BLOG · UNIT ECONOMICS
A human-handled call costs $7.16. An AI-handled minute costs between $0.07 and $0.15. US adoption went from 8% to 34% in two years, and two-thirds of the market is still unserved. The margin is real — and it is not where most operators think it is.
The voice AI margin is one of the few genuinely large arbitrages still open in software. It is also routinely misunderstood by the people selling it, usually in the same direction: they assume a good gross margin on minutes is the same thing as a good business.
Here is the arithmetic, and the threshold where it stops working.
A human-handled call costs approximately $7.16. An AI-handled minute costs between $0.07 and $0.15. Even at a generous five-minute call, the AI side lands near $0.35 to $0.75 against $7.16 — roughly a tenfold spread on the raw transaction.
The market context explains why operators find this irresistible. Voice AI is a $22.5 billion market in 2026 growing at 34.8% CAGR, with voice agents specifically growing at 39%. US adoption moved from 8% to 34% between Q1 2024 and Q1 2026. Contact-centre labour savings are projected around $80 billion in 2026.
And crucially: 66% of the market is still unserved. Two thirds of buyers have not adopted. That is the structural reason the arbitrage is still available.
It is not in the minutes. Per-minute resale runs $1 to $3 against a cost of $0.13 to $0.24, which is an 85% to 95% gross margin on the voice layer. Those are excellent numbers and almost entirely irrelevant to whether the business works.
The reason is that the voice layer is a small share of what the customer is actually buying, and a large share of what they will churn over. Operators who price against the minute margin discover this around month four, when a client with low volume is paying less in total than the cost of keeping them supported.
An 85% gross margin on minutes is compatible with losing money on every client. That is not a paradox, it is a fixed-cost problem.
This is the calculation that should precede every voice deployment, and it is simple enough to do on paper.
A deployment carries fixed monthly costs — the platform, the monitoring, the number, the human attention required to keep it accurate. Those costs do not scale with call volume. The per-minute margin only accumulates with volume. Below some call count, the fixed cost exceeds the margin and the client is structurally unprofitable no matter how well the agent performs.
Our own rule follows directly: if call volume is too low for per-minute economics to beat fixed cost, walk away. Not because the client is bad, but because the deployment cannot be priced in a way that survives — and an arrangement that loses money on a working deployment will eventually be ended badly by one side or the other.
The pre-qualifying question is therefore: how many inbound calls per month, and what is your average handle time? A business taking 200 calls a month and one taking 2,000 are not the same opportunity, and the difference is not 10x — it is the difference between viable and not.
Median time to first client in this category is 18 days, which sounds encouraging. Active clients at six months average 4.7. Read together, those two figures describe a business that is easy to start and hard to accumulate — consistent with the churn pattern seen across the sector.
The 66% unserved share is real, but it is not a homogeneous pool. A meaningful portion of it is unserved because the volume does not clear the fixed cost, and those buyers will never be profitable clients for anyone. Distinguishing them before the sale is most of the operator's job.
Model it in this order, and stop at the first one that fails.
One: monthly call volume times average minutes per call gives total minutes. Two: total minutes times the per-minute margin gives gross margin. Three: gross margin minus fixed monthly operating cost gives contribution. Four: contribution minus the support attention the client will actually consume gives the real number.
Step four is the one operators skip, and it is where the surprises live. A client whose agent needs weekly correction consumes hours that never appear in a per-minute model, and those hours come out of the margin computed in step two.
Ask what happens at your volume, not at the demo's volume. A vendor who cannot tell you your expected monthly total, including fixed costs, has not modelled your deployment — they have modelled a generic one and are hoping you fit.
And ask what the per-minute cost becomes if your volume drops by half, because for most businesses volume is not constant, and a pricing structure that only works at peak is fragile in a way that will surface in a quiet month.
We will model one inbound call flow against your real volume and tell you plainly whether it clears fixed cost before anyone builds anything.
Run the deployment check →One workflow, modelled end to end. Yours to keep either way.
A free deployment readiness check on one inbound call flow. We model your per-minute cost against your real volume and tell you whether the economics clear fixed cost.
Run the deployment checkOne workflow. Real arithmetic. No commitment.