// Straight answer
Short answer
Largely yes. Current large language models read Lebanese dialect and Arabizi (Arabic written in Latin letters and numbers) reasonably well, including messages that mix Arabic, English and French. But accuracy varies by phrase, voice notes are harder than text, and mistakes are costly. Test with your own real messages and set clear rules for handing a conversation to a human.
Answered by Tarek Abou Rjeily, AI consultant & builder in Beirut, for businesses in Lebanon and the Gulf.
Modern large language models were trained on a lot of Arabic and on informal writing, and they cope with Lebanese dialect much better than the rule-based chatbots of a few years ago. They can usually follow Arabizi, where '3' stands for the letter ayn and '7' for ha, and they can handle a single message that switches between Arabic, English and French. That mixing is how many Lebanese customers actually write.
In Invisible CRM, a Beirut brokerage tool built by Tarek Abou Rjeily, the owner asks questions in plain words, Arabizi included, such as 'meen badou yest2jer?', and forwards client messages in English, French or Arabizi. A message like 'shu fi 3andak b Jounieh?' is saved as a Jounieh buyer with no budget, because none was given. The system records only what was actually said.
Dialect and spelling are not standardised, so the same word may be written three ways, and a model can misread slang, sarcasm, place names or a number written in an unusual way. Prices and quantities are the dangerous cases: 'mitten' and a stray digit can change a budget completely. Treat anything involving money, dates or commitments as something to confirm, not assume.
Voice notes add a second step. Speech-to-text has to transcribe the audio before a model reads it, and dialect, background noise and code-switching typically lower accuracy compared with typed text. Invisible CRM transcribes voice notes with Whisper speech-to-text before reading them, and its answers cite the messages they came from, so a misheard detail can be traced back and corrected.
Do not accept a demo in clean Modern Standard Arabic. Collect 30 to 50 real, anonymised messages from your own chats, including the awkward ones, and run them through the system you are considering.
Set the handoff rules before launch. A sensible default: the agent can answer routine questions and collect details, but a person takes over for complaints, negotiation, payments, medical or legal questions, anything the model is unsure about, and any message it cannot read confidently. A system that says 'let me check with the team' is better than one that confidently invents an answer. Tarek Abou Rjeily's practice does not build chatbots that answer customers entirely on their own with no human handoff.
You may not need an AI chatbot at all. If most of your messages are the same few questions, the quick replies in the WhatsApp Business app can cover them, with no model and no builder involved.
Generally yes for common phrases. Rare spellings, slang and numbers inside words are where it slips, so test with your customers' actual messages and have the system quote the original text next to anything it extracts.
Usually, and that is a strength of current models. Check that it replies in the language the customer used, and that names, areas and prices survive the language switch intact.
Less reliable than text. Transcription errors, noise and dialect can change meaning, so keep the transcript visible, flag low-confidence parts, and let a person confirm anything involving money or dates.
Not automatically. Gulf dialects differ from Lebanese in vocabulary and spelling, so retest with local messages from Dubai, Riyadh or Doha instead of assuming Lebanese results carry over.
Send the question on WhatsApp and I'll tell you honestly whether it needs building, buying, or fixing by hand.
Ask on WhatsApp →