In 2026, an AI receptionist for a clinic in India costs 800 to 5,000 rupees a month plus 3 to 9 rupees a connected minute on a platform product, 15,000 to 35,000 a month as a managed service, or 10 to 30 lakh to build your own, with usage of 2 to 4 rupees a minute on top. Those are the market's numbers. The more useful ones are further down: what a month of calls actually costs each way for a real clinic, and the volume at which building starts to beat buying, which for a single clinic it usually does not.
We build voice agents for a living, and we run one of our own, so this guide has an interest to declare. It also has a rule: most of the pages ranking for this query are written by companies selling a receptionist product, and every one of them concludes you should buy theirs. Our conclusion is that a single clinic should almost always buy, and that the cases for building are specific and worth knowing.
The quick-reference table
| Option | What you pay | Live in | Fits |
|---|---|---|---|
| Platform product (ConnectAI, Patient Square, HuskyVoice, Voiceoc and similar) | 800 to 5,000 rupees a month, plus 3 to 9 rupees a connected minute on most plans | Days | One clinic, standard booking, Hindi and English, an existing calendar or practice software the product integrates with |
| Managed service | 15,000 to 35,000 rupees a month all in, usually voice plus WhatsApp plus a human safety net | 2 to 4 weeks | A clinic that wants the outcome and does not want to run software |
| Custom build | 10 to 30 lakh once, then 2 to 4 rupees a minute of usage and maintenance | 6 to 10 weeks | Hospital groups and chains, your own hospital information system, languages the platforms serve badly, or an agent inside your own patient app |
| A second receptionist | 18,000 to 30,000 rupees a month, more in metros, plus hiring and turnover | Weeks | The honest alternative, and the right one when call volume is low |
The per-minute figures are for connected time. A clinic taking 1,200 calls a month at two and a half minutes each is buying about 3,000 minutes, which is the number to carry into the arithmetic below.
What an AI receptionist actually does at a clinic
The name undersells it and oversells it at the same time. It undersells because a good one does more than answer: it books against real slot availability, sends the WhatsApp confirmation, calls the day before, and chases the no-show. It oversells because it is not a receptionist. It cannot triage, it must not give medical advice, and it should hand an emergency to a person without asking a single question.
| It should do this | It should not do this |
|---|---|
| Pick up on the first ring at 9pm, during a rush, and on Sunday | Decide whether a symptom needs a same-day visit |
| Understand what the caller wants, in the language they use | Answer a clinical question with anything but "the doctor will call you" |
| Book, reschedule or cancel against the real calendar, with no double booking | Book an appointment for someone who says it is an emergency |
| Answer the boring questions: timings, address, fees, which doctor sits when, what to bring | Handle an angry or distressed caller past the first sentence |
| Send the WhatsApp confirmation within a minute, and the reminder the day before | Recognise nothing: your repeat patients should be greeted by name or routed to a person |
| Hand off cleanly, with the transcript, to a person or to a callback queue after hours | Make something up when it does not know; it should say so and hand off |
The most useful test of any product is the right-hand column. Call the demo number, say "I think my child has swallowed something", and see whether it tries to book a slot. That one call tells you more than the feature list.
The cost math, done properly
Take a three-doctor clinic in a tier-two city: 1,200 calls a month including reminders, about 3,000 connected minutes, and a front desk of two people who are with walk-ins for a good part of every hour.
| Option | Monthly cost | What is included |
|---|---|---|
| Platform product | 10,000 to 32,000 rupees | 3,000 minutes at 3 to 9 rupees, plus the subscription |
| Managed service | 15,000 to 35,000 rupees | Voice, WhatsApp and a human safety net, run for you |
| Custom build, usage only | About 7,500 rupees | 3,000 minutes at roughly 2.5 rupees of speech, model, voice and telephony |
| Custom build, all in | 25,000 to 40,000 rupees | Usage plus maintenance at 15 to 25 percent of a 15 lakh build per year |
| Second-shift receptionist | 22,000 to 30,000 rupees | A person from 7pm to 11pm, if you can hire and keep one |
Two things follow from the table. For one clinic, a custom build is the most expensive option on a monthly basis before you count the 15 lakh you spent building it, because maintenance on a custom voice agent costs more than the platform's whole subscription. And the comparison everyone makes, agent against receptionist, is the wrong one. The agent is not cheaper than your receptionist; it is cheaper than the second receptionist you never hired, the one who would have taken the 8pm calls that currently go to voicemail and then to the clinic down the road.
Where building wins is scale. A hospital group with twenty locations is buying 60,000 minutes a month, which at platform rates is 1.8 to 5.4 lakh, against usage of about 1.5 lakh plus a maintenance line that does not grow with locations. Somewhere in the low thousands of calls a month the arithmetic reverses, and above that the platform's margin is paying for a build every year. The method for usage costs, stage by stage, is on our voice AI page.
The number to manage, whichever option you pick, is cost per booked appointment: monthly cost divided by appointments the agent booked or saved that a person would not have. Managed services in India report 120 to 280 rupees per appointment recovered. Track your own, because it is the only figure that tells you whether the thing is working.
The five things that decide whether it works
- Latency and interruption. Callers forgive a wrong answer faster than a two-second silence. A silence sounds like a machine, and once a caller decides they are talking to one they stop cooperating. A good agent starts replying in well under a second and stops talking the instant the caller talks over it. Running Crodo, our own voice assistant, taught us that this plumbing, not the model, is where voice products are won; the honest version is in four products in, what we got wrong.
- Hinglish, not Hindi. Products demo clean Hindi and clean English. Patients speak a sentence that starts in Hindi, names the doctor in English and ends in Marathi. Provider quality on that mixture varies far more than the marketing suggests, so test on recordings of your own callers, elderly ones included, before you sign anything.
- Real slot availability. An agent that cannot read and write your actual calendar, whether that is practice software, a hospital information system or a shared Google Calendar, is an IVR with a nicer voice. Ask which systems the product writes to, not which ones it "supports".
- The hand-off. Emergencies, clinical questions, distressed callers and your best patients all need a person, and after hours there may not be one. The design question is what happens then: a callback queue with a promised time, a transfer to an on-call number, or a dropped patient. Only the last one is unacceptable, and it is the default on a surprising number of products.
- Consent, recording and where the data lives. Recording a call needs consent in the conversation, in the caller's language. India's Digital Personal Data Protection Act treats patient details as personal data with notice, consent, purpose and deletion obligations, and its core rules are scheduled to come into force through 2027, so build to them now rather than retrofitting. Ask where recordings and transcripts are stored, for how long, and who inside the clinic can listen to them.
Build or buy: the decision table
| Your situation | Buy or build | Why |
|---|---|---|
| One clinic, standard booking, Hindi and English | Buy | The platform's per-minute margin is smaller than a build's maintenance |
| Two to five clinics on the same practice software | Buy, or a managed service | Still under the volume where a build pays |
| A hospital group, or a chain past a few thousand calls a month | Build, or an enterprise vendor on your terms | The margin is now paying for a build every year, and integration is the hard part either way |
| Your own hospital information system with no product integration | Build the connector at least | Without slot availability, nothing else matters |
| Languages or dialects the platforms handle badly on your recordings | Test first, probably build | Provider choice per language is the fix, and platforms rarely let you choose |
| An agent inside your own patient app | Build | That is AI integration, not a receptionist product |
| Under a few hundred calls a month, all answered today | Neither | Fix the out-of-hours message and call routing instead |
Whichever column you land in, keep your phone number. Products that require you to port the number to their platform are building a switching cost. The agent should sit on top of the number you already print on the board outside.
What a custom AI receptionist costs to build
A clinic receptionist is an acting agent with a voice front end: it books, changes and cancels things in your systems, so it carries all the guardrail work of any acting agent, plus telephony, speech and latency engineering on top. That puts a custom build at 10 to 30 lakh in the Indian market. Here is how the middle of that range breaks down for a hospital group's receptionist across locations, the way we would estimate it after discovery.
| Work package | Person-days |
|---|---|
| Discovery: the calls that matter, recordings of real ones, the hand-off rules | 5 |
| Speech provider bake-off per language, scored on your recordings | 6 |
| Prototype on a test number, scored on accuracy, latency and completion | 10 |
| Telephony, number set-up, interruption handling and the latency budget | 10 |
| Booking actions: slot rules, doctor schedules, read and write to the hospital system | 15 |
| WhatsApp confirmations, reminders and no-show follow-up, with template approvals | 8 |
| Escalation: emergency detection, clinical questions, after-hours callback path | 8 |
| Consent, recording storage, retention, deletion and the audit log | 6 |
| Dashboards and the morning summary for each front desk | 6 |
| Testing on noisy recordings, adversarial calls, load | 8 |
| Staged rollout, after-hours first, then overflow, then everything, and a month of tuning | 10 |
| Total | 92 |
At a blended 16,000 rupees a person-day that is about 15 lakh, and with two engineers it runs nine to ten weeks, inside the 6 to 10 weeks we quote for voice projects. The prototype on a test number comes first for a reason: if your callers' recordings score badly on every speech provider, you find out in week three for a fraction of the price, and the answer might be a WhatsApp-first receptionist instead.
The costs that are not in the quote
| Cost | Typical size | When it surprises you |
|---|---|---|
| WhatsApp Business API | A fraction of a rupee per utility message plus your provider's fee, and template approvals that take days | When the confirmation flow is ready and the templates are not |
| SMS registration | Registering your sender and templates on the DLT platform TRAI requires | When the reminder SMS silently fails to deliver |
| Practice software with no API | A connector, or in the worst case a person re-keying bookings | When "integrates with your calendar" turns out to mean Google Calendar only |
| Extra languages | A test set and a voice per language, days each | When a second location's patients speak something the first location's did not |
| Recording storage and deletion | Small in money, real in work | When a patient asks for their recordings to be deleted |
| The front desk's new habit | Training, and a morning review of what the agent did overnight | Week two, when nobody has looked at the summary |
| The people it hands off to | An on-call number, or a callback promise someone must keep | The first after-hours emergency |
How to test a vendor demo
Ten minutes on the phone beats an hour of slides. Call the demo number and:
- Talk over it mid-sentence. It should stop and listen.
- Speak Hinglish with the television on. Then do it again as your oldest patient would.
- Ask a clinical question. The only good answer is that the doctor will call you.
- Say it is an emergency. It should connect you to a person, not collect your details.
- Ask for a slot that is full. It should offer the nearest real alternative, not book it anyway.
- Give a wrong name for a booking. See how it handles not finding you.
- Ask what happens to the recording. The answer should be specific.
- Ask whether it sits on your existing number, and who owns your data if you leave.
Then ask for the last thirty days of analytics from a clinic your size. Real numbers or no deal.
How to spend less without a worse receptionist
- Start with after-hours and overflow only. Let the front desk keep the daytime calls; give the agent the ones nobody answers. You get the win with the least risk, and the recordings to tune on.
- Three languages, not ten. Add the fourth when the usage data asks for it.
- Confirm and remind on WhatsApp first, and call only the patients who do not respond. A message costs a fraction of a minute of voice.
- Keep the number, keep the staff. The agent's job is the calls the clinic was losing, not the ones it was taking.
- Measure per booked appointment, not per minute. A cheaper agent that patients hang up on is the expensive one.
Questions people ask
How much does an AI receptionist cost in India?
Platform products for clinics advertise 800 to 5,000 rupees a month plus 3 to 9 rupees a connected minute, which for a clinic taking around 1,200 calls a month works out at 10,000 to 32,000 rupees. Managed services run 15,000 to 35,000 a month all in. Building your own costs 10 to 30 lakh, with usage of 2 to 4 rupees a minute, and only pays at hospital-group volume.
Can an AI receptionist replace my receptionist?
No, and be wary of anyone who says it can. It replaces the calls nobody was answering: after hours, during rushes, on Sundays. Your front desk still handles the daytime, the walk-ins, the repeat patients and everything the agent hands off. The clinics that get value from it keep their staff and add bookings.
Does it work in Hindi and regional languages?
Hindi and English are well served. The harder problem is the mixed speech real patients use, and the difference between speech providers on that mixture is large. Test any product on recordings of your own callers before you commit, and expect a second location with a different regional language to need its own testing.
Is an AI receptionist worth it for a small clinic?
If you miss calls, yes, and most clinics miss more than they think; count the voicemails and the calls that rang out for a week before deciding. If you take under a few hundred calls a month and someone answers all of them, it is an expensive way to do what is already happening, and a better out-of-hours message is the honest fix.
How much does it cost to build an AI receptionist?
10 to 30 lakh for a custom build with an experienced Indian team, over 6 to 10 weeks, plus 2 to 4 rupees a minute of usage and maintenance of roughly 15 to 25 percent of the build cost a year. The estimate above shows where the days go. For one clinic, that money is better spent on a platform; for a group with its own systems and volume, it pays for itself inside a year.
Is an AI receptionist compliant in India?
It can be. Record consent in the call itself, in the caller's language; keep recordings and transcripts in India with a retention limit and a working deletion process; register SMS senders and templates under TRAI's rules; and get WhatsApp templates approved. The Digital Personal Data Protection Act's core obligations are being phased in through 2027, so treat them as requirements now rather than later.
Does it work with Practo or my clinic software?
Only if the product writes bookings to it, and the answer varies by product and by software. Ask for a demonstration of a booking landing in your actual calendar, not a list of logos. If your practice software has no API, a custom connector is the first thing to price, whichever option you choose.
Getting a real number
If you run one clinic, the right next step is a platform trial tested the way the checklist above describes, and we are happy to say so. If you run a group, a hospital, or software patients already use, send us a few sentences about your call volume, your systems and your languages, and we will tell you within one working day whether a build makes sense and roughly what it would cost. The process is on our voice AI development page.