Arabic & MENA
Arabic AI: why dialect matters more than Modern Standard
A model that handles formal written Arabic will still fail a customer typing Lebanese. Nobody writes Modern Standard Arabic to a shop. If your assistant only handles MSA, it will be abandoned, and the failure is invisible in testing because the people testing it write formally.
There is a gap between “supports Arabic” and “works for Arabic-speaking customers”, and almost every vendor claim sits on the wrong side of it.
We shipped an Arabic assistant into a Lebanese retail business: text and voice, taking real customer orders. Here is what we actually ran into.
Nobody writes Modern Standard Arabic to a shop
Modern Standard Arabic is the language of news broadcasts, official documents and formal writing. It is what models are most heavily trained on, because it is what most written Arabic text on the internet is.
It is not what anyone types into a chat window.
A Lebanese customer asks for the cheapest olive oil in Lebanese. An Egyptian customer asks in Egyptian. These are not accents, they are substantially different in vocabulary and structure. A system tuned for MSA handles the formal phrasing and then stumbles on the way people actually speak.
The cruel part: this failure is invisible during testing. The people testing your system are usually staff who know it is being evaluated, so they write carefully and formally. It passes. Then real customers arrive writing the way they text their friends, and it falls apart.
Test with real messages, not written ones
The single most useful thing we did was build the evaluation set out of actual customer messages, taken from the WhatsApp conversations the business was already having before any of this existed.
That gave us the real distribution: the abbreviations, the missing vowels, the switch into French or English mid-sentence, the voice notes.
If you are commissioning Arabic AI, ask what the evaluation set is made of. If the answer is “we wrote some test questions”, the system has not met your customers yet.
Arabizi is not an edge case
A large share of messages arrive in Latin characters with numerals standing in for Arabic letters. The 3 for ع, the 7 for ح.
This is not rare, and it is not a niche younger habit. Plan for it. A system that only accepts Arabic script will silently fail a meaningful fraction of your customers, and they will not tell you, they will just stop.
Code-switching is the default
One message will contain Arabic, French and English. This is completely normal in Lebanon and across much of the Gulf, and it is normal within a single sentence, not just within a conversation.
Anything that tries to detect one language and route accordingly will make the wrong call constantly. The system has to handle the mix as the expected case.
Voice changes the problem
Voice notes are how a lot of people prefer to communicate, and they carry dialect much more strongly than text. Someone who types semi-formally will speak completely colloquially.
If you support voice, you have committed to dialect, whether or not you planned to.
Right-to-left is a build decision, not a translation task
This is where a lot of “we support Arabic” claims fall over.
Proper RTL is not a mirrored stylesheet. It touches:
- Layout direction, and which elements should not flip (logos, media controls, numbers)
- Number formatting, and whether to use Arabic-Indic or Western numerals, which varies by market
- Mixed-direction strings, where a Latin product name sits inside Arabic text
- Search that tolerates spelling variation, since the same word is written several defensible ways
- Fonts that actually render Arabic well at small sizes
Retrofitting all of that after building left-to-right costs more than doing it first. We build Arabic-first when the primary audience is Arabic-speaking, and add English as the second direction.
What this means commercially
Two things worth knowing if you are buying:
The bar is low, so the advantage is real. Most regional competitors claim Arabic support and ship MSA-only. If your assistant actually handles how your customers speak, that is a difference customers feel immediately.
Cost control matters more here. Dialect handling often needs a larger model or more context, which costs more per call. Meter it. On our build, assistant usage is capped per customer and tied to loyalty tier, so the owner knows the monthly ceiling before it arrives.
The questions to ask a vendor
- Is your evaluation set made of real customer messages, or ones you wrote?
- Which dialects, specifically? “Arabic” is not an answer.
- What happens with Arabizi?
- What happens when one message contains three languages?
- Show me the RTL interface, not a description of it.
We have shipped this, not just scoped it. If you want to see how the Lebanese assistant was built, the case study is here, and we work across Lebanon, the UAE and Saudi Arabia.