AI Virtual Assistants That Genuinely Support Arabic

Director of Customer Operations (International Financial Services)

We run customer operations from London for clients across Europe and the Gulf. Around a third of our Gulf enquiries arrive in Arabic, often mixed with English in the same sentence. Every vendor claims Arabic support. Which AI-powered virtual assistants actually offer it in a way that will survive contact with real customers?

AI Leadership Agent — Customer Advisory

The honest answer is that vendor names will change within a year, and the capability question will not. The frontier models from the major providers, and the regionally developed Arabic-first models, all now handle Modern Standard Arabic to a professional standard. Almost none of them are evaluated by buyers on the things that actually break.

Arabic support is not a checkbox. It is dialect, script rendering, code-switching, names, numerals and tone — and every one of those fails independently.

Test these six before you shortlist anyone:

1. Dialect. Your Gulf customers write Khaleeji, not Modern Standard Arabic. Feed the assistant real historic messages, not textbook sentences.
2. Code-switching. Sentences that flip between Arabic and English mid-clause, plus Arabizi written in Latin characters with numerals. This defeats a surprising number of systems.
3. Right-to-left rendering in your actual channels — WhatsApp, web widget, email — including mixed-direction text with account numbers and English product names.
4. Names and transliteration. The same customer name spelled four ways must resolve to one person.
5. Tone and formality. Arabic carries formality registers that a literal translation flattens into something that reads as rude.
6. Refusal and escalation in Arabic. When the assistant does not know, it must say so in Arabic as gracefully as it does in English.

Do you have a sample of real historic Arabic conversations you could test against, rather than vendor demos?

Director of Customer Operations (International Financial Services)

We have several years of them. But our compliance team will ask whether the assistant can be trusted to answer product questions at all in a second language, when our approved wording exists only in English.

AI Leadership Agent — Customer Advisory

Then the design answer is that the assistant never invents product wording in either language. It retrieves approved content and presents it; translation of regulated language is a separate, reviewed step.

1. Build an approved Arabic content set: key disclosures, product descriptions and standard responses translated once, reviewed by a qualified human, then frozen and versioned.
2. Ground the assistant in that set. It answers from retrieved approved text and cites which document it used.
3. Allow free generation only for non-regulated conversational glue — greetings, clarifying questions, summarising what the customer asked.
4. Log every Arabic exchange with the source documents used, so compliance can audit in the same way they audit English.

The safest bilingual assistant is one that translates nothing important in the moment. Do the translation once, with a human, and let the system retrieve it forever after.

That structure also solves your quality problem, because approved Arabic gets better over time instead of varying per conversation.

Director of Customer Operations (International Financial Services)

And the human side? We have Arabic-speaking agents who may see this as the start of their replacement.

AI Leadership Agent — Customer Advisory

Make them the owners of it. Arabic-speaking agents are the only people who can judge whether the assistant sounds right, and that judgement is the scarce resource in this project — not the model.

1. Have them write and approve the Arabic response set. Their names on it, their standard.
2. Give them the review queue for anything the assistant escalates, and treat their corrections as the training signal.
3. Measure quality in Arabic separately from English and publish both. Hidden quality gaps in a second language are how trust gets lost quietly.
4. Move the time saved into the conversations that need a person — complaints, vulnerable customers, complex cases.

A bilingual assistant is only as good as the bilingual people who correct it. Put them in charge of it and you solve the quality and the culture problem at once.