Buying AI

How Much Does an AI Chatbot Cost in Saudi Arabia?

Quotations for the same brief routinely differ by an order of magnitude, and the reason is almost never margin. It is that the word "chatbot" covers four different products. Here is how the cost is actually built.

15 min read July 30, 2026

Ask three vendors in the Kingdom to quote for "an AI chatbot for our website" and you will get three numbers that do not resemble each other. It is tempting to read that as some vendors being expensive and others being competitive. Almost always, it means the three of them have quoted for three different products, and the brief was not specific enough to tell them apart.

This article is about how the cost is assembled rather than what any particular vendor charges. If you understand the components, you can read any quotation, see which parts have been included, which have been quietly left out, and which will arrive later as a change request. That is worth more than a price list.

The four products people call a chatbot

Before any discussion of money, settle which of these you are actually buying. The cost difference between the first and the last is not a percentage. It is a multiple, and it is justified.

Four things sold under one word

1Scripted / decision-tree botButtons and fixed replies. No language understanding at all — it matches a keyword or the customer presses an option. Cheap, predictable, and unable to answer anything nobody anticipated.
2FAQ bot with language matchingUnderstands phrasing rather than exact keywords, and maps a question to one of a prepared list of answers. Better, but the answer set is still finite and hand-written.
3RAG assistantReads your actual documents at the moment of the question and answers from what it found, with a citation. New material means uploading a document, not rewriting the bot. This is what most buyers picture when they say "AI chatbot".
4AI agentEverything above, plus the ability to do things — check an order in your system, create a ticket, book a slot. Costs more because it touches your systems and therefore inherits your security and approval requirements.

The honest question to ask internally is which tier your volume actually needs. A large share of contact is answerable at tier three.

Each tier includes the one before it. Most disappointing deployments are a tier-one purchase made against a tier-three expectation.

One-time build versus monthly subscription

Two commercial models dominate the market here, and they suit different buyers.

A subscription — sometimes described as SaaS — bundles the platform, hosting and model usage into a recurring fee, usually banded by conversation volume or by number of agents. It starts cheap and it never stops. Configuration is typically yours to do, and the ceiling on customisation is whatever the platform allows.

A one-time build charges for the implementation itself: the discovery, the knowledge base, the integrations, the Arabic tuning, the testing. It costs more on day one and leaves you owning something specific to your organisation. It still carries a running cost, because nothing that talks to customers runs unattended.

Worth remembering

The dangerous quotation is the one with a large build fee and no stated running cost. Every deployed assistant has ongoing cost — hosting, model usage, monitoring, content updates. A quote that omits it has not removed the cost; it has just not told you about it yet.

A third model is worth naming because it appears often in government and large enterprise procurement: a build fee plus an annual support and maintenance percentage, tendered together. It is the most transparent of the three, which is presumably why the sector that scrutinises contracts most has settled on it.

What actually sits inside the number

Whichever model you buy under, the same underlying components have to be paid for by somebody. A quotation is really a statement about who is paying for which.

Where the effort goes on a typical RAG deployment

Knowledge base: gathering, cleaning, structuring

30%

Integration with existing systems

22%

Language work: Arabic tuning and testing

18%

Platform configuration and interface

16%

Testing, review and go-live

14%

Illustrative structure based on how implementation effort distributes, not a published benchmark. Your own split moves with how tidy your documentation is.

Proportional effort, not price. The shape is what matters: the largest single cost in a serious deployment is usually the knowledge work, not the software.

That first band is the one buyers consistently underestimate. An assistant is only as good as the material it reads, and most organisations discover during implementation that their policy documents contradict each other, that three versions of the returns policy are in circulation, and that the authoritative one lives in somebody’s inbox. Resolving that is real work and it is unavoidable. It is also, incidentally, valuable on its own — several organisations get more out of the documentation clean-up than they expected to get out of the assistant.

The running costs, itemised

Four recurring lines exist on every deployment, however it was sold.

Model usage. Every message costs something to process. Where the model runs changes the shape of this cost dramatically: a hosted commercial API bills per unit of text and scales linearly with your traffic, while a model run on infrastructure you control has a largely fixed cost that does not care whether you send it a hundred messages a day or ten thousand. Below a certain volume the API is cheaper. Above it, the fixed cost wins, and the crossover is worth calculating rather than assuming.

Hosting and infrastructure. The application, the database, the search index. Modest for a text assistant, materially higher if you are running speech or your own model, and higher again if data residency requirements mean it must stay inside the Kingdom on specific infrastructure.

Content maintenance. Prices change, policies change, products launch. Someone has to keep the knowledge base current. In a RAG system this is document management rather than development, which is precisely the point — it can be owned by the team that owns the policy rather than by engineering.

Monitoring and improvement. Reading what customers actually asked, finding the questions that were answered badly, fixing them. This is the line most often dropped from a budget and the one most correlated with whether the deployment is still used in year two.

What pushes a quote up

If two quotations differ sharply, the difference is usually concentrated in a small number of these.

Cost drivers, and whether they are usually negotiable

	Raises cost	Often negotiable
Integrations with internal systems (CRM, ERP, ticketing)	Yes	No
Voice — speech in and speech out	Yes	No
True Arabic dialect handling, not only formal Arabic	Yes	No
Data residency inside the Kingdom	Yes	No
Number of distinct assistants or departments	Yes	Yes
Custom interface design rather than a standard widget	Yes	Yes
Volume of documents to process at launch	Yes	Yes
Single sign-on and role-based access	Yes	Yes

The four that are rarely negotiable are rarely negotiable for a reason: they are engineering scope, not packaging. A vendor who drops them without reducing scope has moved the cost rather than removed it.

Read this before the commercial conversation rather than during it.

Two of those deserve a note. Voice is not a feature toggle — it adds speech recognition and speech synthesis, both of which are considerably harder in Arabic than in English, and both of which have their own running cost. Expect it to be a separate line, and be suspicious if it is not.

Data residency is the one that most often reshapes a project in the Kingdom. If your compliance position requires that personal data stays on infrastructure in a particular jurisdiction, several otherwise-cheap options are simply unavailable to you, and the honest quotations will reflect that. Establish this requirement before you shortlist, not after.

What this actually costs, with numbers

Everything above is structure. Here is our own pricing against it, published in full rather than held back for a call — because an article titled "how much does this cost" that never answers the question is not much use.

Elbi deployment pricing, in Saudi riyals

	Setup (one-time)	Monthly Maintenance Cost
Single department — website assistant, Arabic and English, no integrations	SAR 50,000–75,000	from SAR 10,000
Multi-department — one or two integrations, live-agent handover portal	SAR 100,000–150,000	from SAR 15,000
Enterprise — voice, single sign-on, several integrations	SAR 180,000–250,000	from SAR 20,000
Government or regulated — processing and residency inside the Kingdom	SAR 300,000–500,000	from SAR 30,000

Read the maintenance column carefully: twelve months of it is the same order as the build itself — between 0.7 and 2.4 times it, depending on the tier and where in its range the build lands — so it, not the setup fee, is what you are actually signing up to.. Maintenance is billed six-monthly or annually and includes support and the service level. All figures exclude VAT.

Setup is a range because scope moves; maintenance is a floor because usage does. Compare a vendor on both columns together — a low setup beside a high monthly is a different proposition from the reverse, and only one of them is a one-off.

What the table does not price. Voice, WhatsApp, each additional integration and each additional department are quoted per case rather than at a flat rate, because the cost of connecting to a system depends almost entirely on that system — a documented REST interface and a twenty-year-old internal database are not the same job, and pretending otherwise in a price list only moves the disagreement to month three. Ask for them itemised on your own quotation; a vendor who cannot break them out has not scoped them.

Two things about that table are worth saying out loud, because they are the parts a buyer would otherwise have to work out.

The maintenance cost is the real cost. Twelve months of maintenance always exceeds the lower end of the build, and at the first tier it is roughly double it; only at the two largest tiers, and only when the build lands at the top of its range, does the build still lead after a year. That is deliberate and it is the honest shape of this kind of system — an assistant that answers customers needs its model running, its content maintained and its bad answers found and fixed, every month, forever. A proposal shaped the other way round — a large build fee and a token monthly — is describing a project that is going to be abandoned.

These are not chatbot-platform prices and should not be compared to them. A cloud chatbot subscription costs a small fraction of this, and for organisations whose requirements it meets, it is the right purchase. What the figures above buy is a different product: answers grounded in your own documents, Arabic handled as its own language, and conversation processing that stays inside the Kingdom with no external model provider involved. If none of those three matter to you, we are the expensive option and we would rather say so on this page than in month two.

What WhatsApp adds

Messaging is usually quoted separately, and it has a cost structure that surprises people because part of it is not paid to your vendor at all.

Business messaging on WhatsApp runs through the official Business Platform, which has its own commercial terms set by the platform operator and billed on conversations rather than on individual messages. That charge exists regardless of which vendor built your assistant. On top of it sits the vendor’s own work: connecting the channel, handling the different session rules, adapting answer formatting for a phone, and routing escalations back to a human on the same thread.

Worth remembering

Ask explicitly whether a WhatsApp quotation includes the platform’s own conversation charges or excludes them. Both are legitimate. Only one of them is the number you will actually pay.

What to ask before you accept a quotation

These questions are deliberately awkward. A vendor who answers them cleanly is usually worth more than one who quoted lower.

Nine questions that separate quotations

1Which of the four tiers is this?Scripted, FAQ, RAG or agent. Get it named. Everything else follows from the answer.
2What is the total running cost in year two?Not year one, which is often discounted. Year two is the real number.
3Who owns the knowledge base and the conversation data?If you leave, what leaves with you. Get it in the contract rather than the pitch.
4Where does personal data physically sit, and who else processes it?This is a PDPL question before it is a commercial one, and the answer may eliminate options.
5What happens when the assistant does not know?A vendor who has not designed the failure path has not deployed at scale.
6How is Arabic tested — and in which Arabic?Ask to see it handle dialect and mixed Arabic-English, not a formal sample.
7What is included in support, and what is a change request?The boundary between these two is where budgets are lost.
8Can we see it answering from our own documents before we sign?A short pilot on your real material tells you more than any demonstration.
9What does it cost to add a second language, channel or department later?Cheap to ask now, expensive to discover later.

Ask them in writing and keep the answers. Most disputes at month six are about something on this list that was discussed verbally.

What an Elbi deployment includes

For transparency about our own position rather than as a price list: an Elbi deployment is a tier-three or tier-four system in the terms above, and it is quoted as an implementation with a stated running cost rather than as a per-seat subscription.

What is inside the scope: the assistant itself and the language model behind it; the retrieval layer that grounds answers in your documents with citations; the knowledge base built from your material; Arabic and English handled as first-class languages rather than one translated into the other; the escalation path to your own people; and the analytics that let you see what customers actually asked and where the assistant struggled. Integrations with the systems you already run are scoped per project, because the cost of connecting to a system depends entirely on that system.

What is outside it, and should be outside anybody’s: a guarantee that automation will deflect a particular percentage of your contact. Nobody can know that before seeing your traffic, and a vendor who quotes it is quoting a number they made up.

If you want the components of your own quotation talked through — including the parts that are not ours — that is a conversation worth having before you shortlist rather than after.

Common questions

How much does an AI chatbot cost in Saudi Arabia?

There is no single figure, because "chatbot" covers four different products: a scripted button bot, an FAQ bot with language matching, a RAG assistant that reads your documents, and an AI agent that can act in your systems. The cost difference between the first and the last is a multiple rather than a percentage. Establish which tier you are buying before comparing any two quotations.

Is it cheaper to pay monthly or to pay once for a build?

A subscription starts cheaper and never stops; a build costs more on day one and leaves you owning something specific to your organisation. Neither removes the running cost of model usage, hosting, content maintenance and monitoring. A quotation with a large build fee and no stated running cost has not eliminated that cost, only omitted it.

Why do vendor quotes for the same brief differ so much?

Usually because the brief did not specify the tier, the integrations, the language requirement or the data residency position, so each vendor scoped something different. The four biggest differentiators are integrations with internal systems, voice, genuine Arabic dialect handling, and a requirement to keep data on particular infrastructure.

What does adding WhatsApp cost?

Two separate things. The official WhatsApp Business Platform has its own conversation-based charges set by the platform operator, payable whoever builds your assistant. On top sits the implementation work of connecting the channel, handling its session rules, reformatting answers for a phone and routing escalation back to a person. Ask whether a quotation includes the platform charges or excludes them.

What is the most underestimated cost?

The knowledge base. Gathering, cleaning and reconciling the documents an assistant will answer from is typically the single largest block of effort in a serious deployment, because most organisations discover their policies contradict each other. It is real work, it is unavoidable, and it is valuable independently of the assistant.