Algeria AgencyArtificial intelligence
Two quotes for the same chatbot: what actually moves the price
Two suppliers price the same need at one and at five. The gap almost never comes from technology: it comes from five scope decisions.
An equipment distributor in Blida asks for three quotes for an assistant on their page and their messaging app. They get three unrelated amounts, three different timescales, and not one common line that would let them be compared. The cheapest promises the most.
The gap does not come from the model: everybody calls the same suppliers. It comes from what each of them understood the scope to be, and from what nobody wrote down. This article names the five decisions that move the price of an assistant and the two lines that appear in no quote.
It does not discuss amounts. The part that can be calculated — monthly usage — is calculated with your own figures in this site’s calculator, and the refusal to publish a rate grid is covered elsewhere, with its reasons.
The quotes do not describe the same work
Before comparing two amounts you have to check they cover the same thing, and they almost never do. A quote may cover going live with an assistant answering five questions on one page, or going live with one answering fifty questions on four channels while checking stock.
Both are called "chatbot" and both are honest. What is missing is the line saying which of the two was priced, and that line is yours to write: you decide the scope, not the supplier.
The test that settles it fits in one sentence to send to all three: "how many distinct questions does your price cover, and what happens to the fifty-first?" The answers will make three quotes comparable in ten minutes.
The five sections that follow are the five decisions to take before asking for a price. Taken, they make quotes legible; left open, they come back as a variation six weeks later.
One remark before going into them: a short quote is not a simple quote. A single line reading "assistant development" covers exactly the five decisions below, taken by somebody other than you and without your knowing which. The length of a quote is no sign of seriousness, but the number of things it names is.
First decision: how many questions, really
Scope is counted in distinct questions, not in features. Five questions often cover half a shop’s traffic: hours, address, availability, price, delay. Fifty cover almost everything and cost far more than five times as much, because each question adds an answer to write, a refusal case to define and a test to maintain.
The curve is not linear and it has a visible knee. The first questions are written once and serve thousands of times; the last are edge cases arriving three times a month that take as much work as the first.
The practical consequence is to produce the list before the quote, with each question’s estimated volume. Twenty lines are enough, and they are collected from your own messages in an hour — that is the check in section 11.
A supplier who agrees to price without that list is pricing their hypothesis, not your need. It is not dishonest; it is simply unverifiable, and it is the first source of difference between two quotes.
Second decision: where the answer comes from
An assistant can answer from three sources, and they do not carry the same go-live cost. The first is a written corpus: your pages, your product sheets, a document drafted for it. The second is your catalogue or your database, queried live. The third is a third-party system — stock, management software, an ordering platform.
The jump in price sits between the first and the second, not between the second and the third. Writing a corpus is editorial work; wiring up a live source adds an integration, authentication, a behaviour when the source fails, and a cache whose duration has to be decided.
What that changes for you is the question to ask: "which source is this answer wired to, and what does the system say when the source does not answer?" A quote that does not separate the two will probably charge for the first and deliver the first, while you were expecting the second.
Many businesses discover at this stage that half their questions can be answered from a written corpus, and that the other half does not need an assistant at all — it needs the stock figure to be right.
Third decision: how many channels
Each channel adds an integration, rules of its own and a queue to watch. The second channel does not cost half the first: it costs about as much, because what gets reused is the corpus and not the machinery.
Channels do not share constraints either, and what each messaging app imposes is a subject of its own. What matters here is budgetary: ask for a price per channel rather than a package, and start with the one your customers already write to most.
The decision that saves most is often to delay a channel rather than negotiate its price. An assistant live on one channel for two months produces the log of unanswered questions that will make the second channel half as expensive to open.
The opposite trap exists: opening a channel nobody in the business watches. A channel with nobody behind it produces an invisible queue, and it is the worst purchase on this list because it costs money and damages reputation at the same time.
Fourth decision: the languages
Understanding French only, understanding standard Arabic, understanding written darija, replying in one or the other: those are four distinct decisions and they do not cost the same.
What costs is not the language itself but the spelling variance it brings, and the corresponding work is described in detail in what darija adds to the work. What this section adds is the consequence for a quote: ask for the line to be separated, because it is the one most often underestimated.
There is a free decision and it is usually the right one: understand all three languages and reply in two. The customer does not need to be read in the register they write in, they need to be understood when they write.
If your traffic is mostly French with product references, the "languages" line in your quote should be close to zero, and a supplier pricing it heavily is describing traffic other than yours.
Fifth decision: what has to be true before answering
This is the least visible decision and the one that moves the price most. Answering "we have that model in stock" assumes the stock figure is right at the moment of the answer. Answering "your order leaves tomorrow" assumes access to order status.
Every check of that kind turns an answer into a query against a system, with everything that implies: access, latency, behaviour when unavailable, and a decision about what to say when the data is doubtful.
The question to put to a supplier is this: "which of my twenty questions need a real-time check?" On most lists the answer is three or four — and those three or four are often half the price.
The compromise that works is to accept a more modest answer in those cases: say what is published, and have a person call back. It is less impressive in a demonstration and it is what holds in service.
What is never in the quote: the revision
An assistant that goes live is tuned on real traffic, and that tuning is not after-sales service: it is a stage of the work. The first weeks produce the log of unanswered questions, and that log is the only material that says what the first version missed.
The revision is counted in days and it arrives between the second and the sixth week. A quote that does not contain it describes a delivery rather than a go-live, and the difference will be paid either as a variation or as quiet degradation.
So ask for three separate lines: go-live, revision after real use, and ongoing upkeep. A supplier refusing to separate them makes their quote incomparable, which is sometimes the intention.
The third line is the first to be cut when a budget tightens, and it is the one to keep. An assistant with no upkeep degrades at the rate your catalogue and your vocabulary change, which is to say quickly.
If a supplier still refuses to separate them, one door remains: ask for the revision as a fixed-price option, with a date. An option priced in advance can be set against a competitor’s and triggered when you decide, whereas a variation is negotiated once the system is already live and your position is at its weakest.
Nor is this: the human queue
An assistant that refuses to answer — and it must refuse — produces a queue of conversations for a person to take up. That queue is work, it has a volume, and it appears on no line of any quote.
Its volume depends directly on the five decisions above: a narrow scope produces a bigger queue, a wide scope produces a smaller queue and a higher price. That is the real trade-off here, and it is posed in staff hours as much as in dinars.
The useful question before signing is therefore: "with this scope, how many conversations a day will reach us, and who takes them?" A serious supplier can answer to an order of magnitude; one who answers "none" is describing an assistant that refuses nothing.
The cost of that queue is also what makes the comparison with a salary honest. The calculator gives monthly usage; the queue gives what is left to do, and the two together give the only figure that decides.
Comparing two quotes that describe different work
Once the five decisions are taken the comparison becomes possible and it fits on one sheet: for each quote, the number of questions covered, the sources wired up, the channels, the languages, the real-time checks, and whether the three lines from section 7 are present.
That table reveals a regular pattern: the cheapest quote covers the least, and the most expensive often covers the same as the middle one. The price gap between the extremes is almost always found in two cells, not in general quality.
It reveals something else, more awkward: two quotes can cover exactly the same scope and differ by the presence of the revision. The one that includes it looks dearer and costs less, because the other will charge it as a variation at the moment you can no longer arbitrate.
That table is the only deliverable we ask a prospect to produce themselves before talking to us. It turns a conversation about technology into a conversation about their own scope, and it survives the choice of supplier.
The cap, usage, and the part that can be calculated
Everything above concerns the go-live price. Monthly usage is another matter: it is calculated, from your volumes and your supplier’s public rates, and the calculator does that arithmetic without any price being published here.
What to know before opening it is that usage depends on the scope decided above: more questions wired to live sources means longer conversations, therefore more tokens. The two halves of the budget are not independent.
The ratio between them often surprises. On a moderate-volume customer-service assistant, go-live weighs more than usage in the first year and the reverse from the second. A quote that talks only about usage describes the second year; one that talks only about go-live describes the first.
So do the calculation over twelve months, and redo it over twenty-four. The two horizons do not rank the proposals in the same order, and that is often the only thing that changes a choice.
The check: twenty questions, one hour
Take a hundred real inbound messages and write the list of distinct questions they contain. You will rarely find more than twenty, and often fewer than fifteen. Note beside each how many times it appears.
Then mark three things: those whose answer is already written somewhere, those needing a real-time check, and those no rule can settle. Those three columns are your scope, your integration line and your human queue.
The reading is direct. If the first column dominates, your project is editorial work and it will cost little. If the second dominates, your project is integration work and has to be priced as such. If the third dominates, there is no project: there is a team to reinforce.
That sheet is also what you send to the three suppliers. It makes their quotes comparable, and it protects you from the one gap no negotiation recovers: two prices that do not cover the same work.
What we do, and what we refuse
We make the section 11 sheet with you, take the scope and the three quote lines from it, and price the revision separately from go-live. We leave you the list of questions and their volumes, in a format usable to consult another supplier.
We refuse to price without that list. A price given on a description of a need is the price of our hypothesis, and it will be revised at first contact with your real messages — which puts the disagreement at the worst moment rather than the best.
We publish no range, here or anywhere, and no ratio between our own tiers either. "The top tier is four times the entry one" is a rate card with the units removed, and it ages exactly like a rate card.
What you can do without us is the most useful part: the sheet, and the order of the channels. Many businesses find there that one channel and twelve questions cover most of their traffic, and the list of what an assistant must refuse will spare them the rest.
Frequently asked questions
Why give no range at all, even a wide one?
Because a wide range is wrong in both directions: it reassures when it is low and puts people off when it is high, without saying which of the five decisions it corresponds to. The sheet in section 11 gives you a scope, which is what a price needs in order to mean anything.
Is the cheapest quote necessarily inadequate?
No, and often the opposite: it covers the narrow scope that handles half the traffic. The real risk is not its amount, it is comparing it with a quote covering something else without the difference being written down.
Should the revision after a few weeks be paid for?
Yes, and it is better in the quote. A supplier who gives it away will do it quickly; one who charges it as a variation will do it when you have no choice left. What matters is that it is a planned line, whatever its amount.
How many questions have to be covered for this to pay?
The question runs the other way: cover the ones that recur most first, and stop when the next appears less than once a day. The threshold is your volume rather than a general number, and it is read off the sheet in section 11.
Can we start small and extend later?
That is the order we advise, on one condition: that the corpus and the log of unanswered questions belong to you and are exportable. Without that, extending means starting again with the same supplier, which is no longer an extension but a dependency.
Does the price depend on which model is used?
Far less than on the scope. The choice of model changes monthly usage, which is calculated, and barely changes the go-live work — which is what the five decisions in this article make vary.
Where we come in
Your twenty questions, with a volume beside each, beat a specification: they make three quotes comparable in ten minutes.
- We collect the list with you from a hundred real messages, and it stays with you, exportable.
- The quote comes out as three separate lines — go-live, revision, upkeep — so it can be set against another.
- Questions needing a real-time check are marked apart, because those are the ones that make the price.
Until that list exists we hand over no amount: pricing a description would be selling you our supposition, revised on the day you can no longer arbitrate.
Read next
A chatbot that understands darija: what it adds to the work
Darija is not expensive because it is difficult. It is expensive because it is never written the same way twice, and that is settled on your own messages.Chatbot: what it must refuse to answer
An assistant is judged on its refusals, not its answers. What the list has to contain, and what happens when nobody wrote it.Which messaging channel, and what each one imposes
The channel decides the rules, the deadlines and what you are allowed to send. What to know before choosing where to reply.
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.