Most comparisons of voice agents end up in the same spreadsheet: three vendors in three columns, a row for each capability, and a price at the bottom. The column with the lowest number wins.
The spreadsheet is usually useless, and it is useless for a reason that has nothing to do with the numbers. The prices are not in the same unit. One vendor sells minutes, one sells resolved calls, one sells seats, one sells a flat monthly fee. Putting them side by side and reading the bottom line is choosing between a price per kilo and a price per litre.
The unit is not a detail in the contract. It is the contract. It decides who carries the risk when volume moves, what the vendor makes money on the system doing, and what happens to the invoice the day the deployment starts to work.
When it works, the bill grows
Twilio's second-quarter figures are a useful illustration. The company reported 1.5 billion dollars in revenue, up 22 percent year on year. Voice revenue grew more than 20 percent. Self-service voice – the customers who set something up themselves without going through a sales meeting – grew more than 50 percent, which the company itself describes as an indicator of experimentation among smaller businesses.
In the same review, Twilio says two other things. CEO Khozema Shipchandler describes the market as being in "very early innings" and stresses that the great majority of customer interactions are still handled by humans. And the company points out that rising carrier pass-through fees are putting pressure on its smallest customers, because costs are becoming harder to forecast.
Those are not three pieces of news. They are one piece of news seen from three angles. In a market where price follows usage, "it works" and "the bill is growing" are the same event. Twilio mentions a customer that went from low six-figure quarterly spending to a six-million-dollar annual run rate. That is a success story in the vendor's quarterly deck and a budget line at the customer.
Neither of those is unreasonable. The point is that the relationship between them is rarely settled during procurement, because the unit being traded in is barely discussed at all.
Four units, four distributions of risk
| Unit | The vendor profits from | Volume risk sits with | Where it breaks down |
|---|---|---|---|
| Per minute | Calls lasting longer | The buyer | A slow agent costs more than a fast one |
| Per resolved call | Calls being completed | The vendor | Someone has to define "resolved" |
| Per seat or user | More people getting access | The buyer | Measures the headcount the agent removes |
| Flat monthly fee | You using less than you pay for | The vendor | The risk premium is paid in quiet months too |
The left-hand column is the only one most quotes state. The right-hand column is the only one that shows up in production.
The minute pays for time, not for outcome
A concrete calculation makes it clear. A call that takes sixty seconds today, and an agent that spends twenty seconds extra confirming, repeating and searching: that is a third more cost for exactly the same task. Nothing in the contract gives the vendor a reason to want those twenty seconds gone.
Worse, a call that resolves nothing is billed on the same terms as one that does. The customer who calls again the next day is billed again. A unit that measures duration cannot, in principle, distinguish between a call that was completed and one that merely ended.
It is the same weakness that makes answered-call rates look so good in reports: both count activity. The number that actually says whether something was resolved is how many people call back about the same thing within a day – and that is not a number any vendor invoices against.
The counterargument is stronger than it sounds
The obvious conclusion is that you should pay per resolved call instead. That is too quick, for four reasons.
First, someone has to define "resolved". In practice that someone is the vendor, in an annex, in wording the buyer reads once. Whoever defines the unit controls the invoice. An agent that signs off with "I've logged your request" can be resolved by the definition and unresolved by any reasonable standard.
Second, the minute is the only unit the buyer can verify without trusting the vendor. A call has a start time and an end time, and both sit in a telephony log held by someone other than the party issuing the invoice. None of the more "fair" units has that property. Paying per resolved call means accepting the vendor's own count of what it managed to do.
Third, a flat monthly fee moves the risk to the vendor, who prices it in. You pay the risk premium in the quiet months as well, and a flat fee with a fair-use clause is a per-minute deal on a delay. It is a real advantage when volume is unstable, but it is bought.
Fourth, outcome-based units also penalise the handover that ought to happen. An agent paid for finishing has a built-in reason not to transfer. The route out to a human is the part of the setup that tolerates a bad incentive least, because it is rare, expensive to redo, and invisible in every number except the customer's.
So the minute is not the wrong unit. It is an incomplete one, and the difference matters for what you do about it. An incomplete unit gets completed in the contract, with a cap, an alert threshold, or an agreed review at a given volume. A wrong unit has to be replaced.
The cost that isn't software
One more thing separates voice agents from the rest of the software market, and explains why the prices do not behave like software prices.
Underneath the agent sits a carrier minute. It is a regulated, metered, invoiced commodity bought from an operator, and that part of the cost does not fall because the models get better. Twilio's comment about rising pass-through fees is about exactly this layer – and it came from the company with the strongest negotiating position in the entire market.
The practical consequence is a question that is easy to ask: how much of the price is telephony, and how much is the agent? Only one of those halves has a historical tendency to get cheaper. A quote that cannot answer the split does not control it either, and that is worth knowing before signing a three-year term.
Four quotes in four units, converted to one number
The comparison can be made. It needs three figures, and all three have to come from the buyer's own operation.
How many calls of the relevant type arrive per month. How long they take today. What share of them come back within a day.
With those three, all four quotes can be rewritten as cost per resolved call, and only then do they stand in the same unit. The calculation is rough and the answer will be wrong. But it will be wrong in the same direction for all four, and that is all it takes to rank them.
Those three numbers cannot be estimated in a meeting, either. They have to be measured, and measured before the agent is switched on. That is the part of a pilot almost everyone skips, and it is why almost no pilot can be used to negotiate on price afterwards. Without a baseline the buyer has nothing to hold the vendor's figures against, and the negotiation collapses into an argument about the unit price alone – the number that matters least.
The unit is a build instruction
Finally, this is not only a procurement question. The pricing unit is also an instruction to the vendor's product organisation, and it works slowly and reliably.
A vendor paid per minute builds for coverage: more call types, more languages, more time spent in the agent. A vendor paid per resolved call builds for completion – and for a narrow definition of completion, since that definition is the revenue. A vendor on a flat fee builds for predictability, and predictability in practice means saying no to the difficult call types.
None of this is bad faith. It is simply that over two years the incentive beats the intention, and it does so without anyone deciding that it should.
Threll.ai sells per minute. The entry tier is 699 NOK a month at 4 NOK a minute with no included minutes; the tier above is 2,499 NOK with 500 minutes included. So we sell in the unit this piece has just taken apart, and the reason is worth saying out loud: the minute is the only thing a customer can count themselves from day one. That does not make it the right unit. It makes it an auditable one, and the two are not the same.
Either way the question to any vendor, ourselves included, is the same. Do you make more money when the agent gets better, or when it stays on the line longer?
