AL-2026-01 · White paper
Voice AI economics for Indian businesses: the per-minute cost stack
What a minute of automated conversation actually costs, layer by layer, and why the carrier leg sets the floor no software can go below.
Buyers evaluating voice AI are usually given a single per-minute number and no way to reason about it. This paper decomposes that number into its layers, identifies which of them are software costs and which are bought capacity, and explains why the floor sits where it does.
1. Why the headline number hides the decision
A per-minute price bundles at least three unrelated things: the telecom leg carrying the audio, the compute running the agent, and the number the call arrives on. They have different cost structures, different scaling behaviour, and different parties responsible for them. A buyer comparing two vendors on the bundled figure alone is comparing two sums whose terms may not correspond at all.
The practical failure this causes: a quote that looks cheaper per minute because it excludes dialing, discovered at the first invoice.
2. The layers
2.1 The carrier leg
Every minute of conversation is a minute of telephony, carried by licensed carrier partners. On Voiceplix this is ₹0.65 per minute, and it is the same rate that applies to a plain click-to-call from Astrosia CRM with no AI on the line at all.
That equivalence is the most useful fact in this paper. It isolates the telecom cost from the software cost precisely, because the identical leg is available with and without an agent on it. Anything above ₹0.65 is what the automation costs; anything at or below it is not a real quote.
2.2 The agent
The agent layer is recognition, reasoning, and synthesis, running for the duration of the call. Unlike the carrier leg, its cost per minute falls with volume. The fixed components amortise, and capacity can be planned rather than burst.
This is why published rates form a ladder rather than a single number. Smaller packs sit higher on that ladder; the deepest published tier on Voiceplix is ₹2.65 per minute all-in, inclusive of the ₹0.65 carrier leg. Self-serve packs run up to 20,000 minutes a month, above which pricing is quoted.
2.3 The number
An Indian virtual number is ₹250 per month, charged per number rather than per minute, and issued only after telecom KYC is filed in the company's name. For a business running one line this is a rounding error against usage; for one running many, it is a real line item that scales with footprint rather than with conversation volume.
2.4 Tax
18% GST is added at checkout. Every figure in this paper is exclusive of it.
3. Putting it together
| Layer | Basis | Scales with | Controlled by |
|---|---|---|---|
| Carrier leg | ₹0.65 / min | Minutes | Licensed carrier partners |
| Agent | Balance of the all-in rate | Minutes, with volume discount | Vendor |
| Number | ₹250 / month | Number of lines | Vendor, post-KYC |
| GST | 18% | Everything above | Statute |
At the deepest published tier, a business running 10,000 minutes a month on one number pays ₹26,500 in minutes plus ₹250 for the line, before GST. Of that, ₹6,500 is telecom the vendor bought and ₹20,000 is the automation.
How to read a competing quote: ask what the carrier leg costs separately. If the vendor cannot separate it, the quote is either excluding dialing or bundling it opaquely, and the two are not distinguishable from the outside.
4. What actually moves the total
The instinct is to negotiate the per-minute rate. In practice, on the deployments we have seen, the larger lever is minutes consumed, and that is a function of how the agent behaves, not what it costs.
- Turn latency. An agent that is slow to respond causes callers to repeat themselves, which lengthens calls. The latency budget described in the turn budget note is therefore also a cost control.
- Language handling. A call where the agent mishandles code-mixed speech gets restated, and restatement is billable minutes. See code-mixing is the default.
- Clean handoff. An agent that cannot recognise it should transfer burns minutes before the human conversation starts anyway.
A rate that is ten percent lower on an agent that produces twenty percent longer calls is more expensive. This is the comparison buyers most often fail to make.
5. What this paper does not cover
Human cost. Replacing or augmenting a calling desk changes staffing, and that arithmetic is specific to the business, its wage structure, and what it does with freed capacity. We have deliberately kept this paper to the vendor-side stack, because that is the part where the numbers are the same for everyone and can be stated without guessing.
6. Sourcing and currency
Rates in this paper are the published Voiceplix rates as of the date above, and pricing changes. The authoritative figures are on the Voiceplix pricing page; where this paper and that page disagree, the pricing page is correct.