TL;DR: Before you sign with an AI agency in the UAE, get clear answers on six things: who owns the code, prompts, accounts and data; where your data is hosted and who sees it; what the system does when it does not know; how success will be measured and from what baseline; who maintains it after launch and at what cost; and what leaving looks like. The answers tell you whether you are hiring engineers or renting a product with a new logo. Ask them in writing.
Why the questions matter more than the demo
The questions to ask an AI agency before signing are the ones a demo cannot answer: ownership, data handling, failure behaviour, measurement, maintenance and exit. Every agency in Dubai and Abu Dhabi can show you a chatbot answering questions on a screen. What separates a system you can run your business on from a demo you regret is what sits underneath, and you find that out by asking, not watching.
This list is written for a business owner or operations lead, not an engineer. Each question comes with what a good answer sounds like and what a weak one sounds like. You do not need to understand the technology to tell the difference; you need to notice whether the answer is specific to your business or generic to the product.
Who owns what when the project ends?
Ask this first, because everything else depends on it. The system you are buying has several parts: the code, the prompts and configuration, the knowledge base built from your documents, the accounts (WhatsApp Business Account, model provider, hosting, workflow tools), and the data that accumulates as customers use it. For each part, ask: is it in my company's name, and could another team take it over without you?
A good answer names each part and says which are yours. The WhatsApp Business Account is in your Meta Business account. The model and hosting accounts are yours or transferable. The code is delivered to a repository you control, with documentation. The prompts and knowledge base are yours. A weak answer talks about "your solution" without naming the parts, or reveals that the WhatsApp number and accounts live under the agency's business, which means leaving the agency means losing the number your customers have saved.
Where is our data, and who can see it?
Your documents, customer messages and CRM records will be stored and processed somewhere. Ask where, in which countries, by which third parties, whether any of it is used to train anyone's models, and how it is deleted when you leave. If you are in healthcare, finance, legal or government work, ask specifically whether an in-country or in-region hosting option exists and what it changes about the design.
A good answer draws the path of a message from your customer's phone to the reply, naming every system it passes through, and can explain what changes if data must stay in the UAE. A weak answer is "it's all secure" or "it's on the cloud". Check the current UAE data protection guidance for your sector alongside whatever the agency says; the agency should be able to tell you what their design assumes, but your obligations are yours.
What happens when the AI does not know?
Every AI assistant will be asked things it cannot answer. The interesting question is what it does then. Ask: where does the assistant's knowledge come from, what does it say when the answer is not there, when does it hand off to a person, where does that handoff land, and what is it forbidden from doing on its own?
A good answer describes grounding in your documents, a written scope of allowed and handoff topics, a shared inbox where staff take over, and a list of actions that need validation or human approval. Ask to see a conversation where the assistant declined to answer; a serious agency will be proud of it. A weak answer is "it learns from your data and gets smarter", which describes nothing. We wrote about the mechanics in how to keep control when AI answers your customers.
Show me a failure
Ask the agency to describe a project where something went wrong and what they changed. Ask what the most common failure is for the kind of system you are buying, and how they detect it. Ask what the log looks like and who reads it.
A good answer is specific and a little uncomfortable: a knowledge base that answered from an old price list, an integration that broke when a CRM changed its API, a handoff that nobody picked up over a weekend, and the check that now exists for each. A weak answer is that nothing goes wrong. Language models are probabilistic and connected systems change; an agency that has never seen a failure has not run anything for long.
How will we measure whether it worked?
Ask what will be measured, how, and against what baseline. For a WhatsApp assistant, that might be the share of enquiries answered without a person, handoffs per day, and how long handoffs wait. For a document automation, it might be documents processed, exceptions raised, and corrections made by staff. Ask when the first review happens and what changes if the numbers are poor.
A good answer commits to instrumentation and a review, and is honest that the numbers depend on your data and your team's use of the system. A weak answer promises a percentage or a savings figure before anyone has looked at your operation. Be cautious of guaranteed outcomes in a proposal; they tend to come from a slide deck rather than from your business.
What does the system actually consist of?
You do not need to understand the answer in depth, but you should ask for it and notice whether it is specific. Ask the agency to name each component between a customer's message arriving and the reply going out: the channel, the model, where your documents are searched, what checks run before an action, what is logged, and which tools are third-party products.
A good answer is a short list of named parts and what each does, including which are off-the-shelf and which are built for you. Using a workflow tool or a provider's inbox is not a problem; hiding it is. A weak answer is a brochure. If every question about how it works is answered with what it does, you are probably buying a reseller's product, and you should price it as one.
Who maintains it, and what does that cost?
An AI system is not finished at launch. Documents change, WhatsApp templates need updating, model providers change their models, and the CRM you integrated with will update its API. Ask who does this work, whether it is included or charged separately, what response time you can expect when something breaks, and whether you could take maintenance in-house or to another team.
A good answer distinguishes a warranty period from ongoing support, explains what each covers, and is clear that recurring third-party costs, such as model usage and template messages, are passed through rather than bundled. A weak answer is silence on maintenance, or a monthly fee with no description of what it buys.
What drives the price?
You should not expect a price before scoping, but you should expect the agency to explain what moves it. Ask what makes a project like yours cost more or less, and how they charge: fixed price per scoped project, retainer, hourly, or usage-based.
A good answer names the drivers, typically the number of integrations, the state of your documents, the languages, the volume of messages and any hosting or compliance requirements, and prefers a fixed price after scoping so the estimation risk sits with them. A weak answer is a price quoted on the first call without questions, or open-ended hourly work with no defined deliverable.
Who will actually do the work?
Ask who will build the system, whether they are employees or subcontractors, where they are, and who you will talk to when something is wrong. Ask whether the people in the sales meeting are the people who will build it.
A good answer names the engineers, or at least the roles, and is straightforward about subcontracting if it happens. A weak answer avoids the question.
What does leaving look like?
Ask for the exit in writing before you start. What do you keep, what documentation is handed over, which accounts transfer, how long does your data stay on the agency's systems, and what does a handover to another team involve?
A good answer is that leaving is boring: everything was in your name from the start, the documentation exists, and deletion happens on a stated timeline. A weak answer is any hesitation.
The questions on one page
| Question | Good answer sounds like | Weak answer sounds like |
|---|---|---|
| Who owns the code, prompts, accounts and data? | Each part named; accounts in your name; code in your repository | "Your solution", with accounts under the agency |
| Where is our data and who sees it? | The path of a message drawn end to end; in-region options explained | "It's secure" or "it's on the cloud" |
| What happens when the AI does not know? | Grounding, a written scope, a handoff into a shared inbox, gated actions | "It learns and gets smarter" |
| Show me a failure | A specific incident and the check that now exists | "Nothing goes wrong" |
| How will we measure it? | Named metrics, a baseline, a scheduled review | A guaranteed percentage before seeing your data |
| What does it consist of? | A short list of named components, built and bought | A brochure |
| Who maintains it and at what cost? | Warranty vs support explained; pass-through costs itemised | A monthly fee with no description |
| What drives the price? | Integrations, documents, languages, volume; fixed price after scoping | A number on the first call |
| Who does the work? | Named people or roles; honest about subcontracting | Avoids the question |
| What does leaving look like? | Boring: you keep everything, deletion on a timeline | Hesitation |
How Soluvide answers these
We would rather be asked these questions than not, so here is our position in brief. Accounts, code, prompts and knowledge bases are in the client's name from day one; the WhatsApp Business Account sits in your Meta Business account, not ours. We describe the architecture of every system we build in plain terms, including which parts are off-the-shelf, and we explain hosting and data paths before the proposal, with in-region options where a sector needs them. Every customer-facing assistant launches with human-in-the-loop controls and a log. We scope first and send a fixed-fee proposal after a free conversation, with recurring third-party costs passed through and itemised.
What that looks like in practice is on our AI integration and consulting page, which covers workflow mapping, model selection and guardrails, and on AI chatbots and agents for customer-facing systems. If you want to test us on the questions above before anything else, the free automation audit is a low-commitment way to do it.
Bring this list to your next vendor meeting, including ours. Message us on WhatsApp and ask any of them; we will answer in writing.
Questions
Frequently asked.
Who owns what when the project ends. Ask whether the code, prompts, knowledge base, WhatsApp Business Account, model accounts and hosting will be in your company's name, and whether you could hand the system to another team without the agency's help. If the answer is vague, everything else the agency promises depends on their goodwill rather than your contract.
Where this applies