TL;DR: In Abu Dhabi a customer may write "هلا، عندكم موعد باجر العصر؟", "Hi, can I book for tmrw?", or both in one line. A useful AI chatbot understands Emirati and Gulf dialect, Modern Standard Arabic, English, mixed messages and Arabizi; replies in the customer's language in a register that suits your brand; answers only from your approved information; and hands over to a person when it is unsure. Getting there is a design and testing job on your real messages, not a model choice.
How Abu Dhabi customers actually write
- Emirati and Gulf dialect for everyday questions: availability, prices, directions, "is it open now?".
- Modern Standard Arabic in formal messages, complaints and anything official.
- English, especially from residents and businesses.
- Mixed messages that switch language mid-sentence.
- Arabizi, Arabic written in Latin letters and numbers.
- Voice notes, which many customers prefer to typing.
A chatbot built and tested only in English, then "translated", fails on most of these.
Where Arabic chatbots usually go wrong
| Failure | What the customer sees | The fix |
|---|---|---|
| Stiff formal Arabic for casual questions | A reply that sounds like a government letter | Agree a register with examples and test against it |
| Wrong language switch | An Arabic question answered in English, or the reverse | Reply in the language of the last message unless the customer asks otherwise |
| Misreading dialect words | An answer to a different question | Test with real dialect messages from your inbox and add them to the evaluation set |
| Invented answers | A confident wrong price, timing or policy | Answer only from approved content and hand over when the answer is not there |
| Broken right-to-left formatting | Numbers, prices and English names jumbled inside Arabic text | Format and test RTL output on WhatsApp and the web |
| Ignoring voice notes | "Sorry, I cannot read this" | Transcribe Arabic and English voice notes before answering |
How a good bilingual assistant is built
- Knowledge first. Your services, prices, timings, locations and policies, in Arabic and English, approved by your team. See how to prepare your data for an AI knowledge base.
- A register guide. A short page of example replies in the tone you want, in both languages.
- Rules for the edges. What the assistant never answers, when it hands over, and how it apologises.
- An evaluation set from real messages. Anonymised messages from your own inbox, including dialect, mistakes and mixed language, scored by native speakers on your team.
- Official WhatsApp. The WhatsApp Business Platform with approved Arabic and English templates on a number you own.
- Monitoring. Weekly review of conversations it could not handle, feeding back into the knowledge base.
The engineering behind this is in why Arabic breaks LLMs and the broader Arabic AI chatbot guide.
Questions to ask a vendor
- Can you run your assistant on fifty of our real messages, in Arabic, English and mixed, before we sign?
- Who on your team reads Arabic and reviews the answers?
- How do you handle Arabizi and voice notes?
- What does it do when the answer is not in our information?
- Can we read and approve the replies for sensitive topics in both languages?
How Soluvide builds Arabic assistants in Abu Dhabi
Soluvide is headquartered in Abu Dhabi (trade licence CN-6428683). We build Arabic and English AI chatbots for Abu Dhabi organisations on WhatsApp and the web, tested on your own messages in dialect and formal Arabic, answering only from your approved information, with a clean handover to your team. Send us a few anonymised messages from your inbox and we will show you how the assistant would answer them.
Talk to us
Need an assistant that handles real Arabic?
We build assistants that follow Gulf dialect, Arabizi and conversations that switch language mid-sentence, and answer only from wording you approved.
Questions
Frequently asked.
Current large language models understand Emirati and wider Gulf Arabic reasonably well, including common expressions and mixed Arabic and English. Understanding is rarely the problem; the risks are replying in stiff formal Arabic, switching languages wrongly, misreading Arabizi, and inventing answers. Those are fixed by design and testing on your real messages, not by choosing a different model.
Where this applies