Choosing an AI provider in Algeria: 7 questions

To choose an artificial intelligence provider in Algeria, do not judge the pitch — it has become the same everywhere. Ask questions whose answers can be checked: which client uses the AI every day, where the model runs, who owns the code, how delivery will be validated. A provider who delivers answers with names and documents; a provider who presents answers with a demonstration.

The Algerian AI market filled up in two years. Most players now use the same words: sovereignty, on-premise, bespoke AI, data that never leaves the country. Those words no longer set anyone apart. What sets a provider apart is what can be verified.

1. Which client uses your AI every day?

This is the most important question, and the most revealing. A demonstration proves a piece of software can work once, under chosen conditions. A client using it every day proves it holds up against real data, real users and real failures.

A good answer gives a client name, a sector, and a readable description of what was delivered. A worrying answer talks of "many confidential projects" without being able to name a single one, or describes products "being finalised".

2. Where exactly does the model run?

"On-premise" has become a sales argument; check what it covers. A genuinely local model runs on your server, with no outbound network call at all, and keeps working when the internet is down.

The test is simple: ask what happens if the network cable is unplugged. If the assistant stops answering, the model is not on your premises. The question matters all the more since decree 25-320 provides for a national interconnection separate from the internet network.

3. Who will own the code?

Two models coexist. In the first, you rent access to a platform: the code stays with the provider, and your business depends on its survival, its prices and its choices. In the second, the bespoke code is handed over to you: you can develop it, change provider, or carry on alone.

A good answer is written into the contract: code handed over on full payment, data exportable at any time in common formats. If the answer is only spoken, it does not exist.

4. How will delivery be validated?

Most disputes are born of a misunderstanding about what was owed. You avoid it with a requirements document in which every requirement carries a unique reference, carried over unchanged into an acceptance test plan. A function counts as delivered when you accept its test — not when it has been demonstrated.

Ask for an example of a requirements document and a signed acceptance certificate from a past project. A provider who has them produces them; a provider who does not explains why they are unnecessary.

5. What is the real status of each product?

A catalogue can mix three very different things: what runs at clients, what was delivered once on assignment, and what exists only as a mock-up. Shown side by side without distinction, they look equal.

A good answer distinguishes these states explicitly, product by product. It is the mark of a provider who does not need to inflate the offer to convince.

6. How are access rights partitioned?

An AI that reads all your documents can reveal their content to any user if partitioning is poorly designed. The decisive technical question is this: is the access boundary enforced before the model reads the data, or only at display time?

Filtering at display protects nothing: the model has already read, and a clever question can make it reveal what it was not meant to show. Only partitioning enforced before reading fits a classification of data by sensitivity level, such as the one decree 25-320 introduces.

7. Will your teams be trained?

Software nobody knows how to use produces no value, and it is the most common cause of failure. Training should not be an option: it is part of the delivery.

A good answer describes sessions run on your premises, on your own data rather than a sample set, with a written programme and a precise duration.

Warning signs

What you hear	What it may hide	The question to ask
"Our clients are confidential"	No clients	Just one, even anonymised, with what was delivered?
"Our product is being finalised"	A product that does not yet exist	Who uses it today?
"100% sovereign"	A model hosted online	What happens without the internet?
"Everything is included in the subscription"	Code that will never be yours	Is the code handed over to me?
"We'll adjust as we go"	A scope that will never be fixed	Where is the requirements document?
"A demonstration is worth more than a long speech"	No production deployment	May I speak to a user?

None of these signs is disqualifying on its own. Several together should postpone the signature.

How we answer these seven questions

We would rather answer here than in a meeting, so you can check before calling us.

• Clients: sixteen published case studies, named, with what was delivered. Eleven of our AIs are in daily service at clients.
• Execution: on your servers, with no outbound network call. The deployment in detail.
• Code: the bespoke code is handed over to you, and your data stays exportable.
• Validation: referenced requirements, adversarial acceptance, signed certificate. An example of an accepted project.
• Product status: each AI shows its status — in service, delivered on assignment, or prototype.
• Partitioning: enforced before reading, by role.
• Training: five courses, eighteen days and one hundred and twenty-six hours of programme, on your premises and on your data.

Ask these same questions of every provider you consult, including us. It is the only way to compare offers that all look alike on paper.