Most people meet AI through a chat window in a browser. The model runs on someone else's servers, in another country, under another country's law. For a student or a small shop, that is fine. For a ministry, a hospital or a regulated company in Oman, it often is not.
This is why we started Hosn, a BHD Group company that installs AI at the customer's own site in Oman. This essay explains when on-premises AI makes sense, and what an institution should put in writing before it buys any AI system.
Start with the data
The first question is not "which model is best". The first question is "where can this data go".
Some data cannot leave the building. Patient records, case files, internal correspondence and anything classified fall in this group. If the data cannot leave, then the AI must come to the data. That means a model that runs on hardware inside the institution, on its own network.
Some institutions go one step further and need an air gap: no connection to the internet at all. Then updates, new models and logs need a written procedure, because nothing arrives by itself.
Test the Arabic first
A model that writes good English can still write weak Arabic. It can misread Omani names, mix up dates, or answer a formal letter in the wrong register. So Arabic must be tested on the institution's own approved material, before the model is chosen.
At Hosn we test Arabic and English behaviour on representative samples first. Then we write the acceptance criteria down. We choose the model only after that test.
Write three boundaries down first
Before any purchase, an institution should have three things in writing:
- The site boundary. Where the system runs, which network routes it uses, where data is stored, and who administers it.
- The language boundary. Which Arabic and English tasks it must do, tested on real samples, with criteria that decide pass or fail.
- The commercial boundary. What the price covers, what support costs, and how updates arrive.
If a vendor cannot write these three down, the institution is buying a demo.
Start small, size from real demand
A sensible first step is one department and one workload: for example, search across an approved document library in Arabic and English. Measure how people use it. Then size the hardware from that measured demand, not from a forecast.
What this means for Oman
Oman has a clear interest in AI that works in Arabic and keeps national data inside the country. On-premises AI is one way to get both. It is not the answer for every task.
More on Hosn: hosn.om.