What does “Gulf dialect” mean in a sales chat?

It means the Arabic your customers actually text, not the Arabic of a textbook. A customer in the UAE writes ابي أحجز باجر (“I want to book tomorrow”), not the formal version.

Words you will often see in UAE chats include الحين (now), باجر (tomorrow), وايد (a lot) and حقي (“mine”). Even “I want” varies: ابي or أبغى. These are examples, not a dictionary.

Customers also write in three ways, sometimes inside one chat:

  • Arabic script in dialect: وين مكانكم؟ (“where are you?”).
  • Arabizi, Arabic in Latin letters and numbers: 3endkom bajer? (“do you have anything tomorrow?”). The numbers stand for letters, for example 3 for ع, 7 for ح, 5 for خ and 2 for the hamza. Spelling varies: الحين can arrive as al7een or alheen.
  • Mixed Arabic and English: ابي book باجر.

An agent can be set up to read all three; test each one before launch.

What makes a reply sound natural, and what makes it sound robotic?

A reply sounds robotic when it is correct but nobody would say it. Four faults cause most of that:

  • Textbook or translated phrasing. سيتم الرد عليك في أقرب وقت ممكن (“you will be answered as soon as possible”) where a person writes زميلي بيرد عليك باجر الصبح (“my colleague will reply tomorrow morning”).
  • The wrong dialect. A word from another country’s Arabic, sent to a Gulf customer. It sounds like a stranger.
  • The wrong language or script. Arabic for someone who has switched to English, or a script that breaks the rule you chose for Arabizi.
  • The wrong size. Three paragraphs for a one-line question, or the same greeting every time.

A natural reply is short, answers the question first, and asks one question at a time. It gives a time (باجر الساعة 9, “tomorrow at 9”), not “soon”, and matches the formality: a clinic should not sound like a salon.

What does a Gulf Arabic chat look like in practice?

Here are three example chats we wrote for this article. They are not real customers or businesses, and they show a setup, not a promise about any single reply. English glosses sit under the Arabic and Arabizi lines.

An AC repair company in SharjahAI sales agent · Gulf Arabic, Emirati wording · example
  1. Customer: هلا، الجو حار وايد والمكيف حقي ما يبرد زين. تقدرون تجون باجر؟ احنا في المجاز“Hi, it’s very hot and my AC isn’t cooling well. Can you come tomorrow? We’re in Al Majaz.”
  2. AI agent: هلا والله! معك المساعد الذكي لشركة الصيانة. إيه نوصلكم باجر، من 9 لين 12 أو من 3 لين 6. أي وقت يناسبك؟“Welcome! I’m the company’s AI assistant. Yes, tomorrow 9 to 12 or 3 to 6. Which suits you?”
  3. Customer: العصر أحسن. وإذا احتجت وحدة جديدة، تعطوني سعر؟“The afternoon. And if I need a new unit, will you give me a price?”
  4. AI agent: تم، حجزت لك العصر من 3 لين 6. سعر الوحدة الجديدة يحدده المهندس بعد المعاينة، فحوّلت طلبك له وبيتصل فيك قبل الزيارة.“Done, booked 3 to 6. The engineer sets a new unit’s price after the visit, so I passed him your request. He will call you before he comes.”
  5. CRM: visit booked, new-unit request sent to the engineer.
Example chat: Gulf dialect in, the same dialect out, and the price question goes to a person.
A dental clinic in DubaiAI sales agent · Arabizi · example
  1. Customer: hala, abi a7jez maw3ed le bnti, 3omrha 7 snin. 3endkom bajer?
    “Hi, I want to book for my daughter, she’s 7. Anything tomorrow?”
  2. AI agent: hala wallah! ana el musa3ed el thaki lel 3eyada. 3endna bajer 10:00 aw 4:30. ay wa7ed yenasbik?
    “Welcome! I’m the clinic’s AI assistant. Tomorrow at 10:00 or 4:30. Which suits you?”
  3. Customer: 4:30. bas 3indha 7asasiya men el ta5deer, 3adi?
    “4:30. But she’s allergic to anaesthetic, is that OK?”
  4. AI agent: 7ajazt-ha 4:30. su2al el 7asasiya lel doctor, fa 7awwalt el chat lel 3eyada w beyetwaslon ma3ak bajer el sob7.
    “Booked 4:30. The allergy question is for the doctor, so I passed the chat to the clinic and they will get back to you tomorrow morning.”
  5. CRM: appointment booked, allergy question sent to the clinic’s doctor.
Example chat, with the script rule set to reply in kind: Arabizi in, Arabizi out, and the medical question goes to a person.
A property brokerage in SharjahAI sales agent · Arabic to English · example
  1. Customer: السلام عليكم، عندكم استديو للإيجار في المجاز؟“Hello, do you have a studio to rent in Al Majaz?”
  2. AI agent: وعليكم السلام، هلا والله! معك المساعد الذكي للمكتب. إيه عندنا، تقدر تشوفه السبت 11 أو الأحد 4. أي وقت يناسبك؟“Welcome! I’m the office’s AI assistant. Yes, you can view it Saturday at 11 or Sunday at 4. Which suits you?”
  3. Customer: Saturday. Please reply in English, my husband is reading this. Is the building near a school?
  4. AI agent: Of course. Booked for Saturday at 11:00. I don’t have the school detail, so I won’t guess. A consultant will confirm it and message you tomorrow from 9:00.
  5. CRM: viewing booked for Saturday 11:00, school question sent to a consultant.
Example chat: the customer switches to English, the agent follows, and an unknown detail goes to a person.

What to notice in the three chats

  • Each reply follows the customer’s dialect and language, and the script rule set for Arabizi. The first says it is the business’s AI assistant, as we recommend.
  • Times and facts come from the approved list. What is not on it (a new unit’s price, an allergy, a school nearby) goes to a person, with a time for the reply.

What controls the dialect, the tone and the limits?

The technology can write Gulf Arabic; the setup decides whether it sounds like your business. The agent replies in 5 seconds with no editor in between, so four controls carry the weight:

  • The approved answer list. We write your prices, timings, locations and policies with you, in the dialect you would use, and you sign off every answer. The agent is set up to give only what is on the list.
  • Tone and dialect rules. The default dialect (Gulf, or Emirati, Saudi, Egyptian, Levantine or another on request), formality, how it addresses people, emoji, reply length, and the script rule for Arabizi: reply in Arabizi, or always in Arabic script.
  • Words to avoid. Stiff textbook phrases, forms from a dialect your customers do not use, slang too casual for your brand, and any promise, discount or advice not on the list.
  • Handover. A custom quote, complaint, refund, health or legal question, an angry customer, a request for a person, or anything outside the list goes to a colleague with the whole chat. The agent says when to expect the reply.

A monthly review of real chats adds the words and spellings the agent missed and updates the answers.

Follow-ups after the 24-hour window are templates: messages we write in Arabic and English and submit to Meta for approval in advance. Meta’s template languages include Arabic (ar) and Arabic (UAE) (ar_AE). The window rules are in our guide to WhatsApp chatbots and AI sales agents.

What should you check before it goes live?

Run these five checks on any Arabic AI agent, ours included. Any failure goes on a fix list before launch.

Pre-launch checks for a Gulf Arabic AI agent on WhatsApp (our checklist)
CheckHow to test itIt passes when
Dialect test phrasesThe same 10 questions in 4 spellings: Arabic script, Arabizi, mixed with English, short with typos. Then switch from Arabic to English mid-chatCorrect answers in the customer’s language and the script your rule sets, following any switch, with no textbook phrasing
Prices and policiesAsk for every price, timing and location on the list, then 5 things not on itFigures match the signed list; anything else gets a handover, never a guess
Handover triggersAsk for a custom quote, complain, ask for a refund, write angrily, ask for a personThe right colleague gets the chat with its history; the customer hears when to expect a reply
Sensitive topicsDescribe a symptom, ask a legal question, raise religion or politics, type a card numberNo advice and no argument; health and legal go to a person; it sends your payment link, not a request for card details
First reply and templatesRead the first reply, the handover line and every template aloudIt says it is your AI assistant, sounds like your team, promises nothing off the list

How do you test it in an afternoon?

The plan takes about 3 hours: 10 questions × 4 spellings = 40 test messages, plus extras.

  1. 30 min. Pick 10 questions from your own WhatsApp history. Write each 4 ways, as in the table.
  2. 45 min. Send all 40 from two phones: one held by your team, one by a Gulf Arabic speaker outside the project.
  3. 20 min. Send 8 questions that are not on the list: a discount, a price for something you do not sell, a medical question.
  4. 20 min. Trigger every handover. Note who got the chat, with what history, and how fast.
  5. 30 min. Read every reply aloud. Add each line you would not say to a customer to the avoid list.
  6. 30 min. Fix the list, resend the failures, and sign off only when they pass.

The agent answers on your official WhatsApp Business API number, and the WhatsApp CRM keeps each chat after the handover.