

Test the draft

Test the draft runs what you saved in a text conversation with a simulated customer, before you publish it. You can also open it with Test from this stage, in the panel of an AI stage. Tools and transfers are simulated: the test never runs anything for real.

1. 1
   Pick the case

   One of the agent’s test cases, or a quick case: who the customer is, what happens on the call and 1 to 3 things the agent must do. The quick case is saved so you can reuse it.

2. 2
   Pick how it runs

   In Start from you pick the start, with the greeting, or a stage. In What happens at each action you decide whether each tool and each transfer succeeds or fails.

3. 3
   Run the test

   You see whether it met each criterion, how it ended, the path it took through the call flow and the conversation with the stage of each turn. Fixed messages are marked.

Validate, publish and turn on

• Settings and issues lists what is missing. Errors block publishing: a stage with no exits that cannot hang up, a transfer with no path for Failed, sensitive data visible before verification. Warnings are decisions to review.
• Publish freezes a version: the next calls follow it and calls in progress do not change.
• On turns the call flow on or off. When it is off, the agent goes back to its script and the published version is kept.
• Versions keeps every publication: you can view it on the canvas or restore it to the draft to review and publish it again.
• A test call from the browser follows the published and turned-on version, just like a real call. The agent’s tests follow it too, with tools and transfers simulated: they never run anything for real.

With the copilot

• Ask the agent copilot (Improve with AI) or the main Copilot: "propose a call flow for this agent". It reads the script and builds a call flow that fits it: identification with no data and a verified confirmation, notices as fixed messages copied from the script, the main conversation with your script, transfers and endings.
• It shows you the proposal before saving it. When you confirm it, it stays as a draft and a few tests run comparing your current script with the call flow; the copilot explains the result with numbers.
• If your script asks for a notice to be said but does not include its text (for example, because it says the platform says it), the copilot does not make it up: it leaves that message empty, asks you for the exact text and the flow cannot be published until it is there. Once you give it, it completes the draft and runs the tests.
• It also reads and edits it with you: change the instructions or what a stage sees, a fixed message, add or remove steps, restore a version, create the WhatsApp one from the voice one or apply a template. Every change comes as a proposal you confirm.
• If you tell it the agent does not follow the script, shares data with the wrong person or cuts a notice short, it reviews your real calls and tells you whether a call flow fixes it. The agent review also suggests it when the script needs it.

The call flow on WhatsApp

Each agent has two flows: one for calls and one for WhatsApp. At the top of the tab you choose which one you edit with Voice | WhatsApp. Each has its own draft, its own versions and its own publishing, and publishing one does not change the other. If your organization does not have the WhatsApp flow yet, the lens tells you.

1. 1
   Start from the voice flow

   In the WhatsApp lens, Copy from the voice flow brings what you saved for voice, adapted to writing, and tells you what changed. «Transfer to a number» becomes Hand off to a person, «The customer can interrupt» is removed, and messages that talk about the call or the recording are flagged. You can also Start from scratch.

2. 2
   Rewrite what belongs to a call

   Open each flagged step with See the step. On WhatsApp the agent does not greet before the flow: the first stage starts with the person's first message, so it should introduce itself. Ask for amounts to be written in figures: on a call they are said in words, but in a chat they read badly.

3. 3
   Test and publish

   Test the draft runs a WhatsApp conversation with a simulated person who writes. When it looks right, tap Publish: the agent's next WhatsApp conversations follow that version, and the ones in progress do not change.

• Hand off to a person passes the conversation to a team member, a team or the queue, with a summary of up to 240 characters for whoever takes it. With As in the channel settings, it follows the agent's WhatsApp handoff destination.
• An End needs a closing message: without it, the conversation would stay open.
• Tool and Transfer to a number do not run in writing yet: the validator will not publish the WhatsApp flow while it has them.
• Each conversation's data comes from the row of the campaign that opened it or from the contact's attributes. If one is missing, the stage sees it as «not available»: a variable's default value is for testing and never reaches a real person.
• Each conversation follows the version it started with. If 24 hours pass without a reply, the team closes it or it moves to another agent, the next one starts again from the beginning, with the identification.
• Identity is confirmed by the reply to the agent's question. If the person introduces themselves in their first message ("hi, this is Laura, how much do I owe?"), the agent still asks before showing any data: that first message does not answer anything the agent asked.

What happens at each step

Analytics paints on the canvas what happened in the calls of the published version: how many times each step was entered, how long calls stay on average, how many calls ended there and which connection they left through. The How the calls ended panel separates the ones that succeeded, the ones that did not and the failures, with their reason.