An Arabic chatbot is rarely an English chatbot with the interface translated, though that is how most of them are built. The language is the visible part of the problem. The parts that actually break are underneath it.
This is what to expect if you are deploying an Arabic AI chatbot for customer service, and what to test before anyone outside the company talks to it.
Your customers will not write in Modern Standard Arabic
Formal Arabic is what the documentation is written in. It is not what arrives in the message box.
Customers write in dialect — Gulf, Egyptian, Levantine, Maghrebi — and those differ enough in everyday vocabulary that a model tuned on formal text will miss intent on ordinary sentences. They also write in Arabizi, Arabic transliterated into Latin characters with digits standing in for letters that have no Latin equivalent. A system that only recognises Arabic script will treat that as noise.
Then there is code-switching: an Arabic sentence with the product name, the error message and half the technical nouns in English, because that is how people in the region actually talk about software.
The test that matters is not a translated version of your English test set. It is real messages from your own support history, in whatever form they arrived.
Right-to-left is a layout problem that leaks
RTL support is usually declared done when the chat window mirrors. The failures show up at the boundaries, where a right-to-left sentence contains left-to-right content.
An Arabic sentence containing an order number, a Latin product name, a URL or a phone number will render in an order that looks wrong unless bidirectional text is handled properly. Punctuation migrates to the wrong end of the line. Truncation with an ellipsis cuts the wrong side. A number range can display reversed, which in a message about money or dates is not a cosmetic issue.
It is also worth deciding early whether you display Arabic-Indic or Western digits, and then being consistent, because mixing them inside one conversation reads as carelessness.
Names, dates and addresses do not fit the schema
A chatbot that collects information is constrained by the fields behind it, and those fields are usually inherited from an English-language system.
Arabic names frequently run to several components and do not decompose cleanly into first and last. The same person’s name has multiple legitimate Latin transliterations, so matching an existing customer record by name alone will produce both duplicates and false merges. Validation rules that assume Latin characters will reject perfectly valid input.
Dates are the sharper edge. If any part of the business runs on the Hijri calendar — and for government-facing processes in Saudi Arabia it often does — then a date a customer gives and a date your database stores may not be the same date. Decide which calendar is canonical, store that, and convert at the edges rather than hoping the ambiguity never surfaces.
Escalation needs a human who reads the same language
This is the failure that damages the brand rather than the metrics.
A bot that handles Arabic competently and then escalates into a queue staffed only in English has made the customer’s situation worse: they explained the problem once, in their own language, and now have to do it again in someone else’s. If your coverage is genuinely English-only outside certain hours, the honest design is to say so at the point of handover and offer a callback, rather than dropping the customer into a queue that cannot help them.
Context has to survive the handover too — the transcript, in the original language, in front of the agent before they type anything.
WhatsApp changes the shape of the conversation
In much of the Gulf, a customer-service chatbot in practice means WhatsApp rather than a widget on your website, and the channel imposes its own structure.
Conversations are persistent, so a customer may reply to something from three weeks ago with no context restated, and your system needs to find the thread. Business-initiated messages outside the open customer-service window are governed by templates that must be approved in advance — which means the message you want to send has to be drafted and submitted before you need it, in each language you intend to send it in. Opt-in and opt-out have to be real and respected.
Plan the template library as part of the build, not as an afterthought during launch week.
What to measure before you trust it
Run the evaluation separately by language. An aggregate quality score across Arabic and English conversations will hide a weak Arabic experience behind a strong English one, especially if English volume is higher.
Specifically: intent accuracy on dialect and Arabizi inputs, not just formal Arabic; escalation rate by language; and a human-read sample of Arabic conversations scored for whether the answer was correct and whether the register was appropriate. Automation fails quietly, and a fluent wrong answer looks exactly like a fluent right one in the logs.
Our Call Center Agent is built to hand over with context rather than optimise for a containment number. See AI chatbots and virtual agents, or read how to measure chatbot containment rate honestly and what “Arabic-first” actually means.
Common questions
Can we just translate our English chatbot into Arabic?
Translating the interface and the canned responses is the smallest part. Customers write in dialect, in Arabizi and in mixed Arabic-English, and a system trained on formal Arabic will miss intent on ordinary messages. The data model behind the bot — names, dates, validation rules — usually needs changing too.
Which Arabic dialect should the chatbot support?
Start from your own support history rather than a general answer. The dialects your customers actually write in are visible in past tickets and messages, and that set is usually narrower than “Arabic” and broader than the one dialect a vendor demonstrates.
How should an Arabic chatbot escalate to a human?
To someone who reads Arabic, with the transcript in the original language already in front of them. Escalating into an English-only queue means the customer explains the problem twice. If Arabic cover is limited to certain hours, say so at the handover and offer a callback instead.
