Appearance
Create Your First Voice Agent
This guide takes you from an empty account to an agent that is ready for a browser test.
Before you start: you need access to an account (see Getting Started), and the account needs credits before it can make paid calls. You do not need a phone number yet — the first test happens in your browser.
1. Pick one workflow
Decide what this agent is for before you create it. One agent, one job. Good first choices are the workflows with a clear, checkable outcome:
- qualifying a new enquiry,
- booking or confirming an appointment,
- answering a common support question,
- a payment or renewal reminder.
Avoid making your first agent the hardest conversation in your business. You want a workflow where you can tell immediately whether the call went well.
2. Create the agent
In the dashboard, create a new agent and give it a name you will recognise later. Use the workflow name, not a person's name — Site visit qualification beats Riya.
3. Set language and voice
Choose the language the agent will speak and the voice it will use.
CallMangal agents support 32+ languages, including 12 Indian languages such as Hindi, Gujarati and Marathi, and natural mixed Hindi–English speech. Choose the language(s) your customers actually use — not the one that sounds most impressive in a demo.
Test the voice early
Voice is the single thing customers react to most. Listen to it in a browser test before you build the whole conversation, so you are not rebuilding a workflow around a voice you end up changing.
4. Add your business context
The business context (also called the global node) is where the agent learns about your business. Fill it in once and every step of the conversation can draw on it.
Include the things a new employee would need on their first day:
- what your business does, in the words you would use to a customer,
- the services or products the agent may talk about,
- the tone and persona it should adopt,
- facts it is allowed to state — locations, hours, what happens next.
Keep it factual. Anything vague here shows up as vagueness on the call.
→ Detail: Business Context
5. Set guardrails
Guardrails are the explicit rules the agent must follow on every call — kept separate from the conversation itself so they stay readable and reviewable.
Useful first guardrails:
- what the agent must never promise (discounts, delivery dates, legal or medical advice),
- topics it should decline and hand over instead,
- when it must stop and transfer to a person,
- what it should say if it does not know something.
Write these as plain instructions. They are constraints, not conversation.
6. Build a short conversation
Open the agent's workflow in the Visual Builder and lay out the smallest useful version:
- Start Call — how the conversation opens.
- A few conversation nodes — the greeting and the two or three questions that matter.
- End Call — how it closes and what outcome is recorded.
Resist adding branches on the first pass. Get one straight path working, then add the "what if the caller says something else" paths.
7. Decide what you want back
Add extraction variables for the fields you actually need from the call — for example the caller's requirement, budget, preferred callback time, and the outcome.
Name them before you test. If you do not define them, you get a transcript someone has to read; if you do, you get data you can act on.
8. Set the handoff rule
Decide, now, when a person should take over — and configure it. Common triggers are a direct request for a human, a complaint, or anything outside the guardrails.
Caller details and conversation context carry across the handoff, so your team does not start the conversation from nothing.
9. Test it
Do not connect a phone number yet.
Listen for the things that only show up out loud: an awkward greeting, a question the agent asks twice, a place where it talks over the answer, a guardrail that never fires. Fix, and test again.
Checklist before a real call
- [ ] The agent is named for its workflow
- [ ] Language and voice match your customers
- [ ] Business context is filled in and factual
- [ ] Guardrails cover what it must never do
- [ ] The conversation has a clear open, middle and end
- [ ] Extraction variables are defined
- [ ] A handoff rule exists and you have tested it firing
- [ ] The account has credits
- [ ] You have the right to call the numbers you are about to dial