What it takes to run an Arabic AI agent on WhatsApp, Gulf dialect included

A customer in Riyadh sends a voice note at 22:50. Another types “abi maw3ed bukra” in Latin letters. A third starts a sentence in Arabic and finishes it in English. All three want the same thing, an appointment tomorrow, and an agent that was built and tested only on textbook Arabic can stumble on every one of them.

This guide is for owners and operators in the Gulf and Türkiye who want an AI agent to answer customers in Arabic on WhatsApp. It covers why WhatsApp is the channel, what makes Arabic hard for AI, what Meta's WhatsApp Business Platform requires, how to test an Arabic agent before it talks to a single customer, and where the data should live. If you want the basics of what an agent is first, start with AI agents for business.

Why should an Arabic AI agent work on WhatsApp?

Because that is where your customers already write. In the Saudi Internet 2025 report from the Communications, Space and Technology Commission, WhatsApp was the most-used app in Saudi Arabia, used by 94.4% of men and 90.4% of women, ahead of Snapchat and TikTok. In Türkiye, TurkStat's 2026 household ICT survey found that 92.3% of people aged 16 to 74 used the internet and that 90% of internet users used WhatsApp, ahead of YouTube at 77.6% and Instagram at 71.1% (Forbes Türkiye).

For a business, that means the enquiry, the booking, the complaint and the follow-up all arrive in one thread, at any hour, in the customer's own words. A chat widget on your website or a contact form asks the customer to go somewhere new. An agent on WhatsApp meets them in the app they opened to talk to their family. The difficulty is that in the Gulf, the customer's own words are rarely textbook Arabic.

What makes Arabic hard for an AI agent?

Five things separate a real Arabic WhatsApp inbox from the clean text most demos use. Each one needs a decision in the design and a place in the test set.

1. Dialects, not one Arabic

Modern Standard Arabic (MSA) is the Arabic of news, contracts and signs. Customers write the way they speak, and that changes from country to country and city to city. The word for “what” alone can be وش, شو, شنو or ايش depending on where the customer grew up. “I want” is أبي in much of the Gulf, not the textbook أريد. A Kuwaiti customer may write “tomorrow” as باچر, with a letter that is not in the standard Arabic alphabet.

Research benchmarks show that this is still hard for machines. The NADI 2023 shared task, part of a long-running research competition in Arabic dialect identification, asked systems to tell which of 18 countries a tweet's dialect came from. The best system reached a macro-F1 score of 87.27 out of 100, so even the winner confused dialects on part of the test set. The same edition ran two tasks translating Egyptian, Emirati, Jordanian and Palestinian Arabic into MSA, and the organisers concluded that all three tasks remain challenging.

Large language models face the same problem. AraDiCE, a benchmark presented at COLING 2025, found that Arabic-focused models such as Jais and AceGPT outperform general multilingual models on dialect tasks, but that significant challenges remain in dialect identification, generation and translation. What it means for you: don't judge an agent by a demo in MSA. Ask to see it handle messages written the way your customers write, in your country's dialect.

2. English mixed into Arabic

Many customers in the Gulf switch languages in the middle of a sentence. “ممكن أسوي booking للـ cleaning يوم Sunday؟” (an illustrative example) is a normal message. Linguists call this code-switching. A COLING 2025 survey of code-switched Arabic NLP describes the wide range of challenges it creates and notes that Arabic's own complexities make them harder: diglossia (the gap between MSA and the dialects), romanisation, and dialects that have no standard spelling. An agent has to read the mixed message, understand that “cleaning” is a service on your price list, keep the day right, and reply in a language the customer is comfortable with.

3. Arabizi: Arabic in Latin letters

Some customers type Arabic with Latin letters and numbers, a style known as Arabizi. “salam, abi a3rf el as3ar” (illustrative) means “hello, I want to know the prices”, with the number 3 standing for the letter ع. There is no single spelling, the same word can be written several ways, and software that decides the language from the alphabet can mistake the message for English. The survey above lists romanisation among the factors that make Arabic code-switching harder. Put Arabizi messages in your test set, and decide how the agent replies: in Arabic script, or in English if the customer prefers it.

4. Voice notes

Many customers would rather speak than type, so voice notes arrive with background noise, dialect and English words in them. The agent first needs a speech-to-text step that copes with all three. Researchers at MBZUAI built Mixat, a 15-hour dataset of Emirati speakers switching between their dialect and English, because existing speech-recognition resources served Emirati speech poorly. Testing pre-trained Arabic and multilingual systems on it showed their shortcomings on dialectal Arabic and the added difficulty of recognising code-switched speech. Test transcription on real voice notes from your own customers, and when a transcript is unclear, have the agent ask one short question instead of guessing.

5. Right-to-left text in replies and templates

Arabic runs right to left, while phone numbers, prices, times, booking codes, links and English product names run left to right. Most messaging apps set a line's direction from its first letter, so an Arabic message that opens with a name in Latin letters or an English word can display with its parts in the wrong order. Start Arabic templates with an Arabic word (مرحبًا {{1}}, not {{1}} مرحبًا), keep prices and times in short, predictable phrases, and check every template on an iPhone and an Android phone before you submit it. Write each Arabic and Turkish template natively; a template translated word for word from English reads like one.

What does the WhatsApp Business Platform require?

Language is half the job; the channel's rules are the other half. An automated agent needs the official WhatsApp Business Platform, usually through Meta's Cloud API. The WhatsApp Business app is built for people replying by hand, and unofficial tools that drive WhatsApp Web operate outside WhatsApp's terms and put the number at risk. Our WhatsApp booking playbook compares the three options in detail. Four rules then shape every conversation:

• The 24-hour customer service window. When a customer messages you, a 24-hour window opens, and it resets each time they write again. Inside it, the agent can reply freely (Meta: Service messages).
• Templates outside the window. Once the window closes, the business can only message the customer with an approved template in one of three categories: marketing, utility or authentication. A utility template follows up on something the customer did or asked for, and a template that mixes an update with a promotion counts as marketing (Meta: Template categorization).
• Opt-in. Meta requires a person's opt-in before a business messages them. The request must make clear that they are agreeing to messages from your business, name the business and comply with local law. Meta also recommends separate opt-ins for different kinds of messages and clear instructions for opting out (Meta: Get opt-in).
• The 2026 AI rule. Meta changed WhatsApp's business terms in October 2025 to bar general-purpose AI assistants, the chatbots whose AI is the product, from 15 January 2026. Meta said the change does not affect businesses that use AI to serve their customers (TechCrunch).

The ban on general-purpose assistants has since been challenged by competition authorities. In June 2026 the European Commission ordered Meta, as an interim measure, to restore free access to WhatsApp for rival general-purpose AI assistants until its antitrust investigation ends. Türkiye's Competition Board opened its own investigation and ordered Meta, also as an interim measure, to let third-party general-purpose AI chatbots offer their services on WhatsApp (NTV). Neither case is about an agent that answers, books and follows up for one business; that use was allowed throughout.

In practice: reply quickly while the window is open, write utility templates as plain facts about one booking or order, ask for opt-in in Arabic as clearly as you would in English, and keep offers in separate marketing templates sent only to people who agreed to receive them.

How do you test an Arabic AI agent before launch?

Test the agent on what your customers actually send, before any customer sees it. A sound process has six steps:

1. Collect real messages. Export a few hundred recent WhatsApp conversations, with permission and with names and numbers removed. Keep the awkward ones: voice notes, Arabizi, half-English messages, typos, complaints and questions nobody could answer.
2. Write the right answer for each one. Someone who knows the business writes the correct reply or action for every message, including the cases where the right answer is “hand this to a person”.
3. Cover every dialect you serve. If your customers are Saudi, Emirati, Kuwaiti and Turkish, each group needs its own messages in the test set, reviewed by a native speaker of that dialect. One reviewer cannot judge every dialect.
4. Add the hard cases on purpose. Questions outside your knowledge base, requests for discounts the agent may not give, angry messages, medical or legal questions, and messages that try to change the agent's instructions.
5. Score more than understanding. Check that prices, dates, times and addresses are exact in every language, that the tone fits your brand, that templates display correctly right to left, and that every hand-off trigger fires when it should.
6. Re-run the set after every change to the model, the prompt or the knowledge base. Then launch with a person reviewing replies, and widen what the agent does alone one category at a time.

How do you keep an Arabic agent accurate and safe?

Two rules carry most of the weight. The first: the agent answers only from approved knowledge, a written source of prices, services, hours, policies and procedures that a named person keeps up to date. When a question falls outside it, the agent says so and hands over. Dialect makes this rule more important. If the agent misreads a word, an agent tied to approved knowledge has nothing to invent from and can ask again, while an agent left to its general knowledge can produce a confident, wrong answer.

Keep one source of truth for every language. The Arabic, English and Turkish replies should give the same price, the same address and the same policy. Write the facts once and let the agent phrase them in the customer's language, instead of maintaining three knowledge bases that drift apart.

The second rule is a written hand-off. Decide before launch which conversations go to a person, and make sure that person receives the whole thread, including transcripts of voice notes, so the customer never has to repeat themselves. Typical triggers:

• The customer asks for a person, in any language or dialect. In the Gulf that can be as short as “أبي أكلم موظف”.
• Complaints, refunds and anything involving money the agent may not commit to.
• Medical, legal, safety or other sensitive questions.
• Anything outside the approved knowledge, or a message the agent still cannot understand after one clarifying question.
• Voice notes or images the agent cannot transcribe or read with confidence.

Tell customers at the start, in the language they wrote in, that they are talking to an AI assistant. For the wider set of guardrails, from hallucination to prompt injection and permissions, see AI agents for business.

Where should the conversations and data live?

An Arabic agent on WhatsApp handles more personal data than most businesses expect: phone numbers, names, voice recordings and their transcripts, photos, and the details of what each customer asked for. Before you choose a supplier, find out where each of these is stored, who can read it and what happens to it if you leave.

• The WhatsApp account and number. Register them in your business's name, not the supplier's, so the number and its history stay with you.
• Conversations and transcripts. Know which database they sit in, in which country, and how you export them.
• The knowledge base and prompts. They hold months of work on your dialect, your services and your tone. Make sure you own them and can take them with you.
• The model provider. Most agents send text to a language model through an API. Know which provider it is, on which terms, and whether the model can be changed without a rebuild.

Data-protection law can still apply when data is hosted in another country, and health or financial data usually carries extra rules. Our guide to owning vs renting an AI system covers what the law says in the UAE, Saudi Arabia, Türkiye and the EU, how lock-in builds up and what to put in the contract.

Flowramo builds these systems as projects and installs them on hosting you choose and own, connected to your own accounts. See what we build or tell us about the job you have in mind.

An example: Lina, an Arabic receptionist on WhatsApp

Lina is the AI receptionist we built for dental clinics, and each section above shaped how she works. She runs on the official WhatsApp Business Platform and answers patients in their own language: Arabic including Gulf dialect, English, Turkish and the many other languages AI understands. She can also be connected to the clinic's website. She transcribes voice notes before she replies, answers only from knowledge the clinic has approved, and does not diagnose. Urgent, clinical and sensitive cases go to the clinic team with the full conversation. She is installed on hosting the clinic chooses and owns, connected to the clinic's own accounts, so patient data stays on the clinic's infrastructure.

We will publish results when clinics agree to share them. Until then, you can see Lina in detail and try her on WhatsApp from that page.