Every supplier will quote what you described
Search for AI development in this market and you get app-development agencies, each with a page promising custom models, agents and integrations. They are not dishonest. They are simply built to price what you hand them, and the thing you hand them is a description written before anybody checked whether it was the right thing to build.
That is where the money goes here. Not on bad engineering, but on a well-executed build of the wrong specification: a system that does what the brief said, in a company where the brief was a guess. The work I do starts one step earlier, at the specification, and I say plainly when the answer is to buy something instead. If what you need is judgement on the whole technology estate rather than one build, that is a fractional CTO engagement, and I will tell you so.
What is actually being built
Very little of an AI system is the model. The model is a paid interface somebody else maintains. What gets built around it is ordinary software with unusual failure modes: connectors into systems that were never meant to be read by anything but a person, a schema the answers must conform to so your code can trust them, permission to take actions with a clear boundary on which ones, retrieval so answers come from your own material, and a log detailed enough that a wrong answer can be reconstructed rather than argued about.
That is the honest description of the discipline, and it is why the question that decides your budget is not which model, but how many systems it has to reach and how bad a wrong answer is. A build that touches one read-only source is a different project from one that writes into your finance system, even when the demonstration looks identical.
What I build
- Systems that act, not systems that answer. Reading, classifying, extracting and routing the documents and requests that currently move by hand.
- Retrieval over your own material, with a citation on every answer, so staff can check it rather than trust it.
- Internal tools placed inside the screens your team already uses, because a capability behind a new login gets used for three weeks.
- Integration work: the connectors, the permissions, the queue for the cases a human still has to see.
- Customer-facing assistants where that is genuinely the project, which has its own page: conversational AI consulting.
Where the system will live
This is a UAE-specific decision and it belongs at the start, not in the security review a week before launch. The federal Personal Data Protection Law sets the obligations, and the government publishes its overview of the data protection laws that apply, including the separate regimes inside the DIFC and the ADGM. A company inside one of those free zones is not under the same rules as a mainland company, and the difference decides where processing may happen.
The good news is that the infrastructure exists. Both major providers now run local regions, published on the AWS regions list and in Microsoft’s products by region table. The catch, which nobody mentions in a proposal, is that not every managed service is available in every region, and the specific AI service you want may not be in the local one. That single fact decides the architecture, so check it against the list before the design is agreed rather than after.
Your customers are not writing in one language
A system built here has to survive Arabic, English, and the mixture people actually type, and in many businesses Hindi, Urdu and Tagalog as well. That is not a translation task added at the end. Retrieval has to find the right paragraph in an Arabic document, which means the text has to come out of the file properly in the first place and match despite the spelling variants people use. Names, amounts and reference numbers have to survive being read in one script and repeated in another, which is exactly where errors pass unnoticed because the output still looks correct.
The practical requirement is short: put real messages from your own inbox, in every language your customers use, into the acceptance set on the first day. A supplier who has not asked for them has not thought about this.
Acceptance has to be a measurement
A demonstration proves a system can be right once. What matters is how often it is wrong, in which direction, and whether last week’s change made it worse. That means a written set of real cases with expected answers, run on every change and scored the same way each time. Microsoft’s own design and evaluation guide for retrieval systems treats evaluation as a phase of the build rather than a check at the end, and that is the right way to contract for it.
Put the score sheet in the agreement with a number against it. It is the clause that turns acceptance from an opinion into a fact, and the one most often missing from a proposal. The same argument, applied to assistants specifically, is set out in what an Arabic assistant has to pass.
The bill has three parts, and only one is quoted
The build price is what gets compared. The running bill is usage multiplied by volume, plus retries, plus the retrieval index, plus logging, and it rises with success. The third part is maintenance: models are retired and replaced roughly every year, your documents change, your processes change, and a system with no owner degrades quietly. Establish three numbers before anyone quotes: requests per day, the average size of what goes into the model, and the share of cases that still need a human. A full worked version of this arithmetic is in what an assistant costs to run for a year.
What has to be settled before it goes live
- Authority. What the system may do without asking, and what it must never do at all. The four questions that decide this are in connecting a model to a live system.
- Attack surface. The OWASP Top 10 for LLM Applications is the shortest useful list: instructions hidden in the content it reads, data leaving through the output, and permissions wider than the task needs.
- Ownership. The first function of the NIST AI Risk Management Framework is govern, and it is the one that survives translation into any size of company: a named person is accountable for what the system says.
What I will tell you not to build
Anything whose purpose is a demonstration to a board. Anything replacing a process nobody has written down, because writing it down is the project. Anything where a wrong answer reaches a customer or a regulator with no human between and no way to know it happened. And anything custom where a product would do, which is the most expensive error available, because it is paid once to build and again every year to keep alive. If the pilot you already have has stalled, the reasons are usually the four in why UAE pilots never reach production, and none of them is the model.
How the work runs
Four steps. A short assessment that picks the use case, settles build against buy, and names the systems involved. A written specification: data, boundaries, acceptance set, failure behaviour and the running-cost model. A build of one narrow slice that reaches real users rather than a pilot that reaches a slide. Then handover, which means your team can run it, change it and read its logs without me. If you want the decision examined before any of that is committed, start at AI consulting.
Frequently asked questions
How is this different from hiring a development agency?
An agency prices the brief you bring. I question the brief first, and a meaningful share of the time the honest answer is a smaller build, a product you can buy, or nothing at all. After that the delivery looks similar, and I am happy to specify the work and oversee your agency building it rather than replace them. What you are buying is the decision and the specification, not extra hands.
Do we need to train our own model?
Almost certainly not. Training needs a quantity of labelled data most companies do not have, costs a great deal, and the result falls behind the general models within a year. Nearly every business case is solved by giving a general model controlled access to your own material and systems. If a case genuinely needs a trained model, that is a finding from the assessment rather than an assumption at the start.
Does our data have to stay inside the UAE?
It depends on what the data is, which regime you sit under, and what you have told customers. Some categories can be processed abroad with the right agreements in place, some cannot, and a company inside the DIFC or the ADGM answers to a different set of rules from a mainland one. It is answered in the first week, in writing, because it decides the architecture, and an architecture changed afterwards is a rebuild.
How long before something real is in use?
Six to ten weeks for a first slice that real users touch, on a narrow case using data you already hold. Longer usually means the scope was never narrowed or the data turned out to live somewhere nobody had checked. I would rather cut scope than move the date, because eight weeks of real use teaches you more than a larger system still in testing.
What happens when the model we built on is retired?
It will be, roughly every year. The build keeps the model behind a boundary in your own code instead of scattering provider-specific calls through the application, and the acceptance set exists so that changing model becomes a test run rather than a rebuild. Ask any supplier how they intend to handle it, and treat a vague answer as a price you will pay later.