Workflow builder
A voice agent workflow builder your whole floor can read
Draw each step of a call as a node, write the rule that moves it on in plain words, and hear it in your browser before a single customer does. Publish a named version, and live calls keep running the last one until you do.

What is a voice agent workflow builder?
A voice agent workflow builder is a canvas where you lay out an AI phone call as connected steps, each with its own instructions, and write the conditions that move the call from one step to the next.
In Bananaflow each step is a node, each condition is a transition written in plain words, and the whole flow runs as a live voice agent in English and 10 Indian languages.
New to the idea? Read what conversational AI is and how one voice turn works.
The builder sits inside a full platform: see the platform around the builder, a conversational AI platform for Indian calls.
Flows of states and transitions are an old idea. The W3C describes the model in State Chart XML (SCXML), its standard for state machine notation.
How do you build a call flow in Bananaflow?
You start from a Start Call node, add an Agent node for each stage of the call, join them with transitions written in plain words, and end at an End Call node. Four steps.
Open the call with Start Call
Set the greeting as text or as your own recording. Add a pre-call data fetch if the agent should know the caller before it speaks.
Add an Agent node for each stage
Give each one a single objective, its own tools and its own knowledge base files. A Global node sets the persona and tone once for every Agent node.
Join them with transitions
Each transition has a label, a condition in plain words, and an optional line the agent says as it moves on. A transition can loop back to ask again.
End with End Call and send the result out
End Call says goodbye and hangs up, and can extract variables. A webhook sends them to your CRM when the call ends.
Want to talk through your own call flow?
What does the workflow builder do?
Six things that take a call flow from a sketch to something your team trusts on live calls.
Can AI build the workflow for you?
Yes: describe the call in the Build with AI panel and it adds the nodes, prompts and transitions, creates the tools it needs, and saves the agent while you watch the canvas change.
"Add a node that answers policy questions, and a transition for when the customer wants a person."
- A new agent opens with the Build with AI chat beside the canvas.
- It can create HTTP tools and integration tools, set the voice, and read your call transcripts to suggest edits.
- You can still drag nodes, edit prompts and undo or redo by hand at any time.
How do you test a voice agent before customers hear it?
You call it from your browser by voice or chat with it by text, and the canvas highlights the node the call is on at every turn.
The call moves from Confirm renewal to Answer questions, and the view pans to follow it.
- Set test values, such as a customer name and a due date, in Context Variables.
- Set direction to outbound to test the outbound path in the browser.
- In the text tester, edit one turn and rerun it to see a different path.
What happens to live calls when you change the flow?
Nothing until you publish: you edit a draft, live calls keep running the published version, and every published version keeps its name, its note and a diff against the one before.
"v3 – softer opening for late payers", with the one changed prompt marked in the diff.
- Name a version up to 100 characters and add a note of up to 500 on what changed.
- A side-by-side diff steps through each change, next and previous.
- A publish checks that every node and connection is complete first.
How does the agent use your customer data?
Before the agent speaks it can fetch the caller's record from your CRM, and it reads those fields in its prompts as template variables.
"Namaste {{customer_name}}, your policy {{policy_no}} is due on {{due_date}}."
- Fetch on every call, or on inbound only, or on outbound only.
- Your endpoint can use an API key, a bearer token, basic auth or a custom header.
- If your system is slow, the caller hears a ring tone until the record lands.
Can the agent use your own recorded voice?
Yes: record or upload clips in the browser, drop one into a prompt with @, and the agent speaks generated replies for anything off-script.
A Hindi greeting recorded by your best agent, then generated speech for the rest of the call.
- Check the transcription of each clip before you use it.
- Clips are shared across your organisation.
- A word list helps speech-to-text hear your brand names and terms.
How do you check the calls the flow makes?
Add a QA node to score a share of your calls, and read the extracted variables and the summary after each one.
Score 20% of calls longer than 30 seconds, and skip voicemail.
- The QA node sets a minimum call length and a sample rate.
- It returns tags, sentiment and a summary for each scored call.
- Agent and End Call nodes extract variables, which go out by webhook.
Writing the questions your agent asks? See how AI lead qualification calls work.
A renewal campaign, built on the canvas
Rohan Mehta runs a 60-seat team at a BPO in Noida. His insurance client wants 4,000 consented customers called about policies due for renewal in November, in Hindi and English.

Monday, 10:00 am
He types the brief
A new agent opens on Build with AI. Rohan describes the call. It adds a Global node for the persona, a Start Call node, two Agent nodes, Confirm renewal and Answer questions, and an End Call node. It also creates an HTTP tool for the client's policy lookup.
10:40 am
He connects the client's CRM
He sets the pre-call fetch to outbound only, so the prompt can say the customer's name, policy number and due date.
11:15 am
Anjali records the greeting
Anjali, his best Hindi agent, records it in the browser. Rohan drops @renewal_greeting_hi into the Start Call node.
12:00 pm
He writes the rules that move the call
He adds a transition called wants a person that hands the call to his team, and a loop called asks again that sends the call back to Confirm renewal.
2:30 pm
He tests it by voice
He sets customer_name to Sunita, due_date to 14 November and direction to outbound, then calls the agent from his browser. The canvas lights Confirm renewal, then Answer questions. He edits one turn in the text tester and reruns it.
Next week
He publishes, then improves
He publishes "v2 – first live". A week later he publishes "v3 – softer opening for late payers", opens the diff to check that only the Start Call prompt changed, and splits the campaign between v2 and v3. A QA node scores 20% of calls over 30 seconds.
Running the whole campaign? See AI outbound calling for Indian call centers.
How does the builder work with the rest of Bananaflow?
The flow you draw is the agent that takes calls on your own numbers, runs in your campaigns and feeds your reviews.
Your carrier and your numbers
Connect Twilio, Exotel, Plivo or any SIP trunk, and the published flow answers and makes calls on it.
Campaigns, by version
A campaign can split its calls between two versions of the same agent, so you compare a change before you roll it out.
Call review
Every call keeps a transcript and a recording. A QA node adds tags, sentiment and a summary.
Prices cover the voice stack. Your carrier bills its own call charges. See Bananaflow pricing.
AI outbound calling
Run a consented list with retries and calling windows.
Telephony and SIP
Connect your carrier and pick the agent that answers each number.
Calling campaigns
Upload a CSV and let the agent call the whole list.
Call QA
Score each finished call and tag what went wrong.
AI lead qualification
Ask your own questions and book the next step.
Integrations
Give your agent tools it uses while the caller waits: your APIs, MCP servers and apps.
FAQ
Workflow builder questions, answered
Do I need to code to build a voice agent?
No. You build on the canvas or in the Build with AI chat. Code only comes in if you want your own API tools or the REST API and SDKs.
Can a call go back to an earlier step?
Yes. A transition can loop back, so the agent can ask a question again or return to an earlier stage.
Will editing my agent break calls that are running?
No. You edit a draft, and live calls keep running the published version until you publish the next one.
Can I see what changed between two versions?
Yes. Every published version keeps a name and a note, and a side-by-side diff shows each change against the version before.
Can the agent use my own recorded voice?
Yes. Upload or record clips, insert one with @ in a prompt, and the agent speaks generated replies for anything off-script.
How is a workflow different from a single prompt?
A single prompt holds the whole call in one instruction. A workflow gives each stage its own short prompt and clear rules for moving on, which makes it easier to test and to change.
Bring us one process. We'll pilot it on your real calls, with your team as the fallback.
Plan a pilot