Arabic AI agents: why most chatbots fail in Arabic, and what to do about it
Dialect, mixed language messages and right to left layout are where generic tools break.
Translation is not conversation
A tool that translates English intent into Arabic produces replies that are grammatically correct and completely wrong in tone. They read like a government letter. A customer in Dubai writing in Gulf Arabic can tell within one message that nobody built for them.
The same happens in reverse. An agent that only thinks in English will answer an Arabic question with a literal translation of an English answer, and lose the sale.
Mixed messages are the normal case
Real enquiries in the UAE mix languages constantly. A message that opens with an Arabic greeting and then asks about a two bedroom apartment in Latin script and English numerals is one message, not two.
A useful agent detects language per message, not per customer, and holds both without asking the customer to pick a language from a menu.
Where generic tools break
These are the failures we see most often when a business brings us an existing chatbot.
- Formal Modern Standard Arabic replies that sound nothing like the customer.
- Right to left text that renders broken, with punctuation landing in the wrong place.
- Names, addresses and Emirates IDs mangled or dropped.
- No Arabic in the approval screens, the audit trail or the reports your team reads.
What good looks like
Replies in the dialect the customer used. The same facts in both languages. Arabic throughout the screens your team works in, not just the customer facing ones. And a human who can read every conversation afterwards.
How we test it
We run the same script in every language and compare the outcome: the same qualification questions asked, the same slot offered, the same record created, the same handover triggered. If the Arabic path produces a different result, the agent is not finished.