Choosing an AI company in Kuwait: what to ask before you sign

Three proposals are on the table. One is KD 900, one is KD 6,000, and the third comes “after a free discovery workshop”. Every demo looked impressive, every company says it understands Arabic, and none of the proposals explains what happens if the system gets an answer wrong. By the end of this article you will have ten questions that separate them, what a good answer to each sounds like, and what the first two weeks of any serious AI project should hand you.
The short answer
Judge an AI company in Kuwait on four things, not on the demo: whether it will prove the system works on your own data before you pay for the build, who owns what when the contract ends, what it costs every month after launch, and what happens when it is wrong. A company that answers those plainly, in writing, is worth talking to whatever its size. One that answers with a slide about the future of AI is not.
What the first two weeks should hand you
Start with the result, because it is the easiest test of a proposal. Two weeks into a well-run project you should hold five things, whether or not you continue:
- A one-page scope. The one job the system does, who uses it, and what it does not do.
- A data map. Which documents, systems and messages it reads, and what it may change.
- A test set. Real examples from your business, with the correct answer for each and a pass mark agreed in advance.
- A monthly running cost with a ceiling, not only a build price.
- A decision number. The result that means “build it” and the result that means “stop here”.
If a proposal cannot tell you when you will hold these, it is selling a build, not a result. These documents are also yours: if you stop after two weeks, you can take them to another supplier.
The ten questions
| Question | A good answer | A warning sign |
|---|---|---|
| 1. Will you test it on our data before the build? | Yes, on a sample of real cases, with a pass mark agreed first | The demo uses their sample data only |
| 2. Where does our data go? | Names each provider, the region and the retention period | “It is all secure” |
| 3. What does it cost per month after launch? | A monthly figure with a ceiling and what drives it | Only the build price is quoted |
| 4. What happens when it is wrong? | Describes the handover to a person and who is told | “It does not make mistakes” |
| 5. Who owns what at the end? | You own the accounts, the test set, the instructions and the code | Everything runs under their account |
| 6. Which model, and can we switch? | Names the model and how a switch would be tested | Will not say, or says it never needs changing |
| 7. What can it do on its own? | A written list of actions, and which need approval | “It handles everything automatically” |
| 8. How do you test Arabic? | Real messages in Kuwaiti dialect in the test set | A demo in formal Arabic only |
| 9. Who maintains it after launch? | A named service with a response time | Support is “as needed” |
| 10. Who can we call? | A client running something similar who will take the call | Logos on a slide, nobody to call |
Proof on your data, not theirs
Every AI demo works, because the company chose the examples. The question that matters is how it performs on the messages, invoices or documents your business actually receives, including the messy ones. A serious supplier will run a small test on a sample of your real cases before you commit to the build, and agree in advance what score counts as a pass. We explained how to run that test, and why a hundred examples is usually enough, in the 100-example test.
If the supplier also helps you choose which process to start with, ask how they decided. The best first project is usually narrow, frequent and easy to check, which is the argument in which process to automate first.
Data, cost and ownership
Three answers should be in the contract, not in a meeting. First, where your data goes: which AI provider receives it, in which region, and for how long it is kept. We set out the questions in detail in where your company data goes when you use AI in Kuwait.
Second, the monthly cost. An AI system has a running bill for model usage, hosting and maintenance, and it grows with use. A proposal that quotes only the build price hides the larger number over a year; what AI costs after launch shows how to budget it.
Third, ownership. Ask in writing who owns the AI provider account, the instructions the system runs on, the test set and any code. If all of it lives in the supplier’s account, you cannot change supplier without starting again.
When it is wrong, and what it may do
Every AI system makes mistakes. What separates a good build from a risky one is what happens next: whether the system knows when to stop and hand the conversation to a person, who is told, and how fast. A supplier who claims the system will not make mistakes has either not tested it or is not being straight with you.
The same applies to actions. An assistant that only answers questions is low risk; one that can issue refunds, change orders or send messages needs a written list of what it may do alone and what needs a person’s approval. The OWASP Top 10 for large language model applications lists excessive agency and prompt injection among the main risks, and our post on prompt injection explains how an assistant gets talked into things. For a wider checklist, the NIST AI Risk Management Framework is free and readable.
Arabic is a test, not a claim
Almost every AI company in the region says it handles Arabic. The test is whether its test set contains real messages from your customers, in Kuwaiti dialect, with spelling mistakes and mixed English. Ask to see results on those, not on formal Arabic written for the demo. Our post on why Arabic assistants keep handing over covers the failures to look for.
Who actually builds it
Ask who will do the work, by name, and meet them before signing. In many AI proposals the person in the sales meeting is not the person who builds the system, and part of the work goes to a subcontractor you never meet. That is not wrong in itself, but you should know where your data and your project are going, and who to call when something breaks after launch.
Ask for one reference you can phone: a client running something similar, ideally in Arabic, who will tell you what went wrong and how it was fixed. A supplier with nothing to show yet can still be the right choice, but then the paid test on your data matters even more.
What we stand against
The free discovery workshop that ends in nothing but a proposal. It feels generous, but two hours of questions produce no test, no data map and no cost ceiling, only a price for a build nobody has proven. A paid first phase with documents you keep is worth more than a free workshop you cannot use elsewhere.
We also stand against choosing from a “top ten AI companies in Kuwait” list. Those lists rank companies; they cannot tell you which one fits your process, your data and your budget. The ten questions can.
How we work
Our custom AI assistants and AI agents and automation both start with the five documents above, prices are published on each service page, and the accounts are set up in your name. Send us the task that takes your team the most time and we will tell you whether it is worth testing.
Frequently asked questions
How much does an AI project cost in Kuwait?
It depends on the scope, and the monthly running cost matters as much as the build. As one reference point, Tothiq publishes fixed starter prices on each service page, for example KD 1,250 to build a custom AI assistant plus a monthly running fee.
Should we choose a local AI company?
Local knowledge helps with Arabic dialect, KNET and Kuwaiti paperwork, but location matters less than the answers to the ten questions. A local company that will not test on your data is no safer than a foreign one.
What should the contract include?
The scope, the test set and pass mark, the monthly cost and its ceiling, who owns the accounts and code, what the system may do on its own, and the support response time after launch.
Do we need a lot of data to start?
Usually not. Most first projects need a few hundred real examples and access to the documents or systems the job uses. The test in the first two weeks tells you whether that is enough.
How long before we see something working?
A narrow first version usually takes weeks, not months. If a proposal needs months before you see anything you can test, ask what is being built in that time.

Leave a Reply