Most tools that claim Arabic support mean Modern Standard Arabic. MSA is the Arabic of newspapers, official statements, and school textbooks. It is nobody's WhatsApp voice.
A customer in Riyadh types "فين طلبي" or "وش صار على الطلب". A customer in Cairo types "الاوردر وصل فين". Someone on a laptop keyboard types "wesh sar 3ala el order". None of them will get a reply that sounds like a person if the model was tuned on MSA and evaluated on MSA benchmarks.
What Actually Goes Wrong
Comprehension, first. Dialect is not accented MSA. Vocabulary differs, negation differs, question words differ. A model that has mostly seen MSA can misread the intent entirely, and a support bot that misreads the intent gives a confident, fluent, wrong answer.
Register, second. Even when the meaning lands, replying in formal MSA to a casual dialect message reads the way a legal notice would read as a reply to a friend's text. It signals a machine, and it signals a store that could not be bothered.
Arabizi, third. A meaningful share of messages arrive in Latin letters with numerals standing in for Arabic sounds — 3 for ع, 7 for ح. Tools that only tokenize Arabic script treat these as gibberish or fall through to English.
Gulf-Neutral Is a Real Choice
Trying to mimic a specific city's dialect precisely is the wrong target. Get one word wrong and it reads worse than neutral — like someone doing an impression. What works is Gulf-neutral: the vocabulary and rhythm that a Saudi, an Emirati, and a Kuwaiti customer all read as normal, without any of them hearing a costume.
Egyptian matters separately, because a large share of customers and staff in the Gulf write Egyptian and it is different enough that neutral Gulf phrasing reads as foreign to them.
Short Beats Fluent
The most common failure in Arabic support copy is length. A model asked for a helpful reply writes four sentences of courtesy before the information. Real WhatsApp support is short: the answer, one useful detail, an offer to do the next thing. Anything longer reads as a form letter regardless of how good the Arabic is.
This is a design constraint, not a model limitation. Copy the shape of what a good human agent sends, not the shape of what a language model likes to produce.
How To Test It Yourself
Do not evaluate an Arabic support tool with a demo script. Take fifty real messages out of your own inbox — the messy ones, the ones with voice-note transcripts and half-English product names — and run them through. Then ask a native-speaking member of your team one question about each reply: would you have sent this?
That question is a better benchmark than any accuracy figure a vendor will quote you, and it is the one we ask founding-cohort merchants to run during their pilot.