Many of the processes that leaders in the Kingdom want to improve run in Arabic, or in Arabic and English at once. Letters, forms, customer messages, policies and case files. Before any of them can be handed to an AI system, someone has to answer a plain question: how well does it handle our Arabic? The answer is rarely yes or no. It is a map of what the system can do on its own, and where a person needs to read along.
Arabic support in AI models has improved a great deal. Large general-purpose models handle Modern Standard Arabic well, and there are open and regional models built with Arabic as a priority. For most business writing, the question is no longer whether a model can read and write Arabic. It is how well it handles your documents, your terms and the way your customers write.
What tends to work well
- Summarising and drafting in Modern Standard Arabic.
- Answering questions from your own policies and documents, when each answer points to its source.
- Classifying incoming requests and routing them to the right team.
- Reading typed Arabic documents with OCR, with vision-language models to make sense of layout, tables and stamps.
Each of these is ready for daily operations today: a contact centre drafting replies for an agent to approve, a staff help desk answering from HR policy, a shared inbox sorted before anyone opens it, a finance team reading supplier invoices. Each works best when the output is checked on real samples and every answer links back to its source.
Where to check twice
- Handwriting and poor scans still produce errors. Plan a review step for the pages the model is unsure about.
- Dialects vary. A model fluent in formal Arabic can misread a Gulf voice note, or a customer message typed in Latin letters.
- Names and places, and their spellings between Arabic and English, are a common source of failed searches and mismatched records.
- Hijri and Gregorian dates get mixed, and so do Western and Eastern Arabic digits.
- Diacritics are usually missing, so some words can be read more than one way.
Cost needs checking too. On some models the same content costs more in Arabic than in English, because the text is split into more pieces before the model reads it. Measure it on your own text before you size a budget. Check speed as well. A model that answers well but slowly may not fit a service counter or a call centre.
Test on your own material
Public benchmarks are a starting point. They rarely use your forms, your terminology or your customers' way of writing. Before you choose a model, build a small test set from real cases: a few dozen documents or questions, with answers your experts agree on. Include the hard cases on purpose: the worst scan, the longest name, the message that mixes two languages. A test set of clean examples tells you what you want to hear.
Run each candidate against it, in the hosting setup you would actually use. Then read the errors, not only the score. The pattern of mistakes tells you where a person needs to stay in the process, and where the model can work on its own.
Where hosting meets language
Language and hosting are often decided by different teams, and they should not be. If your data must stay in the Kingdom, test the models you can actually host there, rather than the ones at the top of global leaderboards. The best Arabic model you cannot deploy is not an option. Ask also who decides when the model version changes, because a new version can read your documents differently. Keep the test set, and run it again before every upgrade.
Arabic first, not Arabic later.
When a system is tested in English first and in Arabic later, the hardest cases stay hidden until go-live. Test in Arabic from the first week, on your own material, and you will know early where the system is ready and where a person still has to read the page.
Testing models on your own Arabic material is part of our practice in data, sovereign platforms and technology sourcing. Putting them to work in daily operations is the job of our practice in operations and service re-engineering with AI.
