Will an Arabic AI chatbot really understand your Gulf Arabic clients? A test you can run

Why does "supports Arabic" on a vendor page mean so little?

An AI chatbot is software that reads a message and writes a reply using an AI language model. "Supports Arabic" can mean anything from "translates the English reply into formal Arabic" to "reads and writes natural Gulf dialect and handles mixed language". Vendor pages that claim tuning for several dialects give no evidence, and I could not verify any vendor's dialect accuracy number. Treat them all as unverified until you test.

Independent research exists, but it is general, not real estate. A 2025 academic benchmark called AraDiCE tested dialect and cultural ability across Gulf, Egyptian and Levantine Arabic, and found that Arabic-focused models did better on dialect tasks than general multilingual ones, while dialect generation remained a challenge. Another academic dataset, DialectalArabicMMLU, covers five dialects including Emirati and Saudi and reports large differences between dialects. I read the abstract of AraDiCE and saw DialectalArabicMMLU only through a search summary, and I did not read the full papers, so read them as context: dialect is a known weak spot, and nobody has measured it on property chat.

Formal Arabic, Gulf dialect and the other dialects your clients speak

Formal Arabic (Modern Standard Arabic, or MSA) is what is written in news and contracts. Gulf dialect, often called Khaleeji, is what many clients actually type and say. A bot that answers a relaxed Khaleeji message in stiff MSA is understandable but feels cold, like a legal notice. One vendor-written article says general AI models default to formal Arabic unless told to match the dialect. That is a vendor statement, not a study, but it matches what the test below is designed to catch.

Dubai also buys from Egypt, the Levant, Saudi Arabia, Kuwait, Iraq and North Africa, so "Emirati Arabic" is only one of the dialects your leads use. You do not need a bot that sounds Emirati. You need one that understands all of them reasonably and replies in clear, polite Arabic that does not insult anyone.

Arabizi and mixed Arabic and English

Arabizi is Arabic written in Latin letters, with numbers standing in for letters that have no English match, such as 3 or 7. Mixed messages put English property words inside an Arabic sentence, for example asking for a studio in Marina with a budget in English numbers. This is normal in Dubai. A bot trained mainly on formal text can misread both. Ask any vendor to show you these cases live, not in a slide.

Names, numbers and money: where Arabic bots slip in real estate

• Place and developer names: the same community or developer is spelled several ways in Arabic and English. If the bot does not match both spellings to one listing, it gives wrong answers.
• Numbers: Arabic-Indic digits and Latin digits, and spoken forms for money such as "alf" (thousand) and "million". A budget of "900 alf" must become 900,000.
• Currency: dirham versus dollar. Never let the bot assume.
• Payment plan phrases: splits like 60/40, or post-handover plans in Arabic wording.

Money and numbers are where a mistake costs you real trust, so put them in the test.

Right-to-left and WhatsApp formatting checks

Arabic runs right to left. In a chat that mixes both directions, numbers, links and English names can end up in the wrong order or a bracket can flip. Check this on a real phone, in WhatsApp and in any web chat widget on your site, not only in the vendor's screenshot. Look at links, listing references and phone numbers in particular.

The 10-message test script to run on any demo

Run these in order, in the same chat, on the vendor's real system, and write down what happens. Have an Arabic speaker present, ideally someone who works with clients, not only a translator.

1. Formal greeting in MSA. Does it answer in matching formal Arabic?
2. Khaleeji greeting and a relaxed request using a dialect word such as "abi" (I want). Does it reply in a natural tone, not a stiff one?
3. An Arabizi message with number-letters. Does it understand and reply sensibly?
4. Arabic with English property terms, for example a studio request with an English budget figure. Does it keep the facts straight?
5. An area name spelled two ways, once in Arabic and once in English. Does it treat them as one place?
6. Number formats: Arabic-Indic digits, then Latin digits, then "900 alf". Does the budget come out right?
7. A voice note. Does it listen, or does it admit it cannot and offer another route?
8. An angry complaint in dialect. Does the tone stay calm and does it offer a person?
9. A request to speak to a human. Does it hand over at once, with context?
10. A mid-chat language switch from Arabic to English and back. Does it follow without confusion?

Score each message on four things: understood the intent, replied in the right dialect and script, kept every fact correct, and handed over when asked. A vendor who will not run the test live on your phone has told you something.

What the research says, and what vendors cannot prove

To summarise honestly: independent dialect benchmarks exist, and they show uneven performance. None tests real estate chat. Vendor numbers such as "90 percent of conversations are in dialect" or "74 percent accuracy on Gulf speech" came without a named study, a date or a method on the pages I read, and one page also mixed up two different error measures. If a salesperson quotes an accuracy figure, ask for the test set, the date and who ran it. Without those, it is marketing.

Decide your reply language policy before the pilot

Three choices matter more than the model. First, reply in the client's language and script, or always in Arabic script? A reply in Arabic script to an Arabizi message is a choice, and your vendor should let you configure it. Second, tone: formal for high-net-worth buyers and developer-led clients, relaxed for residents. Third, what happens at doubt: the bot should say it will ask a colleague rather than guess. Write these three decisions down.

Human review and handover for Arabic conversations

For the first month, have a native speaker read a sample of live Arabic conversations every week, ideally the first 200 or all that the bot marked as low confidence. Keep a short "bad phrases" list and feed it back to the vendor or into your instructions. Meta, the company behind WhatsApp, also expects automated replies to come with a prompt, clear route to a human, so handover is a rule of the platform, not only good manners. A human handover with the whole conversation attached is the single best safety net for weak Arabic.

Voice notes and calls

Speech recognition in dialect is harder than reading text. This guide is about chat. For calls, see answering calls in Arabic, English and Hindi and finding a native-sounding AI voice. For a voice agent on portal calls, see the AI voice agent guide.

How software and AI help here, and where a native speaker must stay

• Understanding and replying: the AI model reads mixed messages and drafts replies. A person sets tone and the boundary list.
• Name matching: software keeps a table of your communities and developers with every spelling, so Arabic and English names map to one listing. A person maintains the table.
• Quality control: software flags low-confidence chats and keeps the bad-phrase list. A native speaker reads them.
• Anything involving price, advice or promises: always a human.

To try the test yourself, use Claude, an AI assistant, to build your message set. Paste three anonymised real client messages (one Arabic, one Arabizi, one mixed) and ask, "Write 10 similar test messages in the same styles about buying a studio in Dubai, including numbers written as alf and million, and two spellings of the same area." Then run those against every vendor demo. Do not ask Claude to grade the vendor's Arabic for you; a native speaker should.

How to pilot before paying

Decide how much Arabic you really get. If more than about a fifth of your enquiries are in Arabic, make Arabic quality a pass or fail gate. That threshold is my own judgement, not a sourced figure. If it is only a few, an English-first bot with a clean handover to an Arabic-speaking agent beats a weak Arabic bot. Then run a 30-day pilot with a fixed fee and a stop rule, and read how the same bot works in a real chat flow in the WhatsApp AI agent guide. On privacy, ask where chat data is stored and processed, and whether the vendor uses it to train its models; the UAE personal data law (Federal Decree-Law 45 of 2021) is in force, but confirm with a lawyer what that means for your set-up.

How would I approach it?

I would run the 10-message test on every shortlisted vendor, on my own phone, with a native speaker beside me, and drop anyone who fails numbers or handover. I would choose English-first with a good handover over a bot that sounds confident and is wrong. And I would put the Arabic review into a weekly routine for the first month, with one named person responsible.