Hardware Infrastructure
Unit Economics
Can Voice AI Agents Really Run at ₹2 a Minute in India?
6mins
Aryan Kushwaha
Can Voice AI Agents Really Run at ₹2 a Minute in India?
Every voice AI founder in India says the same thing at some point: the American pricing playbook doesn't work here. A ₹20-25 per minute agent might be fine for a US enterprise chasing a $50 customer, but it makes zero sense for a Bharat use case where the entire transaction is worth ₹200. If voice AI wants to reach tier-2 and tier-3 India at scale, it needs to cost close to what a phone call costs, not what a cloud API costs.
The direct answer: getting to ₹2 a minute is technically possible, but only by stripping the stack down to open, self-hosted models and eating a lot of engineering pain that the bundled platforms currently hide from you. The real fight isn't cutting cloud fees. It's whether a stripped-down pipeline can survive a bad network and a Bihari or Malayali accent at the same time.
This piece looks at what actually has to be true for a ₹2/min voice agent to work in production, not just in a demo.
The thesis: cost and quality stop being a straight trade-off once you accept that Bharat telephony is a different problem than US telephony. The startups that win the ₹2/min game will be the ones that engineer for noise and packet loss first and optimize for cost second, not the other way around.
What Does a Voice Agent Actually Cost, and Where Does ₹2/min Come From?
Global per-minute pricing gives a useful anchor. Bundled platforms like Retell, Bland and ElevenLabs typically land between $0.10 and $0.18 a minute once every layer is added up, while a bring-your-own-key setup on Vapi or a raw build on OpenAI Realtime plus Twilio can run anywhere from $0.13 to $0.30 a minute without careful prompt caching. Converted to rupees, that is roughly ₹9 to ₹26 a minute, which is exactly the number Bharat-focused builders are trying to escape.
Every voice call, wherever it runs, is really five things stacked on top of each other:
Telephony: the actual PSTN or SIP trunk minute
Speech-to-text (STT): converting the caller's audio to text
The LLM: understanding intent and generating a response
Text-to-speech (TTS): converting the response back to audio
Orchestration: the glue that manages turn-taking, latency and state
Cloud-hosted, proprietary versions of all five layers is what produces the $0.10-0.30/min number. Getting to ₹2/min means replacing most of that stack with open-source models running on infrastructure you control, so you are paying for compute and bandwidth instead of per-token and per-character API fees. That is the entire premise of the "2-rupee" bet: hyper-optimized local pipelines instead of rented intelligence.
Why Is Bharat Telephony a Harder Problem Than US Telephony?
This is the part that gets skipped in most pricing pitches. A model that scores well on a clean English benchmark can fall apart the moment it meets real Indian telephony conditions, and that has nothing to do with the model's intelligence.
Three things make Bharat calls uniquely punishing:
Regional accents and code-switching. A caller in rural UP might switch between Hindi, English and a local dialect mid-sentence. Most open-source ASR models are trained overwhelmingly on clean, single-language audio, so accuracy drops sharply on this kind of speech even before you touch network quality.
Packet loss and low-bandwidth mobile networks. A large share of Bharat's target users are on 2G/3G-adjacent connections or congested cell towers, where audio arrives clipped or jittery. Cloud ASR built for fibre-quality audio wasn't tuned for this.
Strict latency budgets. Conversational voice AI needs to respond inside roughly 500-800 milliseconds to feel natural. Every millisecond spent recovering from a dropped packet or re-running a failed transcription eats directly into that budget.
Stack all three together and you get the real stress test: does a stripped-down, cost-optimized pipeline hold conversational accuracy when the accent is unfamiliar, the network is unreliable and the latency window is tight. A model that's cheap but keeps mishearing "haan" as "na" isn't actually cheap. It's a support ticket, a churned customer, or a failed collection call.
Does Going Open-Source Actually Get You to ₹2/min?
In principle, yes, and the math is the whole appeal. Once a startup owns its GPU or CPU inference instead of renting tokens, the marginal cost of a call collapses toward pure compute and bandwidth. Quantized, distilled versions of open ASR and TTS models can run on modest hardware, and an efficient small LLM handling a narrow, well-scoped conversation (loan reminders, appointment confirmations, KYC verification) doesn't need frontier-model reasoning to do its job well.
But the honest caveat: nobody has published verified, at-scale numbers showing a production voice agent holding both sub-₹2/min unit economics and high accuracy on noisy regional calls simultaneously. Every public pricing benchmark available today is built around clean, English-first, well-connected telephony. The operators actually running Bharat deployments (BPOs doing vernacular collections calls, agri-advisory lines, government helpline pilots) are the ones who would know whether the accuracy holds up once the cost comes down this far, and that data isn't out there yet. This is the open question this piece can't responsibly fill with a guess.
Cost layer | Bundled cloud stack (approx.) | Self-hosted open-source stack (target) |
|---|---|---|
Telephony | ₹1.2-1.5/min | ₹1.2-1.5/min (largely fixed, carrier-driven) |
STT | ₹0.4-0.6/min | Near-zero marginal cost once hosted |
LLM | ₹2-4/min | Fraction of a rupee with a small, quantized model |
TTS | ₹1-2/min | Near-zero marginal cost once hosted |
Orchestration/platform fee | ₹4-8/min | Removed or replaced with in-house code |
Total | ₹9-16/min | ₹2-3/min, if accuracy holds |
The telephony line is the one nobody can optimize away. Even a fully open-source, self-hosted pipeline still pays a real carrier for the PSTN minute, which puts a floor under how low the total can actually go.
Where Does This Go From Here?
The next 12-18 months will likely split Bharat voice AI into two camps: platforms that keep selling a premium, accuracy-first product to enterprises who can afford ₹15-20/min, and a second wave of builders who go all-in on self-hosted, open-source pipelines for high-volume, low-value use cases like collections, appointment reminders and government helplines where ₹2-3/min is the only price that makes the unit economics work. The winners in that second camp won't be the ones who found the cheapest model. They'll be the ones who solved noisy-accent ASR and packet-loss recovery well enough that "cheap" stops being a euphemism for "worse."
[
FAQ
]
Frequently Asked Questions
[
Browse Articles
]
Browse More Articles
Explore content across the voice AI stack — from infrastructure to real-world applications
