Arabic that customers actually speak: why Gulf dialect matters for AI agents

Three kinds of Arabic in one inbox

Arabic is not a single register. Official documents, news and most of the text AI models learn from are in Modern Standard Arabic (MSA). But people write to businesses in Qatar in Gulf Arabic (Khaleeji), in the dialects of the many Arab nationalities who live here, in English, and very often in a mix of these within the same message.

Qatar's population is also highly international: most residents are not Qatari citizens. An AI agent for a business here will meet Gulf, Egyptian and Levantine Arabic, and many varieties of English, in the same week. Three patterns come up constantly:

• Gulf dialect: «ابي اعرف وين طلبي» (“I want to know where my order is”) rather than the formal «أريد أن أعرف أين طلبي».
• Code-switching: «ممكن تشيك الـ invoice حقي؟», with Arabic grammar around English business words.
• Arabizi: Arabic typed in Latin letters, with numbers for sounds English lacks, such as 7 for ح and 3 for ع, as in “mar7aba”.

What research shows

Researchers at the Qatar Computing Research Institute (QCRI) built AraDiCE, a benchmark of dialect and cultural capabilities covering Levantine, Egyptian and Gulf Arabic, published at the COLING 2025 conference. They found that Arabic-focused models such as Jais and AceGPT do better on dialect tasks than general multilingual models, but that significant challenges remain in identifying, generating and translating dialects, and that performance on dialects still lags behind formal Arabic and English.

Models are improving quickly, and Qatar now has its own Arabic model family, Fanar, developed by QCRI at Hamad Bin Khalifa University; Fanar 2.0 was announced in December 2025. The lesson is not that one model is right for everyone, but that dialect performance has to be measured, not assumed.

Choose a model by testing it on your own messages

Generic benchmarks say little about your customers. A better approach takes a few days:

• Collect 100 to 300 real messages from your channels, with personal data removed, covering dialects, mixed language and Arabizi in roughly the proportions you see.
• Write down the correct outcome for each: the intent, the information to extract, and what a good reply looks like.
• Run several candidate models on the same set, including at least one Arabic-focused model, and score understanding and reply quality separately.
• Have native speakers review the replies. Fluency, tone and politeness matter as much as accuracy in Arabic customer service.
• Keep the set as a regression test, and run it again whenever you change the model or its instructions.

Understand any dialect, answer in the right voice

Understanding dialect and writing in dialect are different decisions. Many organisations want the agent to understand any variety of Arabic but reply in clear, warm formal Arabic, or formal Arabic with a light Gulf touch, because it reads as professional to every Arabic speaker. Consumer brands may prefer a fully Gulf voice. Decide this deliberately, write it into the agent's instructions, and test it with your own customers.

The agent should also reply in the language the customer used, switch when the customer switches, and keep names, numbers and order references exactly as written, whichever script they are in.

Voice channels need extra care

Spoken Gulf Arabic is harder still: speech recognition has to cope with accents, background noise and sentences that change language halfway through. If your agent will answer calls or voice notes, test transcription accuracy on real recordings before anything else, because every later step depends on it.

Where to start

Start with one channel and one high-volume request, such as order status or appointment changes, and build the test set described above. It is the quickest way to learn which model fits your customers. Our AI agents work is built around this kind of evaluation, and the free AI Readiness Check helps you decide where to begin.