Available --:--local Start a project
← All articles

( · 4 min read )

Arabic AI chatbot for business: understanding dialects and Darija, not just MSA

How to build an Arabic AI chatbot for business that handles Modern Standard Arabic, Gulf and Egyptian dialects, Moroccan Darija, Arabizi and voice notes, for customer support on WhatsApp and the web.

  • Arabic AI
  • Chatbots
  • LLMs

Most "Arabic chatbots" are English chatbots with a translation layer. They work in a demo, then fail on the first real customer who writes "salam, bghit n3ref wach kayn livraison l Tanger?" That's Moroccan Darija in Latin letters, and it's a completely normal message.

If your customers are in the Gulf, Egypt or the Maghreb, this is the gap between a chatbot people use and one they abandon. Here's how I approach building Arabic AI assistants that work with how people actually write.

How customers in the region really write

A single support inbox in the MENA region sees all of these:

  • Modern Standard Arabic (MSA), mostly in formal emails and from older customers.
  • Dialects: Gulf (Khaleeji), Egyptian, Levantine, and Maghrebi dialects like Moroccan Darija, which can be hard for other Arabic speakers to follow.
  • Arabizi: Arabic written in Latin letters, with numbers for sounds Latin doesn't have (3 for ع, 7 for ح, 9 for ق).
  • Code-switching: Arabic, French and English mixed in one sentence, very common in Morocco, Algeria, Tunisia and Lebanon.
  • Voice notes on WhatsApp instead of text.

A chatbot that only handles MSA is handling the smallest part of that list.

Why off-the-shelf models struggle

Large language models have improved a lot in Arabic, but their strength is uneven. They're strongest in MSA, weaker in dialects, and weakest in dialects written in Arabizi. They also tend to answer in MSA even when the customer wrote in dialect, which feels stiff and robotic to the person reading it.

Arabic adds technical quirks too: it often takes more tokens than English for the same meaning (so it costs more per message), right-to-left text breaks some interfaces, and numbers, dates and currency formats vary by country.

None of that is a reason to avoid Arabic AI. It's a reason to design for it on purpose.

What a good Arabic assistant needs

1. Detect the language and dialect first

Before answering, the system works out what the customer wrote: MSA, which dialect, Arabic script or Arabizi, and which other languages are mixed in. That decides how it searches your knowledge base and how it replies.

2. Answer in the customer's register

If someone writes in Egyptian Arabic, a reply in friendly Egyptian Arabic (or simple MSA) lands better than formal MSA. The tone rules belong in the system's instructions and get tested with real examples from your inbox.

3. Ground every answer in your own content

The assistant should answer from your policies, product catalogue and past tickets, not from general knowledge. This is retrieval-augmented generation (RAG): the system finds the relevant documents first, then writes the answer from them. For Arabic, the search step needs care, because the same word can be spelled several ways across dialects and scripts. Normalising spelling and searching by meaning rather than exact words fixes most of it.

4. Handle voice notes

Speech-to-text models now transcribe Arabic voice notes reasonably well, including many dialects. Transcribe, then treat the result like any other message.

5. Know when to hand over to a human

The assistant should pass the conversation to a person, with a short summary, when it's unsure, when the customer is upset, or when money or personal data is involved.

6. Test with your real messages

The only test that matters is your own inbox. I build an evaluation set from a few hundred anonymised real conversations in every dialect your customers use, and every change to the assistant is measured against it.

Putting it on WhatsApp

For most businesses in the region, the chatbot lives on WhatsApp, which means the official WhatsApp Business Platform (the API), not an automated personal account. Since January 15, 2026, Meta's terms bar general-purpose AI assistants from that platform, but bots that serve a business's own customers (support, orders, bookings) are still allowed. So scope the assistant to your business and give it a clear handover to a person. I cover the details in WhatsApp AI chatbots for businesses in Morocco.

Cloud API or a model on your own servers?

Both work. Cloud APIs from the large providers give the strongest general Arabic today and are fastest to start with. Open models, including Arabic-focused ones, can run on your own servers, which keeps customer data inside your country (relevant under data laws like Saudi Arabia's PDPL or Morocco's Law 09-08, which I cover in private AI and data protection laws in MENA) and makes costs predictable at high volume. Many good systems use both: a local model for routine messages, a larger cloud model only for the hard ones.

Working together

I'm Driss Briksine, a software and ML engineer based in Morocco. I'm a native Arabic speaker, I work in Arabic, English and French, and I build AI assistants and automation for companies across the MENA region. See my work, or read about where AI automation pays off first for MENA businesses.

If your customers write to you in Arabic, tell me about your inbox and I'll show you what an assistant could handle.