In 2026, an AI voice agent in India costs roughly 3 to 6 rupees a connected minute on a mid-market platform, 2 to 5 rupees a minute if you build and run it yourself, and 8 to 30 lakh to build one. Vendors also sell it per call, at 4 to 15 rupees a connected call, or per outcome, at 8 to 25 rupees for an appointment confirmed or an instalment collected. Published rates across the Indian market run from 2 rupees a minute to about 20.
The spread is not the interesting part. The interesting part is that the advertised number and the number on your invoice are rarely the same, because a voice agent runs four meters at once and most quotes show you one of them. This guide opens all four, in rupees, and then works out what a real month costs.
Two things this guide is not. It is not about clinics specifically, which have their own arithmetic, their own platforms and their own reason to buy rather than build; that is in our AI receptionist for clinics guide. And it is not about text agents on chat or WhatsApp, where the economics are completely different and Meta's fees dominate; that is in our WhatsApp AI agent cost guide.
The quick-reference table
| How you buy it | What you pay | Fits when |
|---|---|---|
| Per minute, platform | 3 to 6 rupees a connected minute in the mid market, 2 at the aggressive end, 12 and up for enterprise Indian-language vendors | Standard flows, volumes under a few lakh minutes a month, and you want it live in days |
| Per call, platform | 4 to 15 rupees a connected and dispositioned call | Short calls with a clear outcome, such as reminders and confirmations |
| Per outcome | 8 to 25 rupees for a successful outcome: an instalment paid, an appointment confirmed, a lead qualified | Collections and sales, where the vendor is willing to carry some of the risk |
| Per agent seat, with bundled minutes | 1,400 to 2,500 rupees a seat a month, plus overage | A contact centre replacing part of a human queue rather than a flow |
| Managed service | 15,000 to 35,000 rupees a month for a small deployment, voice plus a human safety net | You want the outcome and do not want to run software |
| Custom build | 8 to 30 lakh once, then 2 to 5 rupees a minute of usage and 15 to 25 percent of the build a year in maintenance | Your own systems, your own languages, or volume where the per-minute margin funds a build every year |
The build range sits on top of the ranges in our AI agent development cost guide, because a voice agent is an agent with a telephone attached. An answering voice agent, one that knows your catalogue and your hours and never changes anything, is 8 to 14 lakh. One that books, reschedules, collects or updates a record in your systems is 14 to 30 lakh. The difference is the same difference as in text: acting costs more than answering, because acting needs limits, logs and a hand-off.
The four meters, and why a five rupee quote is not five rupees
A text agent bills when someone sends a message. A voice agent bills continuously while the line is open, on four separate meters, and the caller controls the clock. Every second of silence while they look for their order number is a second all four are running.
Here is the arithmetic in rupees for one connected minute, built from the current published rates of the mainstream providers and converted at 84 rupees to the dollar.
| Meter | What it costs a minute | Notes |
|---|---|---|
| Speech to text | 0.60 to 1.20 rupees | About a cent a minute for English on the usual providers. Indian languages and code-switching cost more, and accuracy matters more than price here |
| The model | 0.50 to 1.50 rupees | A short turn-based conversation is a small number of tokens. A model that has to re-read a long instruction on every turn is not |
| Text to speech | 0.80 to 3.00 rupees | The widest line by far. A mainstream Indian or cloud voice sits at the bottom, a premium US voice provider at the top, and it is the line people spend on first |
| Telephony | 0.80 to 1.60 rupees | Domestic Indian calling in 2026, varying with landline against mobile, direction and volume |
| Platform fee, if you use one | 1 to 4 rupees | What the platform adds for the dashboard, the flow builder, the recordings and the support |
| Total, running it yourself | 2 to 5 rupees | Lower at volume with a mainstream voice, higher with a premium voice |
| Total, on a platform | 3 to 6 rupees mid-market | The same four meters plus the fee, which is what you are actually comparing |
This is why the global platforms advertise five cents a minute and bills arrive at seventeen to twenty-seven. The advertised figure is usually the platform fee alone, with speech, the model, the voice and the telephone billed through at cost on top. Neither number is dishonest. They answer different questions, and the only one that matters is the fourth row plus the fifth.
What to ask, in one sentence: is your quoted rate all four meters and the telephone, or is it your fee with the rest billed at cost?
A real month: 20,000 connected minutes
Take a lender's collections and reminders desk: 12,000 calls a month, about 20,000 connected minutes after the ones nobody answers, in Hindi and English with some code-switching, confirming instalment dates and handing anything disputed to a person.
| Route | Monthly cost | What it includes |
|---|---|---|
| Mid-market platform at 4 rupees a minute | About 80,000 rupees | Everything, including telephony and support, on standard flows |
| Enterprise Indian-language vendor at 12 rupees | About 2.4 lakh | Deeper language coverage, dialects, mid-sentence code-switching, an account team |
| Outcome pricing at 15 rupees a confirmation, 6,000 confirmations | About 90,000 rupees | You pay for results, and the vendor carries the risk of a bad conversation |
| Custom build, usage only, at 3 rupees | About 60,000 rupees | The four meters, nothing else |
| Custom build, all in | 85,000 to 1.1 lakh | Usage plus maintenance at 15 to 25 percent of an 18 lakh build a year |
| Eight human callers | 1.8 to 2.6 lakh | Salaries at 22,000 to 32,000 a month, plus supervision, attrition and a floor to sit on |
At this volume the platform and the build are close, and the build is behind once you count the 18 lakh you spent on it. That stays true until one of three things changes: volume rises past roughly 50,000 minutes a month, the flows need your own systems in ways a platform cannot reach, or the language coverage you need is not for sale. Below about 5,000 minutes a month, do not build at all. The maintenance alone will cost more than the platform.
The number to manage, whichever route you take, is cost per completed outcome, not cost per minute. An agent at 3 rupees a minute that finishes a third of its calls is worse value than one at 5 rupees that finishes two thirds, and the cheaper voice is usually the reason the conversation failed.
Where the build money goes: a worked estimate
Tier tables hide the arithmetic, so here is the build behind that collections agent, the way we would break it down after discovery.
| Work package | Person-days |
|---|---|
| Discovery: listening to recorded calls, mapping the flows, agreeing what a finished call means | 6 |
| Collecting and labelling 300 real calls, and building the scorer that says whether a call went well | 9 |
| Telephony: numbers, inbound and outbound routing, transfer to a person, recording and consent | 8 |
| The conversation: prompts, turn taking, interruption handling, repair when the caller is misheard | 14 |
| Languages: Hindi and English with code-switching, tested on real recordings rather than clean audio | 10 |
| Latency work: streaming speech in and out, cutting the gap to where the caller stops noticing | 10 |
| Systems: reading the loan record, writing the promise to pay, raising the dispute | 12 |
| Guardrails: what the agent may commit to, what it must never say, audit log of every call | 8 |
| Hand-off to a person, with the call summary and the caller's intent attached | 6 |
| Compliance: consent capture, recording retention, do-not-disturb rules and calling windows | 6 |
| Testing and a staged rollout, after-hours first, then overflow, then the main queue | 10 |
| A month of tuning on real calls | 12 |
| Total | 111 |
At a blended 16,000 rupees a person-day, that is about 18 lakh, and with two engineers it runs ten to twelve weeks, which matches the 6 to 10 weeks we quote for voice projects plus the tuning month. Two lines deserve attention. Latency and languages are 20 of 111 days between them, and they are the two lines a demo never shows. The conversation itself, at 14 days, is more than the systems work, which is the opposite of a text agent and the thing that surprises people who have built one.
Latency: the number that decides whether it works
One of the three questions Google is already showing our voice page for is which agent has the lowest latency for phone calls in India. It is the right question, and it has an unglamorous answer: the target is under 800 milliseconds from the caller finishing a sentence to the first sound of the reply, and about 1.2 seconds is where people start talking over the agent.
That budget is spent in four places: detecting that the caller has actually stopped, transcribing, thinking, and starting to speak. Streaming all four, rather than waiting for each to finish, is most of the engineering. Hosting in an Indian region rather than a US one is worth 100 to 200 milliseconds on its own, for free, and is the single cheapest latency improvement available to an Indian deployment.
The cost link is direct: a faster stack usually means a smaller model and a mainstream voice, which is also the cheaper stack. Latency and price pull in the same direction here, which is rare and worth using.
Indian languages, and the details that break first
The second question the page is being shown for is which voice agents pronounce Indian rupee amounts, dates and phone numbers correctly. In our experience that is the right thing to test first, because it is where every vendor demo quietly fails.
Ask for a recording of the agent saying these, in Hindi and in English, before you sign anything.
- An amount: 1,24,500 rupees. Indian grouping, and "one lakh twenty four thousand five hundred", not "one hundred twenty four thousand".
- A date: 3 October, and whether 03/10 comes out as October the third rather than March the tenth.
- A phone number: ten digits, grouped the way a person would say them, not read as one enormous number.
- A reference: an order or loan number with letters and digits mixed, spoken slowly enough to write down.
- A name: three names your customers actually have, pronounced the way they pronounce them.
- Code-switching: one sentence that starts in Hindi and ends in English, which is how most of India actually speaks.
None of this is exotic, and all of it is the difference between a call that completes and a call where the customer says "agent" four times. Budget test cases for each, and put them in the scorer.
The seven things that move the price
- Answering or acting. Reading out hours, prices and order status is the cheap row. Booking, rescheduling, collecting or refunding needs limits, an audit log and a hand-off, and prices like an acting agent.
- Minutes, and who controls them. Inbound minutes are set by your customers. Outbound minutes are set by you, which makes outbound the easier bill to forecast and the easier one to waste. Pay for connected minutes only, and check how the vendor counts a call nobody answered.
- Languages and dialects. English and Hindi are table stakes. Each additional language is a week or more of evaluation, not a configuration switch, and dialect coverage is what enterprise Indian vendors charge the premium for.
- The voice. A premium voice provider can be three times the cost of a mainstream one on the text-to-speech line, which is the widest line in the table. It is worth paying for on a brand line and rarely worth it on a reminder.
- How many systems it touches. A loan record, a calendar, a payment link, a CRM. Each is a connector, a permission model and a failure case, exactly as in text.
- Compliance. Recording consent, retention, do-not-disturb registers and legal calling windows. Financial services and health add review by someone outside engineering, and it is not optional.
- Who carries a bad call. A per-minute contract bills you for the agent's mistakes. An outcome contract does not. That difference is worth more than a rupee a minute on a flow you have not proven yet.
Buy, configure or build
| Option | What you pay | Fits when |
|---|---|---|
| An Indian platform with a flow builder | 3 to 6 rupees a minute, live in days | Standard inbound answering, reminders, confirmations. Most first deployments belong here |
| An enterprise Indian-language vendor | 12 rupees a minute and up, usually on a custom quote | Deep dialect coverage, regulated industries, an account team and an SLA |
| A global developer platform | Component costs plus a fee, roughly 4 to 8 rupees a minute in India once telephony is added | You want control of the stack without owning the plumbing, and your flows are unusual |
| A custom build | 8 to 30 lakh, then 2 to 5 rupees a minute | Your own systems, unusual flows, language needs nobody sells, or volume past about 50,000 minutes a month |
We build the last row and will say plainly when the first row is the right purchase, which for a first voice deployment it usually is. Run a platform for two months, learn what your callers actually ask, and then decide whether a build is worth it with evidence rather than a slide. Our voice AI page describes how we run the build when it is.
The costs that are not in the quote
| Cost | Typical size | When it surprises you |
|---|---|---|
| Numbers and KYC | Small, but weeks of waiting | When the build is finished and the number is not cleared for outbound at volume |
| Recording storage | A few thousand rupees a month, more with a long retention rule | When compliance asks for two years of calls and nobody priced the storage |
| Failed and unanswered calls | Charged by some vendors, not by others | On the first invoice, when the call count does not match your own |
| Transfers to a person | A second leg, billed separately | When a quarter of calls transfer and the telephony line doubles |
| The humans behind the hand-off | Training, and a harder case mix | Month two, when the easy calls are gone |
| Re-tuning after a model change | 2 to 5 days, once or twice a year | When the provider retires the model and the timing of the conversation changes |
| Keeping the scorer alive | A few days a month | When a new kind of call appears and nobody added examples |
GST at 18 percent applies to the development invoice and to most platform fees as it does to any software service in India, claimable as input credit if you are registered, but the cash goes out first.
How to read a voice agent quote
Six checks, ten minutes.
- Ask whether the rate includes all four meters and the telephone. If the answer is a fee plus pass-through, ask for a worked example of a real month.
- Ask how a minute is counted. Connected minutes, rounded how, and what happens to a call nobody answers.
- Ask for a recording of the six Indian-language tests above. Not a showreel. Your amounts, your dates, your names.
- Ask for the latency figure, measured from the caller finishing to the first sound of the reply, and where it is hosted.
- Ask what the agent may commit to without a person, and where that is written down.
- Ask who owns the recordings, the transcripts and the prompts. All three should be yours, in writing.
The vendor-selection version of this, with a scoring rubric, is in our guide to hiring an AI development company.
How to spend less without a worse agent
- Cut the silence, not the voice. Barge-in handling and a shorter opening line save more rupees per call than switching to a cheaper voice, and they improve the call.
- Use a mainstream voice for reminders and save the premium voice for the line customers hear when they are angry.
- Send a message first and call only the non-responders. A WhatsApp utility message costs a fraction of a connected minute, and the arithmetic for it is in our WhatsApp guide.
- Host in an Indian region. Cheaper, and 100 to 200 milliseconds faster.
- Route the simple turn to a small model and keep the capable one for the turn that needs judgement.
- Start with after-hours calls. You get the examples and the confidence data on the calls nobody is answering today, which is the lowest-risk volume you have.
- Prove it on a platform before you build. Two months of real calls is the cheapest specification you will ever get.
When not to use a voice agent
- Under a few hundred calls a month. Fix the routing, the hold message and the out-of-hours greeting instead.
- When the call is the product. A sales conversation that depends on rapport is not a flow, and a synthetic voice reads as a downgrade.
- When nobody can hear the recordings. If no one on your side will listen to calls weekly, quality will drift and nobody will notice for a quarter.
- When the caller is in distress. Emergencies and complaints should reach a person in one step, and the agent's only job there is to be fast about it.
Questions people ask
How much does an AI voice agent cost in India?
Three to six rupees a connected minute on a mid-market Indian platform, with the aggressive end of the market at 2 and enterprise Indian-language vendors at 12 and above. Vendors also sell per call at 4 to 15 rupees and per outcome at 8 to 25. Building and running your own comes to 2 to 5 rupees a minute, after a build of 8 to 30 lakh.
What does a voice agent cost per minute to run, in detail?
Speech to text is 0.60 to 1.20 rupees a minute, the model 0.50 to 1.50, text to speech 0.80 to 3.00, and Indian domestic telephony 0.80 to 1.60. That is 2 to 5 rupees before any platform fee, and 3 to 6 with a mid-market fee on top. Text to speech is the widest line and the one to look at first if the bill surprises you.
Why do global platforms advertise five cents a minute?
Because that is usually the platform fee alone, with speech, the model, the voice and the telephone billed through at cost on top. A realistic all-in figure on those platforms is 17 to 27 cents a minute, which is 14 to 23 rupees before Indian telephony is swapped in for American. Ask any quote which of the two numbers it is.
How much does it cost to build a voice agent?
Eight to 14 lakh for an agent that answers questions and changes nothing, 14 to 30 lakh for one that books, collects or updates records in your systems. The estimate above, for a two-language collections agent with telephony, guardrails and compliance, comes to 111 person-days and about 18 lakh. Add 15 to 25 percent of the build a year for maintenance.
Is a voice agent cheaper than hiring callers?
At volume, by a wide margin. Twenty thousand connected minutes a month costs 60,000 to 90,000 rupees on most routes, against 1.8 to 2.6 lakh for the eight people it would take, once supervision and attrition are counted. Below a few thousand minutes the fixed costs dominate and people are cheaper. Compare cost per completed outcome, not cost per minute.
Can it handle Hindi and other Indian languages?
Yes, and the honest answer is that it varies more than any vendor will admit. English and Hindi are well served. Regional languages and mid-sentence code-switching are where the differences show, and they are what enterprise Indian vendors charge a premium for. Test with recordings of your own customers, and test the amounts, dates and phone numbers before anything else.
What latency should I accept?
Under 800 milliseconds from the caller finishing a sentence to the first sound of the reply. At about 1.2 seconds people begin talking over the agent, and above that the call feels broken however good the voice is. Ask where the stack is hosted, because an Indian region is worth 100 to 200 milliseconds against a US one.
Do we need consent to record calls?
Yes, and it is easier than it sounds: capture it in the call itself, in the caller's language, in the first few seconds, and keep the recording of the consent with the recording of the call. Then agree a retention period and make sure deletion actually works before you launch rather than after someone asks.
Getting a real number
Ranges end where your actual calls begin. We price voice projects after a one-week discovery and a two-week prototype run on recordings of your real calls, so the number you approve is based on how the agent handles your callers and your languages, not on how it demos in ours. What that looks like is on our voice AI page.