Algeria AgencyArtificial intelligence
The cost of a call, minute by minute
Four meters run during a voice call and none of them counts the same thing. Confusing them gives an estimate wrong by a factor, not by a margin.
A service centre in Oran wants a voice agent for its follow-up calls. The supplier announces a price per minute, the accountant multiplies by the month’s call count, and the result bears no relation to the invoice that will arrive.
The reason is not dishonesty: it is that there is no such thing as a price per minute. A call handled by a voice artificial intelligence system runs four distinct meters, three abroad and one here, and none of them measures the same quantity.
This article takes a three-minute call apart, meter by meter, and names the minutes you pay for while no conversation happens. It gives no rate: the arithmetic runs on your figures in the calculator, and the decision to have a voice agent at all is covered elsewhere.
A three-minute call, taken apart
Take an ordinary call: the customer asks where their order is, the agent understands, checks, answers, and hangs up. Three minutes on the telephone meter. Four different invoices are being written during those three minutes.
The first is transcription, which bills audio. The second is the model, which bills text read and written. The third is speech synthesis, which bills what it speaks. The fourth is the telephone line, which bills the duration of the connection.
Those four meters neither start nor stop at the same moments, and above all they do not move in proportion to one another. A call where the customer talks a lot and the agent little is expensive in transcription and cheap in synthesis; a call where the agent recites a long confirmation does the reverse.
The practical consequence is that a price "per minute" can only be a retrospective average, and it is valid only for the shape of calls it was computed on. Yours will be different, and the sections below say how.
Transcription bills audio, not speech
The first meter measures minutes of audio sent, and it is a distinction that costs money. A second of silence, a second of background noise and a second of speech are billed identically.
That changes two things. The first is that line quality matters financially as much as technically: a call dragging because of a poor connection costs more in transcription than a clean call with the same content. The second is that end-of-speech detection — the moment the system decides the customer has finished — is not only a comfort setting, it is a line on the invoice.
A setting that is too patient leaves the meter running after every sentence; one that is too eager cuts people off and lengthens the call through the repetitions that follow. Both cost, in opposite directions, and it is one of the rare settings where the financial optimum and the experience optimum coincide.
What that means for a quote: ask whether billing is on audio sent or on call duration. Both exist, the gap between them is real, and nobody mentions it unprompted.
There is a case where the difference becomes stark and it is common here: the call placed from an area with poor coverage, with dropouts and repetitions. On duration billing it costs what the connection lasts; on audio billing it costs more, because noise and blanks go to the meter like everything else. A sales fleet calling from the road therefore does not get the same invoice as a service receiving calls from homes, at identical volume.
The model does not bill the answer, it bills the re-read context
The second meter is the one everybody underestimates, for a precise reason: it is imagined to bill what the agent says, when it mostly bills what is handed back to it to read at every turn.
A voice exchange runs in turns. At each turn the system returns the conversation history, the instructions, and often the information fetched in between. A ten-turn call therefore does not cost ten times one turn: it costs the sum of histories that grow, which is a curve rather than a line.
That is why a call twice as long costs markedly more than twice as much on this meter. It is also why limiting the length of the instructions — the text describing to the agent what it should do — has a financial effect on every turn of every call.
The setting with the greatest effect here is the one nobody discusses in a demonstration: what is sent back to the model at each turn, and from when is the history summarised rather than re-read? A supplier with no answer to that question has not yet met an invoice.
Synthesis: by character or by minute
The third meter measures what the agent speaks, and it is billed two ways depending on the supplier: by character of text submitted, or by minute of audio produced. The difference looks trivial and is not.
By the minute, an agent that speaks slowly costs more for the same content. By the character, speaking rate is free and sentence length is what counts. The two models push towards opposite choices in how answers are written, and it is better to know which you are paying before writing them.
There is a local effect worth knowing: good voices in Arabic and in Algerian French are fewer and often billed higher than English ones. It is not a general rule, it is a catalogue fact to check at selection time, because it moves the line by tens of per cent.
The practical conclusion is the same as for the previous two meters: the question for the supplier is the billing unit, not the amount. The unit tells you how your invoice will move when the shape of your traffic changes.
Telephony is the only line in dinars
The fourth meter is local and billed here, by an Algerian operator, in dinars. It goes through neither the conversion nor the domiciliation tax, and confusing it with the other three is the commonest error in an estimate done by hand.
That line has another peculiarity: it does not depend on the technology at all. A voice agent, a classic switchboard and a person picking up pay the same minute to the same operator. It is therefore the only line in this article that was already in your budget before the project.
To it must be added what is not a minute: the number itself, its subscription, and sometimes concurrent-call capacity. Those amounts are fixed, they fall even in a month with no calls, and they make up the floor described in section 9.
The separation between this meter and the other three is structural to the point that this site’s calculator enforces it: the part billed here is entered separately, precisely because it is not converted.
The minutes you pay for with no conversation
This is the section we would have wanted to read before our first invoice. Some of the billed minutes correspond to no conversation at all, and they are invisible in an estimate made by multiplying calls by a duration.
There is the ringing before pick-up, depending on how the line is configured. There is the call where the customer hangs up during the greeting, having consumed synthesis and telephony for nothing. There is the wait while the system queries a slow source, where the audio meter keeps running. There is the silence at the end before the hang-up is detected.
Added together, those minutes are not a rounding error, and the share is larger on a follow-up service than on appointment booking — because a customer calling to complain hangs up mid-sentence more often.
The setting that reduces them is dull and effective: a short greeting, fast hang-up detection, and a holding response when a source takes more than two seconds. None improves the system’s intelligence; all three improve the invoice and the experience at once.
The cost of a call is not the cost of a useful call
Not all calls produce a result. Some are wrong numbers, cold calls, customers who wanted something else, or conversations the agent transfers to a person after thirty seconds.
Those calls are billed like the others, and they must not disappear from the calculation. The figure that decides is not the average cost of a call: it is the total cost divided by the number of calls that led to something — an appointment made, information given, a problem solved.
The ratio between the two is your real indicator, and it is often harsher than expected. A system that handles half the calls well and transfers the other half mechanically doubles the cost of every useful call, with nothing looking abnormal on the invoice.
It is also what makes the comparison with a person honest. A person also handles useless calls, and quantities of the same nature have to be compared: cost per completed call on one side, cost per completed call on the other.
What doubles the invoice with no change in volume
Call volume is what everybody watches and it is not the most dangerous variable. Average duration is, because it acts on all four meters at once and acts more than proportionally on the model’s.
Three things lengthen a call without anybody deciding it. An answer that is not understood first time and gets repeated. A slow check that leaves a gap. And an over-talkative agent, whose every answer carries two sentences of courtesy before the information.
The third is the easiest to correct and the most often neglected. Shortening answers by two sentences on a service handling a few hundred calls a month shows on the next month’s invoice, and improves the conversation at the same time.
The consequence for monitoring is that median call duration has to be watched, not call count. A ten per cent rise in volume is expected; a ten per cent rise in average duration is noticed by nobody before the invoice.
The floor: what is paid at zero calls
A voice service has a fixed cost that falls even in a month with no traffic. It holds the number and its subscription, concurrent-call capacity if your operator bills it, and sometimes a minimum commitment with the voice or transcription supplier.
That floor is what makes a voice agent unattractive below a certain volume, and the threshold is higher than people imagine. Below it, the cost per completed call is dominated by lines that do not depend on calls, and a person picking up costs less.
There is one particular case worth naming because it recurs: the seasonal service. A school, a ticketing service or a holiday-trade shop pays that floor twelve months for a use concentrated in two, which multiplies its real cost per call.
The question to ask before signing is therefore: "what is paid in a month where the service takes no calls?" A serious supplier answers in three lines; one who answers "nothing" has forgotten the telephone line.
Comparing with a person: the minute is the wrong unit
The comparison a management wants to make is this one, and it is made badly when posed in minutes. A person on the telephone does not cost by the minute: they cost a month, whether they pick up forty times or two hundred.
The right unit is therefore cost per completed call, over a full month, on both sides. It makes visible what the minute hides: a person is expensive at low volume and becomes economical at high volume, while a system does the reverse because of the floor in the previous section.
One difference remains that this calculation does not capture and that has to be posed separately: the person handles the calls the system transfers, and the system does not handle the calls the person takes outside its scope. The two do not do the same work, and the comparison only means something on the common part.
Our position on this is constant: the calculation serves to rule out projects that make no sense, not to decide the ones that do. A favourable ratio is a necessary condition, never a sufficient reason.
The check: ten calls, a stopwatch
Listen to ten recordings of real calls, stopwatch in hand, and note four things per call: total duration, the time the customer speaks, the time your agent speaks, and whether the call led to anything.
Those four columns are enough to feed the calculator with figures that are yours. Total duration gives telephony, the customer’s speaking time gives transcription, the agent’s gives synthesis, and the number of turns gives the order of magnitude for the model.
The "completed" column is the one that decides. If fewer than half the calls complete, your cost per useful call is more than double your cost per call, and that is the figure to set against a person.
Ten calls are enough because the object is an order of magnitude rather than a statistic. Take them at random from an ordinary week, not the week your service was running a promotion.
What we do, and what we refuse
We take ten of your calls apart with you, draw the four quantities from them, and carry them into the calculator with your suppliers’ public rates and your bank’s real rate. We then set end-of-speech detection, instruction length and the greeting, which are the three invoice levers that also improve the conversation.
We refuse to announce a price per minute. There is none: there is a retrospective average over a given traffic, and presenting it as a rate would be leaving you to multiply one number by another that does not correspond to it.
We also refuse to compare a cost per call with a salary. The two do not reduce to the same unit without going through the completion rate, and that comparison is the one that makes people buy systems costing more than what they replace.
What you can do without us is the sheet in section 11 and the greeting setting. Many services find there that their difficulty is the completion rate, and the weekly listening session answers it better than a change of supplier.
Frequently asked questions
Why give no average price per call?
Because an average is the result of a traffic shape, not a property of the service. A follow-up call, an appointment booking and a complaint have very different durations and speech ratios, and the average of one predicts nothing about the other. The ten-call sheet gives you yours in an hour.
Does line quality really affect cost?
Yes, by two routes: a degraded line lengthens the call, and the transcription meter bills audio, including noise and silence. It is the only cost line in this article that improves by changing operator rather than by changing artificial-intelligence supplier.
Does a voice agent cost more than a written chatbot?
Per conversation, almost always, because it runs four meters instead of one and adds a local line. The useful question is not which is cheaper but which of the two channels your customers actually use — and on that, a service receiving mostly calls has no choice.
What happens if volume doubles?
The three foreign meters double, telephony doubles, and the floor does not move — so cost per call falls slightly. That is the only direction in which these systems have an economy of scale, and it is weaker than the one you will be shown.
Can spending be capped?
Yes, and it has to be arranged before go-live: a monthly threshold, an alert before it is reached, and a decided behaviour beyond it — switching to taking a message rather than an invoice that keeps running. It is a clause, not a feature.
Should we pick a supplier who bills everything together?
It is easier to read and it costs you the ability to compare. A single invoice hides the four meters, so it also hides which one moved when the amount changes — and that is the information that lets you correct anything. If you take a bundled offer, at least ask for a per-item breakdown on the monthly invoice.
Where we come in
Your ten timed calls give the four quantities, and the "completed" column is the one that changes the conclusion.
- Your ten recordings are broken down together, and your measured durations go straight into the tool alongside the rate your bank actually applied.
- The three settings that weigh — end of speech, instruction length, the greeting — are set in front of you.
- The monthly floor is written separately, because it falls in the months with no calls too.
No per-minute rate will come out of here: that number does not exist, and letting you multiply by it would be selling you an average taken from somebody else’s traffic.
Read next
Measuring darija transcription: the protocol, and the figure we do not publish
Nobody publishes an error rate for Algerian darija, us included. Here is the protocol that would produce one, and what to know before reading it.The words transcription does not know
The voice agent understands the sentence and gets the name wrong. That is the opposite of what people fear, and it is what makes an appointment unusable.Voice agent: the thirty-second test
A voice agent does not replace your switchboard. It takes the calls nobody takes — and only the ones that fit in thirty seconds.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.