Unit economics of voice agents
In one sentence
Unit economics of a voice agent is the delivered cost of a single conversation weighed against what that conversation earns, and it carries more weight here than in most software because each exchange runs up a genuine variable cost that scale never removes.
Not to be confused with Pricing models for voice agents.
Definition
Unit economics is what a single conversation costs you to deliver and what it is worth to you.
It matters more for voice than for most software, because every conversation carries a real variable cost that never goes away, no matter how large you grow.
A voice agent behaves financially more like a telecoms or infrastructure operation than like ordinary software. Read the cost stack correctly and everything downstream follows from it: pricing, margin, and the judgement on which customers are worth serving.
The cost stack, per minute of conversation
- Speech recognition, charged per minute or per hour of audio processed.
- Model inference, charged per token, which moves with prompt size, retrieval context, conversation history and response length.
- Speech synthesis, charged per character or per minute of generated audio.
- Transport. Telephony carries a per-minute carrier cost. Browser delivery over WebRTC or WebSockets does not, which removes an entire line.
- Infrastructure: media handling, vector search, storage, logging.
- All-in, this lands around USD 0.11 to USD 0.33 per minute in production.
The cost drivers you can actually control
- Conversation duration, which multiplies almost every line in the stack.
- Prompt size, which drives model cost directly. A bloated system prompt and oversized retrieved chunks are a recurring, invisible cost.
- Model selection. The spread between models on both cost and latency is large.
- Retrieval efficiency. Sending five well-chosen passages instead of fifty mediocre ones cuts cost and improves the answer at the same time.
- Caching and reuse where conversations repeat, which they do more than teams expect.
The margin structure, and why it is unusual
- Gross margin is usage-dependent rather than fixed, which is unusual for software and is the single most important thing an investor or operator needs to grasp.
- A customer with long conversations can be unprofitable at a price that is very profitable for a customer with short ones, at the same rate.
- That makes customer mix a margin variable, and it argues for watching margin per customer rather than reporting a blended average.
The value side, which is different on a website
- Set against a human-handled call at roughly USD 7 to USD 12, an agent that handles the same interaction for roughly USD 0.40 holds an enormous cost advantage, and that gap is the arithmetic a business case is built on.
- On a website the case is usually stronger still. What the widget replaces is often not a cheaper human touch but no touch at all, since the visitor would have left without asking. The value is a conversation that simply would not have happened otherwise.
- Outcome value swings hugely by vertical. A captured lead worth thousands puts the sums in a different world from a support deflection worth a few dollars.
- Winning is costly. A surge in traffic is a surge in spend, which turns rate limiting into a financial control, while a growing corpus pushes retrieval cost and latency up together. A long conversation is doubly punishing, worse to sit through and dearer to run, so shortening it wins on both counts.
Common misconception
That voice agent costs will drift toward zero as models cheapen. Per-token and per-minute prices have dropped a long way, yet usage has climbed faster still. The variable cost structure is durable, and any business betting that it vanishes is standing on ground that has not held up so far.
Why it matters commercially
Unit economics decides which customers are worth serving, what pricing is sustainable, and whether growth improves margin or destroys it. For a voice product it is a first-order strategic question, not a finance exercise done after the fact.
In voice specifically
A text chat runs no speech recognition and no synthesis, so its per-interaction stack is shorter and cheaper. Voice adds two paid stages a text widget never touches, which is part of why the same answer can cost more to deliver by voice than by text.
Where AsqVox fits
AsqVox runs in the browser rather than over a phone line, so its per-minute stack carries no telephony charge, and website conversations tend to run shorter than support calls, which compounds in its favor. On the value side, a website widget is usually measured against no conversation at all, and a single captured lead can be worth far more than a deflected query. AsqVox captures those leads and logs each conversation with its transcript and sentiment in the dashboard.
Visual
What one minute of conversation actually costs
Every conversation costs something. That never stops being true.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
All-in production cost, once speech recognition, the model, synthesis and telephony are stacked, lands around USD 0.11 to USD 0.33 per minute.
USD 0.11 to 0.33 per minindustry rangeIndustry-reported production cost range, 2026 - The delivered cost, not the platform layer. Where in the range a deployment lands is driven mostly by conversation duration and prompt size.
Platform base rates start around USD 0.05 per minute for Vapi, USD 0.07 for Retell and USD 0.09 for Bland.
USD 0.05 to 0.09 per minvendor claimVendor published pricing (Vapi, Retell, Bland), 2026 - The platform layer only. It sits below the all-in figure, which is the number that actually determines margin.
In India, Bolna prices from around Rs 7 per minute falling to around Rs 3 per minute at volume.
Rs 7 to Rs 3 per minvendor claimBolna published pricing, 2026 - A volume-tiered range rather than a single rate. Vendor pricing changes often, so date it when quoting.
A human-handled call costs roughly USD 7 to USD 12 against roughly USD 0.40 for an agent-handled call.
USD 7 to 12 vs USD 0.40industry rangeWidely cited industry range, 2025 - An industry range rather than an audited figure, and it varies substantially by geography and complexity. On a website the more honest comparison is often against no conversation at all.
Containment benchmarks run 70 to 80 percent for mature deployments, 40 to 55 percent on average, and under 35 percent for rule-based systems.
70 to 80% mature, under 35% rule-basedindustry rangeIndustry-reported containment ranges, 2026 - Directional industry bands. Containment feeds the value side of the arithmetic, since an unresolved conversation still costs money to run.
Measured time to first token ranges from around 0.78 seconds for xAI Grok Voice to around 2.98 seconds for Gemini 3.1 Flash Live, so model selection affects both cost and experience, and the two do not always move together.
0.78s to 2.98sindependentArtificial Analysis, via Softcery, April 2026, 2026 - Independent latency measurement. It is a reminder that the cheapest model is not automatically the fastest, and model choice is a cost driver you control.
There is no published benchmark for average conversation duration in website voice deployments, which is the single input most needed to model unit economics for this segment.
-no reliable figureDuration multiplies almost every line in the stack, so this gap matters more than any other. A deployment has to measure its own.
There is no published data on the distribution of gross margin across voice agent vendors.
-no reliable figureWhich is why blended margin claims should be read with care. The interesting number is margin per customer, and almost nobody publishes it.
Provider pricing changes frequently, so any specific rate needs a date and a link to the current pricing page.
-no reliable figureA standing caveat rather than a figure. A rate quoted with no date has an unknown shelf life.
Examples
In practice
A vendor reviews margin per customer and finds one large account running at negative gross margin despite paying the standard rate, because its conversations run three times the platform norm and its system prompt is unusually large. Two changes fix it: trimming the prompt and the retrieved context, and restructuring the agent's opening to reach the question faster. Margin recovers with no price change, and the customer's experience improves because the conversations get shorter.
The everyday version
Unit economics is the cost of one conversation set against its worth. A talking agent, unlike a normal piece of software, spends real money on every single use, and that bill never disappears. The encouraging part for a website is that the thing it competes with is rarely a cheaper phone call. It is the visitor who arrived with a question, failed to find the answer, and slipped away without you ever knowing they came.
Usage
Who says it
- Founders, finance teams and investors, in margin and pricing analysis.
- Platform engineers, when they are optimizing cost per interaction.
- It rarely comes up in a customer conversation, though a customer doing serious evaluation will ask about cost predictability.
Where it turns up
- In board materials, investor updates, pricing decisions and infrastructure planning.
- In vendor selection indirectly, through questions about cost predictability, rate limiting and overage.
Common misuse
- Reporting a blended gross margin with no per-customer variance, which hides the unprofitable accounts.
- Assuming cheaper models will resolve a margin problem. Usage has historically grown faster than unit costs have fallen.
- Ignoring prompt and retrieval size as cost drivers. They are invisible on a bill and substantial in aggregate.
Questions people ask
How much does one voice agent conversation cost?
All-in production cost lands around USD 0.11 to USD 0.33 per minute, once speech recognition, the model, synthesis, infrastructure and any telephony are stacked. Platform base rates are lower, around USD 0.05 to USD 0.09 per minute, but those are the platform layer only. Where in the range a conversation lands is driven mostly by its duration and its prompt size.
Why is voice agent gross margin usage-dependent?
Because every conversation carries a real variable cost that scales with its length. A customer with long conversations can be unprofitable at the same rate that is very profitable for a customer with short ones. That makes customer mix a margin variable, and it is why margin should be watched per customer rather than reported as a blended average.
What drives the cost of a voice agent, and what can I control?
Five things you can act on: conversation duration, which multiplies almost everything else; prompt size, the invisible cost of a bloated system prompt and oversized retrieved chunks; model selection, where the spread on cost and latency is large; retrieval efficiency, where fewer, better passages cost less and answer better; and caching where conversations repeat. Duration and prompt size are usually the largest levers.
Will voice agent costs fall to near zero as models get cheaper?
No. Per-token and per-minute costs have fallen substantially, but usage has risen faster, and the variable cost structure is durable. A business built on the assumption that the cost disappears is building on something that has not held so far. On a website the stronger argument is not that the conversation is cheap but that the alternative is often no conversation at all.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.