Most companies who search for a conversational AI consultant have already been sold a chatbot once. Some have one on the website that nobody uses. Some have a WhatsApp number that answers in fluent English to customers who wrote in Arabic. A few had a pilot that impressed in the demo and was switched off after the first week of real messages.
What they want the second time is not a better platform. It is somebody on their side of the table who can say what the assistant is for, what it must refuse, how it is tested before customers see it, and how it is handed over so the company runs it without the vendor. That is the work I do, for businesses in the United Arab Emirates, in Arabic and English.
What a consultant does that a chatbot vendor does not
A vendor sells a platform and is paid when you sign. A consultant is paid to get the decisions right, and the biggest decisions in a conversational project have nothing to do with which platform you buy.
The first is scope: which questions the assistant answers, which it declines, and which it hands to a person at once. The second is grounding: which of your documents and systems it may read, and what happens when the answer is not in them. The third is the acceptance test: the written set of real messages it must pass before launch, and the score it must reach. The fourth is ownership: who inside your company changes a wording, reads the weekly sample of conversations, and can switch it off on a Friday night.
I write all four before a platform is chosen. The platform comes last, against those requirements, and the choice is usually easy because most products are competent at the mechanics. What separates a good assistant from an embarrassing one was decided in those four documents, not in the vendor demo.
The UAE conversation is not in one language
An assistant built for the Emirates meets customers who write in Gulf Arabic, in English, in Arabic typed with Latin letters, and in a dozen other languages, often inside one message. Most products are benchmarked in English, and the Arabic they handle out of the box is Modern Standard Arabic, which is not what a customer in Sharjah types at eleven at night.
The decision this forces is not “support every language”. It is to decide, per language, what the assistant does. For most companies the honest answer is that it answers in Arabic and English, recognises the rest well enough to route them, and hands over rather than guessing.
I test this with your own inbound messages, not a vendor sample. The method is written up in what an Arabic conversational assistant has to pass before customers see it. On this page the point is simpler: if your customers write to you in Arabic, the assistant is accepted in Arabic, with your messages, before it goes live.
Where the assistant lives
In this market the first channel is almost always WhatsApp, and WhatsApp has rules that shape the design more than people expect. A business can send a free-form reply only inside the 24-hour customer service window that opens when the customer writes in. Outside it, only a pre-approved template can be sent, and every template is categorised by Meta as marketing, utility or authentication first. They are set out in Meta’s template documentation, and they decide, for instance, whether your assistant can follow up an unanswered quotation tomorrow or only today.
The cost rules are changing too. Meta’s pricing page for the WhatsApp Business Platform states that from 1 October 2026 replies inside the service window are charged per message beyond a free allowance of one thousand a month per business number, and that UAE marketing template rates rise on the same date. An assistant that sends five short messages where one complete one would do now spends your money five times. That belongs in the requirements, not on the first invoice.
What I build
Customer enquiry and booking assistants. The questions that arrive dozens of times a day — availability, coverage, status, documents required, what happens next — answered from your own material, in the customer’s language, with a clean route to a person. Where a booking or quotation can be captured, it is captured completely.
Internal knowledge assistants. For your own staff, over policies, procedures, contracts and years of shared drives. Often the better first project, because the users are patient and the mistakes are private.
Assistants that take an action. Checking an order in the ERP, raising a ticket, updating a record, sending a confirmation. This is where the safety work is, because a model that can act can act wrongly. Every action has a boundary, an audit trail and a stop button, and nothing is connected to a live system until the four questions that come before connecting a model have written answers.
What it will never be allowed to do
An assistant is defined as much by its refusals as by its answers. Some items are on every list.
- It never states a price, a delivery date or an availability it did not read from the system that is authoritative for that fact.
- It never gives legal, medical, financial or regulatory advice.
- It never pretends to be a person. It says it is automated when asked and when it matters, consistent with the transparency principle in the UAE Charter for the Development and Use of Artificial Intelligence.
- It never keeps a customer who is angry, confused or asking for a person inside the loop. Those are handover triggers, not conversations to win.
- It never invents. When the material does not hold the answer, the right response is a short honest one and a route to somebody who knows.
The data question comes first
Before any message reaches a model, three questions are answered in writing: what customer data leaves your systems, where it is processed, and how long it is kept. The federal Personal Data Protection Law is in force, and the government’s summary of the data protection laws is the place to start; DIFC and ADGM companies have their own regimes on top.
Conversation logs are personal data and need a retention period. Anything touching identity documents or health may need a model deployed in-region rather than a public API. I work through what the UAE PDPL asks of you before you connect an AI model at the start of every project, not in a security review three weeks before launch.
How the engagement runs
1. Read the real conversations. I start with a few hundred of your actual inbound messages, across channels and languages, and sort them. This shows what it will really be asked, in which language, and how often.
2. Write the scope, the refusals and the acceptance test. A short plain-language document your management can sign: what it answers, what it declines, what it does when it does not know, and the real messages it must pass with the score it must reach. It is also the core of the contract with whichever vendor builds the platform.
3. Build, or oversee the build. For a small company I build it. For a larger one with a preferred platform or an existing supplier, I sit on your side through the build and run the acceptance test myself, on your messages, with the result written down.
4. Hand over and review. Someone in your company is trained to change wording, read the weekly sample and switch it off. The first monthly review is part of the engagement.
A first customer-facing assistant typically takes six to ten weeks from the first conversation to launch, with the first two weeks spent reading and writing rather than building.
What it costs, in shape
Discovery and the written requirements are a fixed fee, useful even if you then decide not to build. The build is quoted fixed against the written scope. The running cost is yours and is dominated by people keeping the material correct rather than by the model; what drives the annual number is set out in what an assistant costs to run for a year.
I do not resell platforms or take commission from vendors. If a platform is the right answer, you contract with it directly.
How you will know it worked
Three numbers, agreed before launch and reported monthly. The share of conversations resolved without a person, which should rise slowly rather than be forced. The share of handovers that were correct — passed on when it should have been, and not otherwise. And the time to the first useful reply outside working hours, which is where the value usually is.
Underneath sit the engineering measures — grounding accuracy, refusal precision, the acceptance score as it is re-run each month — and I report those too. The framing borrows from Google’s People and AI Guidebook and the governance structure of the NIST AI Risk Management Framework.
Conversational work is one part of the wider AI consulting I do in the UAE. If the honest answer after discovery is that an assistant is not your bottleneck, that is what the written recommendation will say.
Frequently asked questions
Do we need a consultant if the platform vendor offers to set it up for us?
The vendor will set up their product well; that is not the risk. The risk is in the decisions before setup — scope, refusals, grounding, the acceptance test and ownership — and a vendor is not the right party to decide what its own product should refuse or how it is tested. If those are already written and signed, you may not need me.
Which is better for our customers, WhatsApp or a chat window on the website?
In the Emirates, WhatsApp is almost always where the customers already are, so it is usually first. But it carries Meta’s template rules and the 24-hour window, and from October 2026 a per-message charge on replies beyond the free allowance, so the assistant has to be designed for it rather than dropped into it. Most companies end up with both, reading the same material; the decision is about sequence.
Can one assistant handle Arabic and English properly?
Yes, if it is tested in both with your own messages rather than assumed. Current models read Arabic well; the failure is that dialect, names, numerals and mixed-script messages are handled inconsistently, and an assistant tested only in English replies in the wrong register to Arabic customers. The acceptance test is run separately in each language and the assistant is not accepted until both pass.
How long does a first assistant take?
Six to ten weeks for a customer-facing assistant over your own material, of which the first two are reading your real conversations and writing the scope and the acceptance test. Anything that takes an action in another system takes longer, because the safety work around each action is where the time goes.
Will we depend on you afterwards?
No, and that is designed in. Someone in your company is trained to change the wording, read the weekly sample of conversations and switch it off, and the first monthly review is done together. After that the assistant is yours. A monthly retainer is optional; the assistant does not stop working without it.