We tested a competitor's Arabic-language banking chatbot a while back — not maliciously, just curiosity about the market — by asking it a few questions a real customer would actually ask, in the way a real customer would phrase them, including some Gulf-dialect colloquialisms. It handled formal Modern Standard Arabic reasonably well. The moment the phrasing got even slightly conversational or regionally inflected, the responses became noticeably worse, sometimes nonsensically so. The vendor's website listed "Arabic support" as a checked box, same size font as every other feature. The actual experience underneath that checkbox told a very different story.
Why this gap exists so often
Most AI systems built primarily for English-speaking markets and then extended to "support Arabic" do exactly that — extend, as an afterthought, usually via translation. The underlying model reasons in English, a translation layer converts the input from Arabic and the output back into Arabic, and the whole system is fundamentally an English-first system wearing an Arabic interface. This works fine for simple, formal interactions. It falls apart quickly for anything that requires understanding regional dialect, cultural context, or the genuine ambiguity that comes with natural spoken Arabic, which varies meaningfully across the Gulf, the Levant, and North Africa.
What "Arabic-first" actually requires
Building genuine Arabic-first AI means the system is trained and evaluated on Arabic as a primary language, not a translated afterthought — including right-to-left interface design done properly (not just flipped CSS), Gulf dialect understanding specifically rather than generic Modern Standard Arabic, and crucially, evaluation done by native Arabic speakers checking actual conversational quality, not just whether a translation is technically accurate. A translated sentence can be grammatically correct and still sound like nobody who actually speaks the language would say it that way.
Why this matters more in the UAE than almost anywhere
The UAE's population is famously diverse, and English functions as a genuine lingua franca here in a way it doesn't in many other Arabic-speaking markets — which paradoxically makes it easier for AI vendors to under-invest in Arabic quality, because the system can appear to function fine when most of the visible testing happens in English. But a meaningful share of customers — citizens, long-term residents, and visitors from across the region — will engage in Arabic specifically, and for a bank or any consumer-facing service, a customer who gets a noticeably worse experience in their preferred language is a real service gap, even if it's invisible in an English-language product demo.
How to actually test for this before you buy
Don't evaluate Arabic capability with formal, written-out test questions a vendor's own QA team probably already tested against. Find a native Gulf Arabic speaker — ideally someone who isn't on the deal team and has no incentive to be generous — and have them have a genuinely casual conversation with the system the way a real customer would, including switching between Arabic and English mid-conversation, which happens constantly in real UAE usage. Watch whether the quality holds up or degrades the moment the interaction gets less formal and more natural.
We build Arabic support this way by default — native Gulf dialect handling, proper RTL interface work, and genuine bilingual switching rather than two separate single-language modes bolted together — not because it's a differentiator we like to mention, though it is, but because anything less doesn't actually serve a meaningful share of the people who'll be using these systems. "Supports Arabic" should mean something specific. Right now, across the market, it often doesn't, and the only way to find out before you've signed a contract is to actually test it properly.