What AI Actually Costs in the UAE

Nexus Labs AI — Wed, 30 Sep 2026 12:58:21 +0000

Three published price guides put the same product — a basic FAQ chatbot — at AED 6,000, AED 12,000 and AED 150,000. Here is why they disagree, what really moves the number, and what to check before you sign anything.
Search for what an AI chatbot costs in the UAE and you will find a dozen 2026 pricing guides, each confident, each formatted like a rate card. Read three of them side by side and the confidence stops making sense.

A Dubai agency puts a basic FAQ bot at AED 6,000–15,000. Another UAE firm puts the same thing at AED 12,000–25,000. A third puts a basic rule-based chatbot at AED 150,000–275,000.

The cheapest guide’s floor and the dearest guide’s floor are twenty-five times apart. Same three words, same country, same year.

PUBLISHED PRICE FOR A “BASIC” FAQ CHATBOT — AED THOUSANDS
Gravitybase
6k – 15k
Emtech
12k – 25k
Appinventiv
150k – 275k
0
100k
200k
300k
Linear scale. The first two bars really are that small.
Entry-tier bands as published by three UAE-focused cost guides, September 2026. Sources listed at the foot of this article.
Why the guides disagree
Three things are going on, and none of them is dishonesty.

“Chatbot” is not one product
A decision-tree bot that answers eleven questions from a script and a retrieval system that reads your policy documents and answers in Arabic are both sold under the same word. The first is a week of work. The second is a data project with a chat window attached. A price guide that does not say which one it is pricing is not telling you anything.

Some AED prices are USD prices wearing a disguise
The dirham is pegged at 3.6725 to the dollar. Take the highest guide’s bands — AED 150,000 / 275,000 / 735,000 / 1,470,000 — divide by the peg, and you get $40,000 / $75,000 / $200,000 / $400,000: four round dollar figures. Those are not UAE project prices. They are a global rate card converted at the border.

Every guide is anchored to its author’s own book
An agency that builds enterprise systems for banks publishes enterprise numbers. A studio that ships WhatsApp bots for clinics publishes clinic numbers. Both are describing their own order book accurately and calling it the market. We are doing the same thing further down this page, and we are telling you so.

What actually moves the number
Scope words like “basic” and “enterprise” carry almost no information. These six things carry all of it, and you can assess every one of them before a vendor quotes you.

Biggest single driver
How many systems it has to touch
An assistant that answers from a document set is one build. The same assistant that also reads live stock, writes to your CRM, raises a ticket and checks a delivery status is five builds. Each integration brings its own authentication, its own rate limits, its own failure modes and its own owner inside your company who has to approve it. Count the integrations in your brief. That count predicts the price better than any other single fact.

Usually underestimated
The state of your content
Retrieval systems answer from what you give them. If your policies live in forty PDFs of varying vintage, three of which contradict each other, somebody has to reconcile that before the system is trustworthy — and it will not be the model. This is the line item that quietly doubles projects. It is also the one you can shrink yourself, for free, before anyone quotes.

Region-specific
Arabic, and which Arabic
Modern Standard Arabic is handled well by current models. Gulf dialect in a voice note, code-switched with English mid-sentence, is a different problem, and it is what your customers actually send. Bilingual quality assurance is real work with real cost: someone has to read the transcripts. Budget for it rather than discovering it in month two.

Regulated sectors
Where the data is allowed to go
If you are in health, finance or government supply, the question of which jurisdiction processes your customer data stops being a technical preference. In-region inference, data-processing agreements and retention controls narrow your vendor options and add cost. Raise it in the first meeting; discovering it at integration is the expensive path.

Smaller than people expect
Volume
Going from one thousand conversations a month to ten thousand multiplies your model bill by ten and your build cost by roughly nothing. The arithmetic is below, and it is less frightening than most buyers assume.

The line people forget
Who owns it in month seven
Products change, policies change, models get deprecated. Someone updates the content, watches the transcripts and fixes what broke. Whether that is your vendor on a retainer, your own team after a handover, or nobody, that decision belongs in the budget. “Nobody” is a choice with a cost; it just arrives later.

The running cost, done properly
Most buyers arrive worried about the model bill and relaxed about everything else. It is usually the other way round. Here is the arithmetic for a support assistant that reads your documents, on published rates for a mid-tier model at $2 per million input tokens and $10 per million output tokens.

Retrieved context and instructions, per turn
~4,000 tokens in
Customer message, per turn
~60 tokens in
Reply, per turn
~250 tokens out
Turns in an average conversation
6
Model cost per conversation
AED 0.23
3,000 conversations / month
~AED 700
10,000 conversations / month
~AED 2,340
Same, with prompt caching switched on
~AED 1,400
Ten thousand conversations a month for roughly the price of one part-time shift. That is the line everybody worries about, and in a well-built system it is rarely the largest one.

What sits above it: hosting and the vector database, the messaging channel itself — WhatsApp is billed per conversation template by Meta, separately from your model spend — monitoring, and the human hours spent reading transcripts and correcting the content. One UAE guide puts total annual upkeep at 25 to 40 percent of the original build cost. That is on the high side of our experience, but a proposal that implies zero ongoing cost is the one to distrust.

Ask any vendor for the month-thirteen number, not the launch number. The gap between those two figures tells you what kind of company you are dealing with.

What we quote, and why
These are Nexus Labs’ own bands, from our own work in the UAE. They are not a market survey, and we are not claiming they are the right price for every firm in Dubai. They are what we charge, published so you have something concrete to compare against.

Scripted FAQ assistant, one channel, no integrations
AED 15,000 – 30,000
Document-grounded assistant, bilingual, one or two integrations
AED 45,000 – 95,000
Multi-system assistant with CRM write-back and handover
AED 110,000 – 240,000
Discovery sprint, fixed fee, credited against the build
AED 12,000
Anything above that range stops being a chatbot project and becomes a systems-integration programme, and should be scoped and governed as one.

What a fair quote contains
You do not need to be technical to audit a proposal. You need it to answer these six questions in writing.

Which systems does it connect to, named individually? Not “CRM integration” — which CRM, which objects, read or write.
Who prepares the content, and is that time in this price? If the answer is “the client provides materials”, that is your cost, not theirs.
What does month thirteen cost? Hosting, model usage, channel fees and human upkeep, itemised.
Where is the data processed, and what is retained? Named region, named retention period.
What happens when it cannot answer? A defined handover to a human is the difference between a helpful system and a complaints generator.
Who owns the code and the prompts at the end? Get it in the contract, not the kickoff call.
A vendor who answers all six in a page is worth more than one who is twenty percent cheaper and answers none.

Stage the money
The UAE is not short of AI enthusiasm. Microsoft’s AI Diffusion Report for the first quarter of 2026 put UAE working-age AI adoption at 70.1 percent, the highest in the world and roughly four times the global average of 17.8 percent. Enthusiasm is not the constraint here. Discipline about what gets funded is.

Widely-cited research through 2025 and 2026 claimed the large majority of corporate generative-AI pilots returned nothing measurable. The exact percentages are contested and the methodologies have been argued over, but the direction is not seriously disputed: most pilots do not become products. The practical response is not to spend less. It is to spend in stages.

Stage one, four to six weeks. One workflow, real users, a number you agreed to move before you started. Small enough that cancelling it costs nothing but the fee.
Stage two. Fund the build only if stage one moved the number. If it did not, you have bought a cheap and honest answer.
Stage three. Budget upkeep from day one, as a line in the operating budget rather than a surprise in month seven.
The firms getting value from AI in this market are not the ones who spent the most. They are the ones who found out early, cheaply, whether the thing worked.

A word on this article
Nexus Labs AI builds these systems in the UAE and is quoted above with its own prices, so read this the way you would read any vendor’s guide: with the comparison-shopping instinct switched on. The difference we are trying to earn your attention with is that the sourced figures are sourced, the unsourced ones are labelled, and where our numbers disagree with someone else’s we have said so rather than quietly published our own and called it the market.

If you want a second opinion on a quote you have already received, send it over. We will tell you what we think of it whether or not it is ours.

DEV Community: Nexus Labs AI

What AI Actually Costs in the UAE