Building AI interfaces that are Arabic-first, not Arabic-translated

Translating button labels is the easy part. The hard part is that your documents, your users' questions and your model's reasoning all mix Arabic and English freely — and most retrieval pipelines silently degrade the Arabic half.

We benchmark every embedding model on mixed Arabic-English corpora before it goes anywhere near production. The differences are not subtle: retrieval quality on Arabic legal and technical text varies by a factor of three between models that look identical on English benchmarks.

Interface direction matters more than teams expect. An assistant that answers in Arabic inside a left-to-right layout reads as an afterthought. Directionality, typography and date formats have to switch together, driven by the conversation's language rather than a global setting.

The UAE market rewards teams that get this right. Users forgive a missing feature; they do not forgive being made to feel like the secondary audience of their own government's or employer's systems.