Algeria AgencyArtificial intelligence
Your first AI invoice: the lines that were not in the calculation
The estimate said one amount and the invoice says another. The gap is almost never an error: it is six lines nobody counted.
A services company in Algiers estimated its assistant before go-live, seriously, using its supplier’s public rates. The first invoice arrives on the 3rd of the following month and it is higher. Nobody can say by how much exactly, because the two documents do not count in the same unit.
This is the most instructive moment of an artificial intelligence project, and it is the one nobody recounts. The gap is almost never a supplier error: it comes from lines that existed and that the estimate had no reason to know about.
This article names them one by one and gives the twenty-minute reconciliation that makes them visible. It completes the calculator, which predicts, and the article on the payment route, which describes how the money leaves; this one describes what comes back on paper.
Two documents that count different things
Your estimate counts conversations. The invoice counts technical units: tokens, seconds of audio, characters, interface calls. Between the two there is a conversion you did not make and the supplier does not show.
So the first thing to do is not to dispute, it is to convert. Divide the invoiced total by your real number of conversations for the month: you get a cost per conversation, which is the only quantity comparable with your estimate.
That figure is often a revelation, in both directions. It is sometimes lower than the estimate because real conversations are shorter than imagined ones; it is sometimes far higher, for the reasons in the sections below.
Keep it, because it becomes your reference. From the second month it is no longer the invoice to look at but that cost per conversation: it is what says whether something changed, whereas a monthly total conflates a price change with a volume change.
One precaution about that first calculation: count conversations the way your system counts them, not the way your team describes them. Is a conversation resumed the next day by the same customer one or two? Does a message with no reply count? Both answers are defensible, neither is better, and the only thing that matters is keeping the same one from quarter to quarter. A denominator that changes quietly produces a variation nobody will be able to explain six months later.
Failed attempts are paid for
A call to a service that does not succeed is often billed anyway. A request cut off by a timeout consumed compute; an answer rejected for not matching the requested format was produced before being thrown away; a retry after failure is a second request.
Those three cases are invisible in an estimate made by multiplying conversations by a unit cost, and they weigh proportionally more in the first month: that is the month when settings are still rough and the system retries a lot.
What that imposes is one question to ask before signing, and only one: "are failed attempts billed?" The answer varies by supplier and by kind of failure, and it decides your first month’s gap more than any rate does.
The good news is that this line falls over time if somebody watches it. If it does not fall, it is not saying the supplier is overcharging: it is saying your system retries in a loop on cases it cannot handle, which is useful information well beyond the invoice.
Prepaid credit that expires
Many suppliers sell credit to be consumed rather than a subscription, and that credit often has an expiry date. The consequence is direct: a quiet month does not save you money, it loses you money.
The mechanism is perfectly legal and it is written in the terms, but it does not show on an invoice — expired credit produces no line, it produces an absence. It is the only item in this article you have to go looking for rather than read.
It has a perverse side effect worth knowing: it pushes you to consume before the deadline, that is, to run processing nobody needs so as not to lose the credit. We have seen it happen, and it is the only time a system has cost money for a purely accounting reason.
The sensible position is to choose pay-as-you-go where it exists, even at a slightly higher unit price, as long as your volumes are not settled. Prepaid becomes rational once you know your cost per conversation, which is to say after three months.
What the bank takes on top of the rate
The calculation you made multiplied an amount in currency by a rate. The line that arrives in dinars contains more: the exchange commission, the fixed charges of the operation, the domiciliation tax — 5 % on transferable royalties under article 123 of the 2025 finance act — and sometimes correspondent charges.
None of that is hidden, and all of it is absent from an estimate made with a rate taken off a financial site. That is why this site’s calculator asks for the rate your bank actually applied rather than a market rate: it already includes part of that gap.
There is a threshold effect that matters at small volumes: the fixed charges of an operation do not depend on the amount, so paying a small amount monthly costs proportionally far more than paying triple quarterly. That is a trade-off to settle with your bank, not with your supplier.
This paragraph describes the state of a regulation at the date of publication and does not replace professional advice. What to take from it is operational: the "conversion" line of your reconciliation has to contain everything the bank took, not only the rate difference.
The line that appears in the second month
Some lines do not appear on the first invoice because they only exist once the system has accumulated something. Storing conversations is the classic example: free while it is empty, billed as it grows.
The same holds for an internal search index, for logging if it is kept at the supplier, and for any form of system memory. Those items are small at first, growing after, and they have no reason to shrink.
The practical consequence is that three invoices have to be read before concluding, not one. The first shows usage, the second reveals accumulation, the third gives the slope. A decision made on the first invoice is made on a third of the information.
It is also what makes a retention policy decided up front useful: how long do you keep conversations, and why. The answer has legal consequences covered elsewhere, and a financial consequence nobody mentions.
Checking the count against your own logs
A usage invoice is verifiable, and almost nobody verifies it. Your system knows how many requests it sent; the supplier says how many it counted. Comparing the two numbers is twenty minutes of work in the first month and five thereafter.
The gap, if there is one, is instructive both ways. A gap against you may come from calls you did not know you were making: a test page left connected, a scheduled job still running, an integration retrying silently. A gap in your favour usually means your log is incomplete.
For that check to be possible, the system has to write a log of its calls from day one, with a counter per type. It is a line of work to ask the supplier for, and it costs an hour if asked before go-live and a week after.
We have never seen a supplier deliberately overcharge. We have several times seen a business paying for calls it did not know it was making, and that is exactly what this reconciliation reveals.
The day the supplier changes model
Catalogues move: a model is withdrawn, another replaces it, prices are revised. What happens then does not read as a price rise, because the new model does not consume like the old one — it can be cheaper per unit and produce longer answers, and so cost more.
That is why the cost per conversation from section 1 is more useful than the advertised rate. It absorbs both effects at once and it is the only figure that answers the question you care about: will next month cost more.
A model change also has a non-financial effect to check at the same moment: the answers change. A system tuned on one model gives different results on the next, and the test set has to be replayed before the drift is observed on customers.
The useful contract clause is therefore not a guaranteed price — nobody guarantees one — but notice. Knowing a month in advance that a model will be withdrawn is worth more than a discount, because it leaves time to redo the trials.
What is on no invoice at all
Two real items will never appear in these documents. The first is the human time of the exception queue: the conversations the system hands over and a person takes up. It is paid in salaries and it varies exactly inversely to the system’s quality.
The second is monitoring time. Somebody has to read the unanswered questions, check the log, reconcile the invoice. It is little — an hour a month is enough — but it is an hour that exists in no budget and gets done by nobody if it is not assigned.
Adding those two to the invoiced total gives the only cost that can be compared with anything else. It is also what makes a comparison with a person honest, and that comparison is covered in detail in the anatomy of a call’s cost.
A business that cannot name anybody for the monthly monitoring hour should know that before signing rather than after. It is the smallest item in this whole file and the one whose absence degrades things fastest.
The reconciliation, once a quarter
The first month takes twenty minutes; after that, once a quarter is enough. The reconciliation fits in five lines: the total invoiced in currency, everything the bank took, the real number of conversations, the resulting cost per conversation, and the gap against the previous quarter.
Those five lines live in a spreadsheet and they tell the system’s story better than any dashboard. A rise in cost per conversation at constant volume is a signal; a rise in the total at constant cost per conversation is not one, it is growth.
The quarter is the right period because it absorbs atypical months — holidays, the start of term, an incident — without letting a drift settle in. A monthly review is noise; an annual one arrives after the drift.
That sheet is also what to produce the day somebody asks whether the system "costs a lot". The question has no answer in the abstract; it has one with four quarters of cost per conversation.
When the gap says to stop
There is a case where the reconciliation leads not to a setting but to a decision: when the cost per conversation does not fall in the second and third quarters while volume is rising.
A system that works becomes cheaper per conversation as it runs, for mechanical reasons: settings sharpen, failed attempts fall, answers shorten. If that does not happen, something is resisting — a badly chosen scope, a slow source, a model ill-suited to your texts.
That finding is more honest than general dissatisfaction, because it is quantified and it points at a cause. It justifies either tuning work or stopping the service, and both beat continuing to pay.
We advise setting that threshold in advance, in writing, before go-live. A business that decided at the start what it would do if cost per conversation did not fall takes that decision coolly; one that discovers it after a year takes it having already spent.
The check: one invoice, three columns
Take your last invoice and fill in three columns: the item as the supplier writes it, the unit it bills, and what you thought you were paying for that item. The third column will be empty on two or three lines, and those are exactly the lines in this article.
Then add the same month’s bank charges — commission, fees, tax — on a separate line. Many businesses discover at that point that half their gap against the estimate comes from the bank and not from the supplier.
Finally, divide the resulting total by your real number of conversations. That figure is your cost per conversation, it is comparable with your initial estimate, and it is the only quantity to follow over time.
The sheet takes twenty minutes and it beats a negotiation. A supplier will not lower their rate; two of the six lines in this article, on the other hand, are corrected at your end within the week.
Do it the first time with whoever pays, not only with whoever runs the system. The two read the same invoice differently: one sees a technical item and the other sees cash leaving with its bank charges, and the gap between those readings is precisely what this article tries to make visible. An hour spent together on the first invoice avoids a year of misunderstanding about what the system costs.
What we do, and what we refuse
We wire up the call log before go-live, do the first reconciliation with you, and leave the five-line sheet in a format that fills itself in each quarter without us. We also put two questions to the supplier in writing: are failed attempts billed, and what notice applies to a model withdrawal.
We refuse to place ourselves between you and your supplier. The account is yours, the invoice arrives with you, and a provider who interposes on that line makes you dependent on an intermediary to read what you are paying.
We announce no typical gap between estimate and invoice. It depends on six lines of which three are peculiar to you, and a general percentage would give the illusion of a forecast where a reconciliation is required.
What you can do without us is the sheet in section 11, from the very first invoice. Many businesses find there that half their gap comes from their bank, and the payment route will help them more than a change of supplier.
Frequently asked questions
Should an invoice higher than the estimate be disputed?
Almost never before doing the reconciliation. In the vast majority of cases the gap is explained by real lines — failed attempts, bank charges, storage — and an ill-founded dispute costs time and a relationship. Check first, dispute afterwards if an unexplained gap remains.
How many invoices before judging?
Three. The first shows usage, the second reveals the items that grow with accumulation, the third gives the slope. Deciding on one invoice is deciding on a third of the information.
Is prepaid cheaper?
Per unit, often yes; in practice, not while your volumes move. Credit that expires turns a quiet month into a straight loss, and it pushes you to consume so as not to lose it. Move to prepaid once you know your cost per conversation, which is to say after a quarter.
How do we know whether we are paying for calls we do not make?
By comparing your log’s counter with the supplier’s. It is the only check on this list that needs preparation: the log has to exist from day one, with a counter per call type. Asked for before go-live, it costs an hour.
What should be done when a supplier withdraws a model?
Replay your test set before switching, and recompute cost per conversation afterwards. A new model that is cheaper per unit can cost more in total if it answers at greater length, and it can also answer differently on cases you had already settled.
Do these lines exist with a local provider too?
Some disappear — the conversion, the domiciliation tax, the bank charges — and the rest remain, because they come from the model supplier sitting behind. A local intermediary does not remove those costs, it folds them into its price, which makes the reconciliation impossible at your end.
Where we come in
The total divided by your real conversations is the only quantity that compares with your estimate, and it takes twenty minutes to work out.
- The call log is wired up before go-live, with a counter per type, because afterwards it costs a week.
- We put the two written questions to the supplier that decide: whether failed attempts are billed, and the notice on a model withdrawal.
- The quarterly five-line sheet stays with you and is filled in without us.
The subscription stays in your company’s name, with no screen between it and the foreign provider: an intermediary position would leave your own spending opaque until we explained it.
Read next
Paying an AI supplier from Algeria
The project is scoped, the team is ready, and the card is declined. That is not a fault: it is the service-import regime.Integrating CIB and Edahabia online payment
Cards are no longer the constraint: there are 21.9 million of them. What is missing is merchants who accept them.Online payment once it is live: reconciliation, refunds, disputes
Getting accredited takes a few weeks. Running the payments every day is the work nobody describes to you before the contract is signed.
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.