Which restaurant chatbots actually understand Arabic dialect and Franco-Arabic?

Most chatbot tools advertise "Arabic support", and mean Modern Standard Arabic written in Arabic script. But restaurant customers in Egypt and the Gulf don't write in Modern Standard. They write in dialect, they mix English into Arabic, and they write in Franco-Arabic. Those are three separate gaps, and "Arabic support" closes none of them. At the end of this article is a five-message test that exposes any tool in minutes.

Gap one: Arabic is not one language in practice

Gulf, Egyptian, Levantine, and Maghrebi Arabic read and sound differently. A tool tuned on one dialect may stumble on a customer from another region, and the difference isn't accent, it's vocabulary: ordering words, quantities, and dish names differ between markets.

The practical effect is that a tool proven in the Gulf market can behave quite differently with an Egyptian customer, even with "Arabic support" enabled in both cases.

Gap two: language mixing inside a single message

Customers routinely write messages like "عايز 2 large بيتزا delivery للمعادي". This is not a language error, it is normal bilingual communication. But many NLP engines treat mixing as confusion, or try to apply a single language model to the whole message, and lose half the meaning.

Gap three: Franco-Arabic

Franco-Arabic is Arabic written in Latin characters, with digits standing in for letters that have no Latin equivalent: 3 for ع, 7 for ح, 2 for hamza. A message like "3ayez 2 burger" contains two digits with two different meanings: the first is a letter, the second is a quantity.

More importantly, Franco-Arabic was never standardised, so there is no official orthography to normalise against. The same word may be written "3ayez", "3ayz", or "aiz" by three customers, all correctly. A tool that matches fixed strings will necessarily fail here.

Where the current tool categories stand

• Rule-based flow builders (ManyChat, Landbot): rely on buttons and text matching, so they avoid the problem rather than solve it. Fine while the customer taps buttons, brittle the moment they type freely.
• Shared inboxes (Tidio, Freshchat): don't attempt understanding at all; language is the human agent's problem, not the tool's. That's a legitimate choice if you have staff to answer.
• Agent platforms (Botpress): do provide genuine Arabic understanding, but tuning it to a dialect and handling mixing and Franco-Arabic remains engineering work you own. There are public reports of difficulty extracting values from Arabic sentences in multilingual configurations.
• Purpose-built systems (Ordr, and us): vary widely. The deciding factor is whether they were built in an Arabic market or translated into one afterwards. Ordr, for one, advertises voice-note transcription in Egyptian and Khaleeji dialects, which is a far more specific claim than "Arabic support" and worth testing rather than assuming.

The five-message test

Don't trust the product page. Send these five messages to any tool you're evaluating and judge for yourself, since each exposes a different gap:

1. "3ayez 2 burger w batates": did it understand that 3 is a letter and 2 is a quantity?
2. "عايز ٢ برجر من غير مخلل": dialect, Arabic-Indic numerals, and an exclusion request.
3. "فيه حاجة من غير مكسرات؟": a question about an ingredient, not an order. Does it answer, or try to add an item?
4. "عايز 2 large بيتزا delivery للمعادي": two languages in one sentence, plus an address.
5. "الأول قولي بتوصلوا فين، وبعدين هطلب": deferred intent. Does it respect the sequence or jump to ordering?

Any tool that handles all five correctly deserves serious consideration. Any tool that fails the first or the fourth will fail with a large share of your real customers, whatever its page says about "Arabic support".

Our position on this

We built Redood around exactly these gaps: the agent detects which register the customer is writing in, whether Modern Standard, Egyptian colloquial, English, or Franco-Arabic, and replies in the same one. We've tested it internally against real menus, but we are pre-launch and have no production accuracy data yet, and we won't claim otherwise.

So the honest recommendation is: run the five-message test on every tool you're considering, including ours. You can run it on your own menu, free and without signing up.