Skip to content

Bananaflow MCP is live.

Read the docs

How to Build an AI Calling Agent in India

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Updated
Cover for "How to Build an AI Calling Agent in India": a white line diagram of an AI calling agent's three jobs on a phone line: from a phone panel labelled Caller, arrows run through a wave panel labelled Hears, a spark panel labelled Thinks and a speaker panel labelled Speaks, and a dashed arc carries the reply back over the top to the caller, drawn over a soft, out-of-focus water lily in sage mint, its pointed petals blurred into veils, with an aqua haze at one end and a peach glow at the other.

An AI calling agent is software that holds a phone conversation: it hears the person, decides what to say, speaks, and acts on what it learned. If you want to make an AI calling agent for your business, the work is less about the AI and more about the plumbing around it: one clear call, a phone number, your data, the rules, and a careful first week.

Picture three businesses. A lender in Pune wants to remind borrowers about an EMI due on the 5th. A clinic chain in Chennai wants to confirm tomorrow's appointments. A D2C brand in Surat wants to call every website lead quickly. Each needs the same nine steps, and each will trip on a different one. The names are made up, the pattern is not.

This guide walks through the steps in order. For each one we say what it is, what it needs and the mistake we see most. Then we compare building it yourself with using a platform, and we say plainly when building is the right call.

What is an AI calling agent?

An AI calling agent, also called an AI voice agent or voicebot, is a program that makes or answers phone calls and talks like a person for one defined job. It is not a recording and it is not a phone menu. It listens, works out what the caller means, and replies in its own words. Our post on what conversational AI is goes deeper, and the glossary has a short entry on the voicebot.

Under the hood, most voice agents chain three parts:

  • Speech to text. The ears. It turns the caller's audio into words as they speak.
  • A language model. The brain. It reads the words, follows your instructions, decides the next move and may call a tool, such as a calendar lookup.
  • A voice. The mouth. It turns the reply back into speech.

Some systems fold the three into one speech-to-speech model. The chained version is the one most teams build and tune, because you can see and change each part.

A fourth piece is the phone line. Calls reach the agent through a carrier. The standard behind most business phone links is SIP, the protocol for "creating, modifying, and terminating sessions", which include internet telephone calls, in the words of RFC 3261. A carrier connects its network to your agent over a SIP trunk.

The steps to build an AI calling agent, in order: pick one call, write the prompt, choose voice and language, connect data, get a number, clear the rules, test, launch small and measure.

What do you need before you start?

Before you build, you need one call type, someone who owns the script, the data that call needs, a phone number, the legal groundwork and a person to review calls. The phone number and the registrations are the two items you cannot speed up yourself, so start them first. Do not plan a launch date until both are in hand.

  • Item: One call type and one outcome

    Who owns it
    Business owner or operations head
    Kind of cost
    None
  • Item: A script and a prompt owner

    Who owns it
    Someone who knows how the best human caller handles the call
    Kind of cost
    Staff time
  • Item: The data the call needs (CRM, calendar, order system)

    Who owns it
    Whoever runs that system
    Kind of cost
    Developer time
  • Item: A phone number and carrier account

    Who owns it
    Operations or IT
    Kind of cost
    Rental and per-minute charges
  • Item: DLT registration and a consent record

    Who owns it
    Compliance or the business owner
    Kind of cost
    Set by the operator
  • Item: A reviewer for calls

    Who owns it
    A team lead
    Kind of cost
    Staff time
  • Item: The voice stack: speech, model, voice

    Who owns it
    You (if building) or the platform
    Kind of cost
    Per-minute usage

How do you build an AI calling agent, step by step?

You build an AI calling agent in nine steps: pick one call, write the prompt, choose voice and languages, connect your data, get a phone number, clear the rules, test, launch small and measure. Each step below says what it is, what it needs and the common mistake.

Step 1: Pick one call and one outcome

What it is. Choose a single call with a result you can count. "Confirm tomorrow's appointment" is a call. "Handle customer service" is a department. For the Chennai clinic chain, the call is "confirm, reschedule or cancel tomorrow's slot." For the Pune lender it is "tell the borrower the EMI is due on the 5th and learn whether they will pay." For the Surat brand it is "call a new website lead and book a demo."

What it needs.

  • A direction. Outbound calls go to people you chose, inbound calls come to you, and they are built differently. See inbound versus outbound calls.
  • A short list of outcomes, decided now: confirmed, rescheduled, cancelled, call back later, wrong number, no answer, transferred. Every report you read later depends on them.

Common mistake. Giving one agent too many jobs. A call that confirms, upsells, collects feedback and answers any question is hard to write and harder to test. Start narrow. You can add a second agent for a second call. Draw the call first: call flow design.

Step 2: Write the script and the prompt

What it is. The prompt is the written brief the language model follows: who the agent is, what it may say, what it must never do, and how each outcome ends. It is where most of the quality comes from.

What it needs. A script owner who has heard good calls, and time to rewrite. We have a full method in our voice agent prompt guide: the sections of a system prompt, how to write rupee amounts and dates so they are spoken right, tools, guardrails and testing. We do not repeat it here.

Common mistake. Writing the prompt like an email. A phone caller cannot scroll back. Short sentences, one question at a time, and the key fact first.

If you would rather not start from a blank page, Bananaflow can draft the call from a description of the use case. The workflow builder can write the first version of the workflow and the tools it needs, and you edit from there. An AI assistant such as Claude Code or Cursor can also create and test agents through the Bananaflow MCP server.

Step 3: Choose the voice and the languages

What it is. Decide which languages your callers speak, pick a voice that sounds right in them, and choose the speech-to-text that hears them well.

What it needs.

  • The languages your list really uses. Pune may need Marathi, Hindi and English, Chennai needs Tamil, and Surat may need Gujarati.
  • A plan for mixed speech. Many callers switch between Hindi and English inside one sentence. Our post on Hinglish voice agents explains why that is hard and what to test.
  • A test with the details that break voices: names, rupee amounts, dates, addresses and pin codes.

Common mistake. Choosing a voice from a demo clip. A voice that sounds warm reading a sentence about the weather can stumble on "Rs 14,250 due on 5 December" or on a Tamil surname. Test with your own data.

On a platform, check whether you can change one part, say the voice, without rebuilding the rest. On Bananaflow, each agent picks its own speech-to-text, language model and voice and can change any one of them without a rebuild. It supports English plus 10 Indian languages (Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu), and the agent can follow the caller between them on one call. See the languages page. A word list for the speech-to-text helps with brand names and jargon.

Step 4: Connect the data and tools

What it is. A call that only talks is a demo. A useful call looks something up or writes something back: today's open slots, an order's status, a lead's name, a note in the CRM.

What it needs.

  • A list of what the agent must read and write mid-call. For the clinic, free slots and a way to book one. For the lender, the due amount and date.
  • An API from your own system for each, with a test you can run before the agent uses it.
  • A knowledge base for answers that should not come from the model's memory, such as your refund policy or clinic timings.
  • A hand-off for what the agent must not handle (a complaint, a dispute, a medical question): a person, with their number ready.

Common mistake. Letting the model guess. If the agent is not given the due amount, it may invent one. Pass the real data into the call, and instruct the agent to say it cannot help rather than guess.

Bananaflow connects agents to your own APIs as tools, to the customer's own MCP servers and to more than 100 apps by one-click sign-in. It has a knowledge base for PDF, DOCX, TXT and Markdown files up to 5 MB each. It can transfer a call to another agent or, on Twilio, Plivo, Telnyx, Cloudonix and Asterisk only, to a person. See the integrations page for the list.

Step 5: Get a phone number and connect it

What it is. The agent needs a real line. Outbound calls leave from a number your carrier gives you, and inbound calls arrive on one. The carrier links to your agent through a SIP trunk or the carrier's own integration.

What it needs.

  • A carrier account that supports your use. Check the number series, because promotional and service calls are now separated by series (see step 6).
  • The connection details for your agent, and a test call from and to a real mobile phone.
  • A caller ID that does not get labelled as spam. See why business calls get marked as spam.

For a worked example with one carrier, read our guide to running an AI voice agent on an Exotel number.

Common mistake. Treating the number as a last-minute task. Ask your carrier early which documents it needs, because onboarding can take time. Start it in week one, in parallel with the prompt.

Bananaflow works with Twilio, Plivo, Exotel, Vobiz, Telnyx, Vonage, Cloudonix, Asterisk and VICIdial, for inbound and outbound. You bring your own numbers, and the carrier bills its own call charges. The telephony page has the detail.

Step 6: Clear the rules for India

What it is. An AI agent is still a business making a phone call. The rules below apply to the business and to its carrier. This guide summarises them. It is not legal advice, and it does not replace the deep pages we link.

These duties are yours and your carrier's. TRAI, DLT, DND and DPDP compliance needs your own checks, whatever software you use.

  • Rule: TRAI's TCCCPR

    What it asks
    The telecom regulator's rules for commercial calls and messages. First notified on 19 July 2018 and amended since, including in February 2025
    Read more
    [TRAI rules for business calls](/blog/trai-new-rules-for-business-calls/)
  • Rule: 140 and 1600 number series

    What it asks
    Since the February 2025 amendment, promotional calls go out on the 140 series and service or transactional calls on 1600, not on ordinary ten-digit numbers
    Read more
    [140, 160 and 1600 series explained](/blog/140-160-1600-series-numbers-india/)
  • Rule: DLT registration

    What it asks
    A business that makes commercial calls registers as a Principal Entity on an operator's DLT portal. TRAI's sender hub lists entity, header, consent and preference registration
    Read more
    [DLT registration for voice calls](/blog/dlt-registration-for-voice-calls/)
  • Rule: DND and the preference register

    What it asks
    Customers can block promotional calls by category and time band, and you must not call numbers that have opted out
    Read more
    [DND](/glossary/dnd/) and [TRAI rules for AI calls](/glossary/trai-rules-for-ai-calls/)
  • Rule: DPDP Act 2023 and Rules 2025

    What it asks
    Notice and consent for personal data, which includes call recordings and transcripts
    Read more
    [DPDP rules for call centres](/blog/dpdp-rules-for-call-centers/)
  • Rule: Call recording

    What it asks
    When recording is allowed and what to tell the person
    Read more
    [Is call recording legal in India?](/blog/is-call-recording-legal-in-india/)

TRAI's own sender registration hub is the place to see the four registration types, and its February 2025 press release is the source for the number series change. The DPDP Act text is on the MeitY site.

One more point. We have not found an Indian rule that forces an AI agent to announce that it is an AI. We still suggest it. A caller who feels tricked hangs up and tells others, and the opening line is cheap to write.

What it needs. A named person in the business who owns compliance, a carrier who knows the rules for your call type, a consent record for the numbers you call, and a way to remove people who ask to stop.

Common mistake. Assuming the AI vendor or the carrier has "handled it." Ask your carrier, in writing, what they check on their side and what is left to you.

Step 7: Test before real calls

What it is. A test is a call you make yourself, on purpose, to see where the agent breaks.

What it needs. A list of cases and the patience to read the transcripts, not only the summary. At minimum:

  • The happy path, ending in the right outcome.
  • A wrong number. Does it apologise, end the call and mark it correctly?
  • An angry caller. Does it stay calm and offer a person?
  • Silence and background noise, such as a road or a market.
  • A language switch. Start in English, answer in Hindi.
  • A voicemail. Does it talk to a machine? See answering machine detection.

Our prompt guide covers how to test and improve a voice agent prompt in detail. For the full test plan, with a case library, run counts and a scoring rubric, see how to test a voice agent.

Common mistake. Testing only in a quiet room, on your own voice. Real callers have accents, children in the background and bad signal. Ask five colleagues who did not write the prompt to call it.

You can test a Bananaflow agent in the browser, by voice or by text chat, before it takes real calls. It also detects voicemail and IVR menus before it speaks.

Step 8: Launch small

What it is. Start with a slice of your list, not the whole thing. Run one campaign, for a few days, with a person watching.

What it needs.

  • A small first batch, such as fifty calls, read the same day.
  • Calling hours that respect your callers. Nobody wants an EMI reminder at 10pm.
  • Retry rules: how many tries, how far apart, and when to stop. See what an auto-dialer does.
  • A person on hand to take transferred calls.

Common mistake. Launching to the full list on day one. If the prompt mispronounces a word or the data feed has a bad column, ten thousand people hear it before you notice.

Bananaflow campaigns run from a CSV with retries, calling time slots and concurrency limits, plus an A/B split between two agent versions. See the campaigns page.

Step 9: Measure and improve

What it is. A live agent is never finished. Each week you read a sample of calls, fix what went wrong and change the prompt or the data.

What it needs.

  • A few numbers: the share of calls reaching each outcome, the transfer rate, average call length and complaints or opt-outs. See call centre metrics.
  • A habit of listening to a sample of recordings each week. Numbers say something is off, and the recording says why.
  • A way to score calls without hearing every one, such as a QA step that tags each call and writes a summary.

Common mistake. Watching only the pick-up rate. A call that was answered and ended badly counts as a pick-up. Watch the outcome.

After each call, Bananaflow stores a transcript, a recording, minutes used and any variables the agent extracted. A QA node scores the call and returns tags, sentiment and a summary. Call dispositions, reports with CSV export, webhooks and BigQuery export are covered on the call QA and analytics pages. Recordings are stored in India, in Mumbai.

Should you build it yourself or use a platform?

For most businesses with a standard call, use a platform. Build it yourself only if you have voice engineers and a flow no platform fits. The honest answer depends on three questions: how standard is your call, how many engineers can you spare, and how soon do you need it live? If you build, see Pipecat for voice agents.

If your call looks like appointment confirmation, lead follow-up or payment reminders, a platform's ready pieces will fit. If your call is unusual, or you must run your own models and servers, building starts to pay off. If you would rather describe the call in plain words, you can build one from Claude Code on Bananaflow.

What you own when you build an AI voice calling agent yourself and what you own when you use a platform. Building keeps every part in your hands. On a platform, the speech, the voice and the hosting are run for you, and the call design, your data, the number and the rules stay yours.
  • : Time to a first working agent

    Build it yourself
    Longer: you join, host and test every part yourself
    Use a platform
    Shorter, as the parts are already joined. A production call still needs a number, the rules and testing
  • : Control

    Build it yourself
    Full. You pick every part and own the code
    Use a platform
    Less. You work within what the platform offers
  • : Engineers

    Build it yourself
    Needed to join, host and monitor the parts
    Use a platform
    Few or none for a standard call
  • : Bills

    Build it yourself
    Separate ones for speech, model, voice and carrier. Several are often in US dollars, so exchange rates move your cost
    Use a platform
    One per-minute bill for the voice stack, plus your carrier's charges
  • : Uptime and fixes

    Build it yourself
    Yours. If the voice or a provider fails at 11am, you fix it
    Use a platform
    Shared. Ask what the platform guarantees and what it does not
  • : Indian languages and accents

    Build it yourself
    You tune and test them yourself
    Use a platform
    Check which languages are supported and what has been tested
  • : Compliance

    Build it yourself
    Yours and your carrier's
    Use a platform
    Still yours and your carrier's. A platform is not a compliance service
  • : Best for

    Build it yourself
    A team with voice engineers and an unusual flow
    Use a platform
    A business with a standard call that wants it live soon

The real cost of building is people-time: engineers who join the parts, tune them for Indian accents and Hinglish, and keep them running.

Before you choose any platform, check which Indian languages it supports and whether it can switch mid-call, which carriers it connects to, whether you can swap the voice without a rebuild, what it does not do, and what your carrier will charge on top.

If you are comparing platforms, our comparison of AI voice agent platforms in India lays out the options. We do not rank them in this post.

How much does an AI calling agent cost in India?

An AI calling agent costs the voice stack per minute, plus your carrier's per-minute charge, plus the time to build and tune it. If you build, add engineer time and separate vendor bills, some in US dollars. If you use a platform, you pay one per-minute rate for the voice stack and the carrier bills its charges on top.

We do not repeat rupee figures here. Our post on telecaller versus AI calling agent cost has the numbers and the sums behind them, and the pricing page has Bananaflow's current rates. A demo is cheap, but a production line is not free.

What mistakes stop AI calling agents from working?

The mistakes that sink AI calling agents are almost all planning mistakes, not model mistakes. Here are the ones we would check first.

  1. Too many jobs in one agent. Split a broad call into separate agents, one per outcome.
  2. No human fallback. Every agent needs a way to hand a caller to a person, and a person who answers.
  3. Testing only in a quiet room. Test with noise, accents and silence.
  4. Ignoring DND and consent. One careless batch can create complaints and carrier trouble.
  5. Launching to the full list on day one. Start small and widen.
  6. Never reading transcripts. A summary hides the sentence that went wrong.
  7. Skipping the data step. An agent that cannot look things up will guess, and a guess on a due amount is a real error.
  8. Picking a voice from a demo. Test on your own names and rupee amounts.

How does Bananaflow fit?

Bananaflow is a platform for the build steps, so you do not have to join the parts yourself. You describe or draw the call, test it in the browser, connect your own carrier and numbers, run a small campaign and read the results. Step 1 and the rules in step 6 stay with you. A good place to start is the workflow builder.

What should you remember about building an AI calling agent?

An AI calling agent works when it does one call well. Pick the call, write a short prompt, test the voice on your own names and amounts, connect the data it needs, get a number early, treat the rules as your own duty, and launch to a small slice of your list. Build it yourself only if you have the engineers and an unusual flow. Otherwise a platform gets a standard call live sooner, and your team's time goes into the part that matters most: listening to the calls and making the next one better.

FAQ

Frequently asked questions

  • Can I make an AI calling agent for free?

    You can try parts of one for free, but a live agent on a real phone number costs money. Speech recognition, the language model, the voice and the carrier's call minutes are each billed, and some of those bills are in US dollars. A demo in a browser is cheap. A production line is not free. Our cost post has the rupee numbers.

  • Can an AI calling agent speak Hindi?

    Yes, and it should also handle Hinglish, the mix of Hindi and English most callers actually speak. Test the voice and the speech-to-text on names, rupee amounts and mid-sentence language switches before you commit. On Bananaflow an agent can follow the caller between English and 10 Indian languages, including Hindi, on the same call.

  • Do I need a new phone number for an AI calling agent?

    You need a number from a carrier, connected to the agent through a SIP trunk or the carrier's own integration. Which series it sits on matters: since the February 2025 amendment, TRAI points promotional calls to the 140 series and service calls to 1600. You bring your own number through your carrier.

  • How long does it take to build an AI calling agent?

    A demo is quick to get talking. A production agent takes longer, because the phone number, the compliance work, the testing and a small launch each take time, and the first two depend on approvals outside your control. We give no fixed timeline, because it varies by business.

  • Can I build an AI calling agent on my phone?

    You can test one on a phone, but the build itself needs a computer and a carrier account. You write the prompt, connect your data, link a phone number and review transcripts, and those jobs are easier on a laptop. On Bananaflow you can test an agent in the browser by voice or text before it takes real calls.

  • Should I build an AI calling agent myself or use a platform?

    Most businesses with a standard call, such as appointment confirmation or lead qualification, are better off using a platform. Build it yourself if you have voice engineers and an unusual flow that no platform fits. Building gives you full control and also gives you the uptime, the Indian-language tuning and the separate vendor bills.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot