Voice AI Vendor Pricing in India 2026: What Gnani, Ozonetel, Bolna, Vapi and ElevenLabs Actually Charge
The procurement lead at a Hyderabad NBFC had four quotes on screen and no way to compare them. One vendor quoted ₹3.50 a minute. One quoted a platform fee plus ₹2.20 a minute. One quoted per successful conversation. One quoted in dollars, per minute, and did not mention telephony at all.
She did the obvious thing and ranked them by the visible number. Six months later the cheapest quote had produced the highest invoice, by a margin of roughly 2.4x over what the finance model had assumed, and nobody had done anything wrong. Every rupee was contractually owed. The rate card had simply been the smallest part of the price.
Voice AI pricing in India is not expensive so much as it is structurally opaque. Vendors price different units, bundle different components, and put the variable costs in places a per-minute comparison never touches. This post decodes how each of the major platforms selling into India actually prices, what their published rate cards leave out, and how to convert any quote into a single comparable number: effective cost per minute, and then cost per outcome.
One thing to state up front. Published list prices move, and enterprise quotes routinely land well below list. Every figure here describes a pricing structure and a range observed in the Indian market, not a contract rate you should hold a vendor to. Use it to ask better questions, not to argue about a number.
Why this got harder in 2026
Three shifts made the comparison problem worse rather than better.
The market split into two pricing philosophies. Indian enterprise vendors, the ones with contact centre heritage, price like telecom and CCaaS companies: platform fee, per-seat or per-concurrency charges, committed volumes, annual contracts. Developer-first global platforms price like infrastructure: usage-based, per minute, self-serve, no commitment. These two models are not comparable on any single axis, and most shortlists contain both.
The model layer got cheaper while the telephony layer did not. Speech recognition and synthesis costs fell sharply through 2025. Indian telephony did not. The result is that the component vendors love to quote, the AI, is now a minority of your real cost on many workloads, and the component they quote least clearly, the telephony last mile, is a growing majority.
Outcome-based pricing arrived and muddied everything further. Charging per connected conversation, per qualified lead or per resolved contact is genuinely better aligned for some workloads. It is also completely incomparable to a per-minute rate without knowing your connect rate, which most buyers do not know at quote stage.
The five things every voice AI quote is actually made of
Before comparing vendors, decompose the quote. Every Indian voice AI bill is some combination of these five, and vendors differ mainly in which ones they make visible.
1. The AI layer. Speech-to-text, the language model, text-to-speech. This is what gets quoted as the headline per-minute rate on developer platforms. It is genuinely commoditising and it is usually the smallest line.
2. Telephony. The actual call: origination or termination, the number, the carrier. In India this is priced in paise per minute and varies by whether the call is landing on mobile or landline, which circle, and which carrier. It is frequently excluded from the headline rate on global platforms and bundled on Indian ones. This is the line that surprises people.
3. Platform and orchestration. Call routing, session management, the dashboard, integrations, the conversation designer. Usually a fixed monthly fee on enterprise vendors, usually invisible and folded into the per-minute rate on developer platforms.
4. Concurrency or capacity. How many calls you can run simultaneously. Enterprise vendors often price this explicitly and it is the line that bites during a campaign burst or a collections push at month end. Developer platforms usually rate-limit rather than charge, until you need a raised limit.
5. The stuff that is not in the rate card at all. DLT registration and template management, number provisioning, recording storage, transcript retention, custom voice, professional services for onboarding, sandbox and staging environments, and support tiers. Individually small. Collectively, often 15 to 30% of the first-year bill.
A quote that shows you only line 1 is not cheaper. It is less complete.
How each vendor actually prices
Structures, not contract rates. Verify current list pricing directly with each vendor before modelling.
| Vendor | Pricing model | Telephony included | Commitment | Where the cost hides |
|---|---|---|---|---|
| Gnani.ai | Enterprise: platform fee plus usage, often per-bot or per-deployment; quoted, not self-serve | Usually bundled or arranged | Annual contract typical | Professional services, per-language or per-bot expansion, concurrency tiers |
| Ozonetel | CCaaS heritage: per-agent or per-seat plus usage, cloud telephony bundled | Yes, it is their core business | Monthly to annual | Seat licences you keep paying for during low season; add-on modules |
| Bolna | Developer-first, usage-based per minute, self-serve entry | Typically not, you bring or buy separately | Low or none | Telephony added separately; scale pricing negotiated |
| Vapi | Developer-first, per minute, plus pass-through for STT/LLM/TTS providers you choose | No, separate telephony provider | None at entry | Pass-through model costs stack on top of the platform minute; USD billing |
| ElevenLabs | Per minute for conversational AI, credit-based for TTS, tiered subscription | No | Subscription tiers | USD pricing and FX; telephony last mile entirely on you; India latency |
| Nurix | Enterprise, quoted, outcome and deployment framed | Arranged | Annual typical | Scope-based; expansion priced per use case |
| Caller Digital | Per minute or per outcome, INR, telephony bundled for India | Yes | Flexible | Published INR rates; see pricing |
Read that table for shape rather than for a winner. The Indian enterprise vendors (Gnani, Ozonetel, Nurix) sell a managed outcome with telephony and compliance handled, priced as a contract. The developer platforms (Bolna, Vapi, ElevenLabs) sell a component you assemble, priced as usage. Both are legitimate. They are appropriate for different buyers, and comparing their headline numbers is close to meaningless.
The specific trap in each model
Enterprise contract pricing traps you on utilisation. You commit to seats, concurrency or volume, and then your actual usage is seasonal. An NBFC collections workload peaks in the first ten days of the month and is quiet after. If you sized concurrency for the peak, you are paying for it during the trough. Ask for burst pricing rather than sizing to peak.
Developer usage pricing traps you on the components you did not price. A Vapi or ElevenLabs quote covers the platform. Your bill also includes the STT provider, the LLM provider, the TTS provider where separate, a telephony provider that can terminate to Indian mobile numbers reliably, and the engineering time to hold it together. Teams routinely model the first number and are then surprised by a bill three to four times higher.
USD pricing traps you on FX and on the India last mile simultaneously. A rate that looks competitive at one exchange rate is a different number in your INR budget a quarter later, and global platforms rarely have strong Indian carrier relationships, which shows up as connect rate rather than as cost. A 10% worse connect rate is a 10% worse cost per outcome regardless of the per-minute price.
Turning any quote into one comparable number
Two calculations. Do both for every vendor on the shortlist.
Effective cost per minute
Take everything you will pay in a year and divide by the minutes you will actually run.
Effective ₹/min = (platform fees + AI usage + telephony + concurrency + storage/retention + DLT/number costs + support + amortised onboarding) ÷ annual billable minutes
The gap between headline and effective is the whole story. Across Indian deployments, headline rates commonly sit in the ₹2 to ₹12 per minute band while effective cost lands between ₹6 and ₹25. A vendor quoting ₹2.20 with telephony and platform excluded is frequently more expensive in practice than one quoting ₹7 all-in.
Two details people get wrong in this calculation. Bill on billed seconds, not talk time: most carriers bill in 30 or 60 second increments, so a 35-second call is often billed as 60 seconds, and a workload of short calls carries a rounding penalty of 20 to 40%. And include failed calls: unanswered, busy and rejected calls consume telephony attempts even when no conversation happens.
Cost per outcome
This is the number that should drive the decision.
₹/outcome = effective ₹/min × avg minutes per connected call ÷ (connect rate × completion rate)
Worked example, an EMI reminder workload:
| Input | Vendor A | Vendor B |
|---|---|---|
| Headline rate | ₹2.50/min | ₹7.00/min |
| Effective rate all-in | ₹9.80/min | ₹11.20/min |
| Avg minutes per connected call | 1.4 | 1.2 |
| Connect rate | 31% | 44% |
| Completion rate on connect | 71% | 82% |
| Cost per completed conversation | ₹62.35 | ₹37.26 |
Vendor A is 64% cheaper on the headline and 67% more expensive per outcome. The difference is almost entirely connect rate and call efficiency, which is a function of carrier relationships, caller ID reputation and conversational quality, none of which appear on a rate card. This is why the per-minute versus per-outcome pricing question matters more than the rate itself.
What goes wrong in vendor pricing evaluations
Comparing a bundled Indian quote to an unbundled global one. The single most common error. Normalise first or the comparison is noise.
Sizing concurrency to peak. Month-end collections and festive campaign bursts are real, but paying for peak capacity for twelve months to serve it for two is a large avoidable cost. Negotiate burst.
Ignoring the minimum commitment. Enterprise contracts frequently carry a monthly minimum that you pay whether you use it or not. In a pilot year this is often the largest single line and it never appears in a per-minute comparison.
Treating recording storage as free. Regulated workloads in BFSI and insurance retain audio and transcripts for years. At scale this is a real line item, and some vendors price egress on retrieval, which turns an audit request into an invoice.
Believing the dashboard's minute count. Reconcile the vendor's reported minutes against your telecom bill in month one. Discrepancies between platform-reported connected minutes and carrier-billed minutes are common and rarely resolve in your favour if you find them in month nine.
Not pricing the exit. Ask what it costs to export your call recordings, transcripts and conversation designs, and in what format. A vendor who charges for your own data on the way out has priced your switching cost, not your usage.
Contract clauses that decide the real price
Six clauses matter more than the rate. We have covered these at length in the seven contract clauses that decide whether ₹3 a minute is cheaper than ₹9, but in summary:
- Billing increment. Per second, or per 30 or 60 second block. On short-call workloads this alone moves the bill 20 to 40%.
- What counts as a billable minute. Does ring time count? Does a call that hits voicemail count? Does a failed transfer count twice?
- Minimum commitment and true-up. What you owe in a quiet month, and whether unused volume rolls forward.
- Concurrency and burst. The cost of exceeding contracted concurrency, and whether burst is available at all.
- Price escalation. Annual uplift clauses, and whether FX movement can be passed through on USD-denominated contracts.
- Data export and termination. Format, cost and timeline for getting your recordings and transcripts out.
For regulated buyers, add data residency. Where the audio is processed and stored has both a compliance consequence and a cost consequence, and the two interact. Our note on voice AI data residency and sovereignty in India covers the compliance side.
A four-week vendor pricing evaluation
Week 1: establish your own numbers. You cannot evaluate pricing without knowing your workload. Pull your actual call volume, average handle time, connect rate by hour and circle, and seasonality across twelve months. Most teams cannot answer "what is our connect rate" at quote stage, which is precisely why they cannot compare outcome-based quotes.
Week 2: normalise every quote. Force each vendor onto the same decomposition: AI, telephony, platform, concurrency, everything else. Where a vendor will not break it out, model the missing line at market rate and tell them you have done so. Ask each for a quote at three volumes: your trough month, your average, and your peak.
Week 3: run the same audio through the shortlist. Give every vendor the same 50 recorded calls from your hardest segment. Measure word error rate on numbers and proper nouns specifically. Then measure connect rate on a live 500-call pilot, because that is the input that dominates cost per outcome and no rate card discloses it.
Week 4: model twelve months, then negotiate. Build the effective-rate and cost-per-outcome model at all three volumes. Negotiate on billing increment, minimum commitment and burst before negotiating on the headline rate. The headline rate is the line vendors expect you to push on and the one where they have the least room.
What changes over the next 12 months
The AI layer keeps getting cheaper and will keep shrinking as a share of the bill. Anyone whose competitive position rests on model cost is in a bad position.
Telephony becomes the moat. Carrier relationships, caller ID reputation management and connect rate optimisation are where cost per outcome will actually be won in India, and they are hard to replicate.
Outcome pricing spreads, and with it a new opacity: who defines the outcome, and who audits the count. Expect the negotiation to shift from rate to definition. Insist on the raw event log, not the dashboard summary.
Bottom line
Voice AI vendors in India are not lying about their prices. They are quoting different things. The Indian enterprise platforms sell a managed outcome with telephony and compliance included and price it as a contract; the developer platforms sell a component and price it as usage; both look cheaper than the other on the axis they choose to publish. Decompose every quote into AI, telephony, platform, concurrency and the unlisted extras, compute effective cost per minute, then compute cost per outcome using your own connect rate. A headline rate of ₹2.50 routinely costs more per completed conversation than a headline rate of ₹7. Negotiate billing increment and minimum commitment before you negotiate the rate.
If you want a like-for-like model built against your own volume and connect rate, our India pricing page publishes INR rates and we will run the comparison with you, including against vendors we lose to.
How to read a vendor's pricing page without being misled
Published pricing pages are marketing documents. Four habits make them readable.
Find the unit before the number. A page showing "$0.05" is meaningless until you know per what. Per minute of conversation, per minute of connected call including ring time, per 1,000 characters of synthesised speech, per session, per agent seat. Character-based pricing in particular is impossible to compare against per-minute pricing without knowing your average utterance length, and Indian-language synthesis consumes more characters per second of audio than English does, which quietly inflates character-billed workloads in Hindi and the southern languages.
Look for what the free tier excludes. Generous free tiers on developer platforms almost always exclude telephony, concurrency above one or two calls, and any production support. They are useful for evaluating conversational quality and useless for estimating cost.
Check the currency and the escalation. USD-denominated pricing exposes your budget to FX movement across a multi-year contract. Ask explicitly whether the vendor can quote and invoice in INR, and whether the rate is fixed in INR or merely converted at invoice date. Those are very different commitments.
Treat "custom" and "contact sales" as information. When a vendor stops publishing at the tier you need, it means the price is a function of your negotiating position rather than a rate card. That is not sinister, but it does mean you should arrive with a modelled alternative and a walk-away number.
Workload shape changes which vendor is cheapest
There is no cheapest vendor, only a cheapest vendor for a given workload shape. Four shapes dominate Indian deployments and they invert the ranking.
Short, high-volume outbound. EMI reminders, COD confirmation, delivery scheduling. Calls of 30 to 70 seconds, enormous volume, month-end concentrated. Here the billing increment dominates everything: at 60-second rounding a 38-second call costs the same as a 58-second one, and a vendor with per-second billing at a higher headline rate wins comfortably. Concurrency burst pricing matters because volume is spiky. See our EMI payment reminder use case for the workload profile.
Long, low-volume inbound. Support, triage, complex enquiry. Calls of 4 to 12 minutes, steady volume. The per-minute rate genuinely dominates here and rounding is irrelevant. Conversational quality and interruption handling matter more than price, because a failed call costs a human callback.
Bursty campaign outbound. Festive campaigns, launch pushes, election-style outreach. Two weeks of extreme volume then nothing. This is where enterprise concurrency commitments are punishing and usage-based platforms win, provided their rate limits can be raised temporarily.
Regulated, recorded, retained. BFSI and insurance workloads where every call is stored for years and retrievable on demand. Storage and retrieval pricing, which nobody looks at, becomes a top-three line item. Ask about egress cost on retrieval specifically, because an audit request that pulls 40,000 recordings should not generate an invoice.
Model your own shape before you shortlist. A team that knows its average handle time, connect rate, monthly distribution and retention obligation can evaluate four quotes in a morning. A team that does not will pick on headline rate and be wrong.
One number to walk in with
Before the first vendor call, compute your own ceiling: the cost per outcome above which the workload stops being worth automating. For an EMI reminder workload that is some fraction of the amount recovered per successful contact. For a lead qualification workload it is a fraction of your cost per qualified lead through existing channels. For a support workload it is your fully loaded cost per human-handled contact.
That single number changes the conversation. It converts "is ₹7 a minute expensive" into "at our connect rate and handle time, ₹7 a minute lands at ₹41 per resolved contact against a ₹95 human baseline", which is a question you can actually answer. It also tells you when to walk away, which is the only real leverage a buyer has.
Frequently Asked Questions
Tags :
