How to read this comparison
Most comparison pages in this category are written by one of the vendors and rank the author first. This one is written by a vendor too, so the useful thing we can offer is method rather than a verdict: every figure below was read off the vendor’s own live page, each cell says “not published” where nothing is published, and our own entry is marked as ours.
The most expensive mistake buyers make here is reading the per-minute rates as like-for-like. They are quoted in the same unit, but they do not cover the same things. A ₹3 platform rate buys the engine; the telephony, the models, the DLT registration and the engineering time to build and run the agent are all still ahead of you. A managed rate is the car with a driver in it. The four categories below exist to make that difference visible before the table does.
Every platform that will take an Indian call
Fourteen vendors that will place or answer a call on an Indian phone line, on the four facts that decide the shortlist.
| Platform | Type | Languages (as published) | Published pricing | Time to live |
|---|---|---|---|---|
| Troika Techours | Managed service | 9+ Indian languages including Hindi, Marathi, Tamil, Telugu, Bengali, Gujarati, Malayalam, Marwadi and English, with code-switching inside a call | ₹10–12 per minute, per-second billing, everything included | Live in 48 hours |
| Sarvam AI | Build it yourself | Speech-to-text in 12 Indic languages, text-to-speech in 11 | ₹3.00–₹3.50 per minute, models included | Under 5 minutes to a first agent, per its own page |
| Bolna AI | Build it yourself | 20+, including Hindi, Hinglish, Tamil and Telugu, switchable mid-call | From $0.06 per minute plus telephony and platform fee | Templates in the dashboard, same day |
| Vapi | Build it yourself | Not stated on its site | $0.05 per minute hosting, models and telephony billed separately | Developer build |
| Retell AI | Build it yourself | "Multiple languages", not listed | $0.07–$0.31 per minute | No-code studio, same day |
| ElevenLabs | Build it yourself | 90+ languages | $0.08 per minute, LLM billed separately | Self-serve |
| OmniDimension | Build it yourself | Not stated on its site | Free tier, paid rates not published | Self-serve |
| Gnani.ai | Enterprise suite | 40+ | Not published | Sales-led |
| Haptik | Enterprise suite | 100+ | Not published, demo on request | Sales-led |
| Yellow.ai | Enterprise suite | 500+ claimed | Free plan, enterprise pricing on request | Sales-led |
| PolyAI | Enterprise suite | Multilingual, not listed | Not published | Sales-led |
| Exotel | Cloud telephony | Not stated on its site | Not published, free trial and custom quotes | Sales-led |
| Ozonetel | Cloud telephony | 24 vernacular languages | Not published | Sales-led |
| Gupshup | Cloud telephony | Not stated on its site | "Get started for free", plan prices not published | Self-serve to sales-led |
Every figure above was read from the vendor’s own live page in November 2026. A cell reading “not published” means the vendor publishes no figure, not that none exists. Rates are quoted in the vendor’s own currency. Check anything you intend to rely on — pricing in this category moves.
Four kinds of vendor, not one market
These vendors are not competing with each other in the way a single table implies. They are four different purchases.
- Platforms you build on. Sarvam, Bolna, Vapi, Retell, ElevenLabs and OmniDimension. You get an API or a studio; you supply the call flow, the telephony, the CRM integration and someone to own it. Rates are the lowest here, and the rate is not the cost — the engineering time is. Right if you have developers and want control.
- Enterprise conversational AI suites. Gnani, Haptik, Yellow.ai and PolyAI. Voice sits inside a larger platform covering chat, web and email, sold through a sales team with pricing on request. Procurement runs in months, not days. Right if you are a bank, an insurer or a large enterprise with a security review to pass.
- Cloud telephony that added voice AI. Exotel, Ozonetel and Gupshup. Their strength is the phone network itself — numbers, routing, contact-centre tooling — with AI agents added on top. Right if the call infrastructure matters to you more than the conversation does.
- Managed services. Troika Tech, and few others in India. Nothing is handed over for you to build: the flows are written, the numbers registered, the campaigns run and the results reported. Dearest per minute, cheapest in staff time. Right if you want calls happening this week and have nobody to assign to it.
What the per-minute rate leaves out
A per-minute rate is rarely the whole bill. Reading the published pages side by side, this is what sits outside the headline number on the self-serve platforms:
- The telephony. Vapi bills hosting at $0.05 a minute and the phone line separately through your provider. Bolna lists telephony and a platform fee on top of its $0.06. Sarvam is the exception that bundles the models into its ₹3.00–₹3.50.
- The models. ElevenLabs charges $0.08 a minute for the agent and bills the language model separately. Vapi bills models at cost. Two agents on the same advertised rate can differ by half again depending on which model sits behind them.
- The numbers. Promotional calling in India needs a 140-series number and transactional calling a 160-series, both after DLT registration. No platform rate includes that work, and it is the step that most often delays a launch by weeks.
- The build. The call flow, the objection handling, the CRM write-back, the retry logic. On a platform this is your engineering time, and in the first three months it is almost always the largest line.
The comparison that survives contact with a finance team is cost per useful outcome — a qualified lead, a confirmed order, a collected payment — not cost per minute. A ₹3 per minute agent nobody stays on the line for is dearer than a ₹12 call that produces an answer.
TRAI, DND and the numbers you dial from
This is the section most comparisons in this category leave out, including the one currently ranking for these searches. It is also the first wall every Indian buyer hits — usually in week one, after the agent already works.
- Number series. Promotional calls must originate from a 140-series number, transactional calls from a 160-series. Using the wrong one is the most common reason a campaign gets cut off.
- DLT registration. The entity, the sender header and the call templates are registered with the operator before volume runs. This is paperwork with a lead time, not a setting in a dashboard.
- DND. Numbers on the Do Not Disturb register may not receive promotional calls, so lists are scrubbed against DND before dialling. That obligation sits with whoever runs the campaign, not with the platform.
- The calling window. Promotional voice calls are permitted between 9 AM and 9 PM. An agent that dials outside it is a compliance problem no matter how good the conversation is.
- AI disclosure. Telling the person they are speaking to an automated system is both the emerging expectation and, in our experience, better for completion than concealing it.
Self-serve platforms hand you the tools and leave the obligations with you. That is a legitimate way to sell software. It is worth knowing before the first invoice rather than after the first complaint.
How to choose between them
Six checks, in the order that saves the most money.
- Test in the language your callers actually use. Not textbook Hindi — the mix your customers speak, with real names, addresses and amounts inside the sentence. Most platforms handle a clean sentence. Far fewer handle a Mumbai address read at speed, or a caller switching language halfway through.
- Settle the phone numbers before anything else. Ask which series the vendor provides, whether DLT registration is included and who does it. A working agent with no compliant number to dial from is the most common way these projects stall.
- Price one connected outcome, not one minute. Use your own numbers: dials to a connect, connects to a qualified answer. Then apply each vendor’s rate, telephony and models included, to that chain. The ranking usually changes.
- Check the handoff to a person. Can the agent transfer a live call, and what does the human receive with it? A transfer that arrives with no context is worse than none, because the customer repeats everything they just said.
- Check where the data lands. A dashboard you have to export is not an integration. Look for a webhook or API that writes the transcript and the captured fields into your CRM while the lead is still warm.
- Decide honestly who will build it. If the answer is “we will find someone”, a platform rate is not your real cost. If you have developers who want control, a managed service is overhead you do not need.
Where Troika Tech fits — and where it does not
Ours is the managed entry, and the trade is explicit. At ₹10–12 a minute with per-second billing we are two to four times the platform rates — Sarvam at ₹3.00–₹3.50, Bolna at roughly ₹5.50. What the difference covers is precisely what those rates exclude: we write the call flows, register the numbers and the DLT templates, scrub against DND, run the campaigns inside the permitted window, and put transcripts into your CRM rather than a dashboard for you to export.
The agents run in 9+ Indian languages and will switch inside a single call, which is how Indian households actually talk. Setup is 48 hours, and the rate is published on this site rather than quoted after a discovery call.
Where we are the wrong choice: if you have engineers who want to own the stack, buy a platform — Sarvam if Indic accuracy matters most, Vapi if you want API control. If you need voice inside a wider enterprise CX suite with a procurement process behind it, talk to Gnani or Haptik. We fit when you want calls going out this week and have nobody internally to assign to it.
Questions buyers ask about AI voice agents in India
Sarvam publishes speech-to-text in 12 Indic languages and text-to-speech in 11. Bolna publishes 20+ including Hinglish, switchable inside a call. Ozonetel publishes 24 vernacular languages and Gnani 40+. Troika Tech runs 9+ Indian languages with code-switching. Vapi, Exotel, Gupshup and OmniDimension state no figure on their sites. Language counts are not comparable across vendors — some count every language the underlying model touches, others only those tested on Indian phone lines. Test the two or three you actually need.
Published rates run from about ₹3 a minute at the low end to $0.31 at the high end. Sarvam publishes ₹3.00–₹3.50 with models included; Vapi $0.05 for hosting only; Bolna from $0.06 plus telephony and a platform fee; ElevenLabs $0.08 plus the language model; Retell $0.07–$0.31. Troika Tech charges ₹10–12 a minute with per-second billing, as a managed service rather than a platform rate. Gnani, Haptik, PolyAI, Exotel and Ozonetel publish no rates at all.
Promotional calls need a 140-series number and transactional calls a 160-series, and both require DLT registration of the entity, the header and the templates before volume runs. Most self-serve platforms let you bring your own number or provision through a partner, and leave the registration to you. Confirm who does this before you sign anything — it is the step that most often delays a launch.
It is legal, within TRAI’s rules for commercial communication. Numbers on the Do Not Disturb register may not receive promotional calls, so lists are scrubbed against DND before dialling; promotional calling is permitted between 9 AM and 9 PM; and the originating number must match the call type. These obligations sit with whoever runs the campaign. On a self-serve platform, that is you.
It depends entirely on who will build it. A platform at ₹3 a minute is cheaper only if you already have the engineering time to write the flows, handle the telephony and maintain it. If that time has to be bought or borrowed it is usually the largest cost of the first quarter, and it appears on no pricing page. A managed service costs more per minute and nothing in staff time.
Most platforms here support live transfer, but what the human receives varies a great deal. The question to ask in a demo is not “can it transfer” but “what does my agent see when the call lands”. If the answer is a ringing phone with no transcript, your customer will repeat everything they just said.
Sarvam advertises under five minutes to a first agent and Bolna offers dashboard templates, but those timings describe a working demo, not a compliant campaign dialling real numbers — DLT registration alone runs longer than that. Enterprise suites are typically measured in months. Troika Tech quotes 48 hours to a live managed campaign.
Because leaving ourselves off would be less honest, not more. Our entry is marked as the managed option, our price is stated, and our section says plainly where we are the wrong choice. Every competitor figure here comes from that vendor’s own published page, so you can check all of it.