An Arabic Chatbot That Understands Gulf Customers: Dialect, Script and Mixed-Language Messages
Customers in Oman and the Gulf rarely write to a chatbot in textbook Arabic. They write in dialect and mix Arabic with English in the same sentence. They spell the same word several ways, and some write Arabic in Latin letters with numbers standing in for sounds. On WhatsApp, many send voice notes instead of typing.
A chatbot tested only on formal Arabic is likely to misunderstand real messages, and customers who get wrong answers stop using it.
What makes an Arabic chatbot work is not a special model. It is a handful of decisions:
- test with real customer messages, not invented ones;
- normalise spelling before searching;
- write the knowledge base in natural Arabic, with the words customers actually use;
- set a clear policy for which Arabic the bot replies in;
- handle numbers, dates, currency and names deliberately;
- evaluate Arabic separately from English.
The short answer
| Challenge | What it looks like | What to do |
|---|---|---|
| Dialect | «أبغى أعرف وين طلبي»، «شحالك» | Test on real messages; add dialect terms to the knowledge base |
| Mixed Arabic and English | «ابي اغير الـ delivery date» | Retrieval that handles both; reply in the customer’s main language |
| Arabizi | “salam, abi a3raf wain talabi” | Detect it; transliterate before searching, or ask once in Arabic |
| Spelling variation | أ/إ/ا، ة/ه، ى/ي، tatweel, diacritics | Normalise before searching |
| Voice notes | Spoken Gulf dialect on WhatsApp | Test speech-to-text on your customers’ speech; confirm before acting |
| Numbers and dates | ٣٥٠ vs 350; “بعد العيد” | Normalise digits; ask for exact dates before booking |
| Names | Mohammed / Muhammad / محمد | Fuzzy matching, plus confirmation on account lookups |
Start from real messages
The single most useful asset for an Arabic chatbot is a few hundred real messages from your own customers. Take them from WhatsApp, your contact form, email, and call-centre notes, with personal details removed. They show what people actually ask, in the words they actually use.
Sort them into the questions the bot should handle and mark the correct answers. That set becomes your Arabic evaluation set. Run every model, configuration and knowledge update against it before it reaches customers.
Do not judge Arabic quality by a vendor demo, or by how the model answers questions you wrote yourself in clean Arabic. The gap between those and real customer messages is where Arabic chatbots tend to fail.
Dialect: understand it, but think carefully before writing it
Modern large language models generally understand Gulf dialects much better than earlier rule-based systems did. They do not understand them all equally well, and the weak spots are regional vocabulary, product nicknames and local place names. The cure is in your data:
- Add the words customers use to the knowledge base, next to the formal terms. If customers call a service by a nickname, the knowledge base should contain that nickname.
- Include dialect phrasings in your test set, and check retrieval as well as answers. A good model cannot answer from a passage that search failed to find.
Writing in dialect is a separate decision. A bot that replies in heavy dialect risks sounding wrong to someone from a neighbouring region. It can also come across as overfamiliar for complaints or financial matters. A practical default is clear, simple Modern Standard Arabic with familiar local vocabulary — the register of a good customer service message — and a light echo of the customer’s greeting. Decide it once, write it into the instructions, and test it with native speakers from your market.
Mixed Arabic and English
Gulf customers switch languages mid-sentence, especially for technical terms, product names and delivery vocabulary: «ابي اغير الـ address», «الـ order وصل damaged».
Three things help:
- Retrieval that works across both languages, so an Arabic question can find an English passage when that is the only source, and the other way round. Better still, keep Arabic and English knowledge at parity.
- A reply-language rule: answer in the customer’s dominant language, and keep established English terms where customers use them.
- Consistent terminology in the knowledge base. If your Arabic interface says «الطلب», the bot should too, not three different words for the same thing.
Arabizi: Arabic in Latin letters
Some customers, particularly those typing on English keyboards, write Arabic in Latin script with digits for sounds Latin lacks. The digit 3 stands for ع, 7 for ح and 5 for خ. A message like “abi a3raf wain talabi” means «أبي أعرف وين طلبي».
You have two options. Convert Arabizi to Arabic script before searching, so retrieval works. Or detect it and reply once in Arabic and English, which nudges most customers into one of them. Either way, include Arabizi in your test set. A bot that answers Arabizi with “Sorry, I didn’t understand” loses exactly the customers who were most willing to try a chat channel.
Normalise spelling before you search
Arabic allows the same word to be written in several ways, and search treats each spelling as a different word unless you tell it otherwise. Standard normalisation steps, applied to both the knowledge base and the customer’s message before searching, include:
- unifying alef variants (أ، إ، آ → ا);
- treating taa marbuta and haa (ة/ه) as equivalent for matching;
- treating alef maqsura and yaa (ى/ي) as equivalent;
- removing diacritics and the tatweel stretching character (ـ);
- normalising Arabic-Indic digits to Western digits for lookups.
These steps belong to search, not to what the customer sees. Keep the original text for display and for the model. Normalise the copy you search with.
Voice notes
On WhatsApp, a voice note is often the natural way to ask. Supporting voice means speech-to-text first, and the accuracy of speech-to-text on Gulf dialect varies between providers. Test it on recordings of real customers in your market, with their consent, before promising it.
Then add one safeguard: before acting on a voice request, the bot should restate it in text. For example: “You’d like to move your appointment to Sunday at 10am — is that right?” Misheard numbers and names are the typical error, and confirmation catches them.
Numbers, dates, currency and names
Small details cause many of the practical errors in Arabic conversations:
- Digits. Customers type both ٣٥٠ and 350. Normalise for lookups, and pick one style for replies that matches your interface.
- Dates. “بعد العيد” or “أول الشهر” are not dates. Before booking anything, ask for a specific date and show it back with the day of the week. Be explicit when a date depends on the Hijri calendar.
- Currency. Customers write ريال, ر.ع. and OMR, and prices in Oman can involve baisa — 1,000 baisa to the rial. Show amounts the way your invoices do.
- Names. محمد may be stored as Mohammed, Muhammad or Mohamed in an English-script system. Use fuzzy matching for searches, and never confirm account details on a name match alone.
The interface: right-to-left and mixed direction
A chat widget that is correct in English often breaks in Arabic. Typical failures are bubbles aligned the wrong way, punctuation jumping to the wrong end of a line, and English product codes scrambling the order of an Arabic sentence. These are layout problems with known solutions, covered in building bilingual Arabic–English software. Test the widget with real mixed-direction messages, not lorem ipsum.
Evaluate Arabic on its own
Report quality per language. A bot that is right most of the time in English and much less often in Arabic will show a healthy average while failing your Arabic-speaking customers.
Keep an Arabic test set that includes dialect, mixed messages, Arabizi and voice. Run it before every change. Have native speakers from your market review a weekly sample of live Arabic conversations. The test set should grow from every mistake they find.
How the rest of the bot is put together — retrieval, tools, guardrails and handoff — is in how a custom business chatbot is actually built.
Frequently asked questions
Can AI chatbots understand Gulf Arabic dialects? Modern language models understand Gulf dialects much better than earlier systems, but quality varies by model, region and vocabulary. Test candidates on a few hundred real messages from your own customers before choosing.
Should our chatbot reply in dialect or Modern Standard Arabic? A practical default is clear, simple Modern Standard Arabic with familiar local vocabulary, which reads naturally across the Gulf without sounding wrong to anyone. Heavy dialect risks seeming overfamiliar, especially for complaints or financial matters. Decide deliberately and test with native speakers.
What is Arabizi, and does the chatbot need to handle it? Arabizi is Arabic written in Latin letters, with digits such as 3, 7 and 5 standing for ع, ح and خ. Some customers use it, especially on English keyboards. Either transliterate it before searching or reply in both Arabic and English, and include it in your tests.
Why does our Arabic chatbot miss answers that are in the knowledge base? Often because of spelling variation. Search treats أ and ا, or ة and ه, as different characters unless you normalise them. Apply normalisation to both the knowledge base and incoming messages, and check retrieval separately from answers.
Can the chatbot handle WhatsApp voice notes in Arabic? It can, through speech-to-text, but accuracy on Gulf dialect varies between providers. Test on real recordings from your market, and have the bot confirm the request in text before acting on it.
How do we measure Arabic chatbot quality? Keep a separate Arabic test set with dialect, mixed-language and Arabizi messages and agreed correct answers. Run it before each change, report results per language, and review a weekly sample of live Arabic conversations with native speakers.
Sources: W3C — Arabic and Persian Layout Requirements · Lewis et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020)
Want a chatbot your Arabic-speaking customers will actually use? ZAWAT builds bilingual AI and automation with Arabic treated as a first language, not a translation. Book a call.