A component-by-component breakdown of voice AI agent cost in India — telephony, STT, LLM, TTS, and the architecture choice that swings the per-minute bill by 10x.
CodeKrypt Bot 6 min readA voice AI vendor will quote you a per-minute rate, and the number will look low next to what a human agent costs. What it usually will not show is that "per minute" is stacking four separate meters — telephony, speech-to-text, the LLM call, and text-to-speech — and that the architecture choice behind those four meters can change the total by 10x before a single call is placed.
This is a breakdown of what a voice AI agent actually costs to run in India in 2026: the components, the architecture decision that dominates the bill, and the build cost on top.
A voice agent call is billed in layers, whether or not the vendor shows them separately:
| Component | What it does | Typical range |
|---|---|---|
| Telephony | Carries the call, inbound or outbound | ₹1–5/min |
| Speech-to-text (STT) | Transcribes caller audio | ₹1–3/min |
| LLM | Generates the response | ₹2–8/min |
| Text-to-speech (TTS) | Synthesizes the agent's voice | ₹2–6/min |
Bundled platforms roll these into one number. Reported India-facing platform rates span roughly ₹4/min on the cheap end to ₹15–26/min once telephony, routing, and support are stacked on top, according to a 2026 pricing breakdown from Edesy. A platform quoting a flat low rate is usually excluding one of the four meters, most often telephony or a routing fee — ask which ones are in the number before comparing it to a competitor's.
The single biggest cost lever is not which vendor you pick — it's whether the agent runs as a cascaded pipeline (separate STT, LLM, and TTS models chained together) or a speech-to-speech model that processes audio natively end to end, like OpenAI's GPT-Realtime line.
Speech-to-speech sounds like the simpler architecture, and it produces noticeably better prosody and lower latency because it isn't round-tripping through a text intermediate. But it is dramatically more expensive, because the model re-processes the accumulating audio context on every turn of the conversation rather than paying once for a short text exchange. On Artificial Analysis's speech-to-speech pricing page, OpenAI's GPT-Realtime-2 (High) tier lists at $1.15 per hour of input audio and $4.61 per hour of output — multiples of what a cascaded stack pays for the equivalent STT, LLM, and TTS calls separately.
A 2026 multi-project voice AI benchmark from DestiLabs tracking 12 production deployments found the same split in practice: cascaded projects landed at $0.07–$0.13 per connected minute, while speech-to-speech projects ran $0.18–$0.21 — roughly double, with the fleet median across all-in connected-minute cost sitting at $0.105. The latency gap was real but smaller than the cost gap: speech-to-speech posted 540–580ms p50 response time versus roughly 700ms p50 for cascaded, a difference most callers on a phone line will not consciously notice.
| Cascaded pipeline | Speech-to-speech | |
|---|---|---|
| Cost per minute | Lower — usually the cheapest viable option | Higher, often 1.5–10x cascaded depending on model tier |
| Latency | Good, ~700ms p50 in production | Best-in-class, ~550ms p50 |
| Debuggability | Text transcript at every step, easy to trace a bad answer | Harder — no clean text intermediate to inspect |
| Voice control | Any TTS voice, swappable independently | Locked to what the model provider offers |
| Best fit | Cost-sensitive, high-volume, scriptable flows (order status, appointment confirmation, collections) | Low-volume, high-value calls where prosody and latency matter more than unit cost |
If your unit economics need to work at call-center volume — hundreds of thousands of minutes a month — cascaded is close to the only architecture that clears the bar. Speech-to-speech earns its premium on calls where the emotional register of the voice is part of the product, not just a nice-to-have.
Published Indian pricing guides for voice AI builds describe three rough tiers, per Edesy's 2026 breakdown:
On top of the build, budget ongoing platform or maintenance cost separately — reported figures run ₹5,000–10,000 a month for a production-tier deployment, before the per-minute usage cost described above. Treat the build quote and the running quote as two different negotiations; a low build price with an unmodeled per-minute rate is the same trap as the "cheap now, expensive at month three" pattern we describe for chatbots in AI chatbot development cost in India.
For India-facing traffic specifically, the telephony leg is where global and India-focused providers diverge most. Twilio's published outbound rates to Indian numbers run roughly ₹1.20–1.50/min, against ₹0.80–1.00/min for India-focused carriers like Exotel, per Edesy's Twilio-vs-Exotel comparison. That is before counting the operational cost of a non-INR billing relationship and the compliance work of routing calls through a carrier that isn't natively set up for TRAI-registered numbers. For a purely India-facing deployment, an India-focused telephony provider is usually the lower-friction default; a global provider earns its place when the same agent also needs to place or receive calls outside India.
The number that makes voice AI attractive is the comparison to human agent cost. A fully loaded human agent in a Tier-1 or Tier-2 city BPO — salary, infrastructure, management overhead, and idle time between calls — runs roughly ₹8–20 per minute, according to Awaaz's 2026 call center cost breakdown. A cascaded voice AI pipeline at ₹6–12/min all-in is genuinely cheaper on a per-minute basis, sometimes by half or more.
The honest caveat: that comparison only holds for calls the agent actually resolves. A voice AI agent that mishandles a call and forces the caller to call back, or escalates to a human anyway, has spent the AI cost and the human cost. Model the ROI against resolution rate at your actual call mix, not against the per-minute rate in isolation — the same trap we describe for inflated automation ROI claims generally in what AI automation actually returns.
If you're scoping a first voice AI deployment, pick the single highest-volume call type your team handles today — order status, appointment confirmation, a collections reminder — and build a cascaded pipeline for that one flow before considering speech-to-speech or a broader multi-intent agent. It is the cheapest architecture, the easiest to debug, and it gives you a real resolution-rate number to model the next flow against instead of a vendor's demo.
Our AI automation page covers how we approach these builds end to end.
A breakdown of AI chatbot development cost in India — why quotes range from under a lakh to tens of lakhs, what changes between tiers, and the running costs most proposals leave out.
AI agent memory cost is the budget line most teams discover only after their context window balloons. What naive context injection actually costs, what tiered retrieval saves, and real 2026 pricing to plan against.
AI agent observability cost is the line item most teams forget until an incident forces it. What evals and monitoring actually cost, broken into tooling, compute and human review, with real 2026 figures.