Every few months a new Arabic language model is announced, and every few months a Saudi leadership team asks us the same question: which one should we use? It is the wrong first question.
Short answer: for most Saudi enterprises, the model is the most replaceable part of the system. What you should decide first is where the system runs, which data it may read, and how you will know it answers correctly in Arabic. Get those right and you can change models in an afternoon.
What "sovereign AI" means in practice
Sovereignty is not a slogan about where a model was trained. For an enterprise it comes down to three questions:
- Where does inference run? In your tenant or a Saudi region, or on a service abroad that you do not control.
- Where does your data go? Prompts, documents and logs are personal or sensitive data under the PDPL as soon as they contain it.
- Who can switch it off? If the provider changes terms, can you move to another model without rebuilding?
Our name, سيادة, means sovereignty, and this is what we mean by it: the client's data stays in the client's environment.
Arabic quality is a measurement, not a feature
Models that score well on general Arabic benchmarks can still fail on your documents: legal phrasing, Gulf dialect in customer messages, mixed Arabic and English in the same paragraph, names written three ways. The only reliable test is your own material.
Before choosing a model we build a small evaluation set from real questions and the passages that answer them, then run each candidate against it. See how we evaluate AI before release.
Retrieval does more work than the model
Most enterprise questions are about your documents, not the world. A grounded system that finds the right passage and cites it beats a larger model guessing from memory. That is why every answer in Hukum-AI resolves to its source passage, and why we build Arabic retrieval from the data up.
A practical decision order
- Pick one workflow and collect fifty real questions with their correct answers.
- Decide where the system must run and what data it may touch.
- Shortlist models that can run there, local Arabic models included.
- Measure them on your set, in Arabic and English.
- Keep the model behind an interface you own, so it can be replaced.
If you want help with step one, start with one workflow.