Why bilingual needs its own requirements
Current language models handle Arabic far better than earlier systems did, and a quick trial in formal Arabic usually impresses. The trouble appears with real users. They write in their own dialect, switch to English for technical terms, type Arabic words in Latin letters and expect a reply that sounds natural where they live. An assistant specified only as supports Arabic will be tested in tidy formal sentences and will meet none of this until launch.
Quality also varies by language in ways that are not obvious to a team that reads only one of them. The requirements below turn supports Arabic into things that can be designed, built and tested.
Specify what users will actually write
| Input | What it looks like | Requirement to state |
|---|---|---|
| Modern Standard Arabic | Formal written Arabic, as in documents and news | Must be understood. Usually the language of your content |
| Dialect | Gulf, Levantine, Egyptian, Maghrebi and others, as people speak | Name the dialects your users write, and test with real examples of each |
| English | Often from expatriate staff or for technical subjects | Must be handled equally well, by the same assistant |
| Mixed messages | An Arabic sentence with English product names or terms inside it | Must be understood as one message, without asking the user to choose a language |
| Arabic in Latin letters | Arabic words typed with English letters and digits | Decide whether it is supported. If so, test it. If not, the assistant should ask politely |
| Informal spelling | Dropped hamzas, varied spellings of the same word, missing diacritics | Search and matching must tolerate it |
Collect real messages before writing this section. Support inboxes, WhatsApp business chats and call notes show what your own customers write, which is better evidence than any general statement about Arabic.
Decide how the assistant replies
- 01
Which language does it reply in?
What to decideThe usual rule is to reply in the language of the user's latest message. State what happens with a mixed message, and let users override it.
- 02
Formal Arabic or dialect?
What to decideReplying in clear Modern Standard Arabic is the safer default for a business, since it is understood everywhere and avoids a dialect that feels wrong to part of the audience. Some consumer brands choose a light local tone.
- 03
How are names and terms handled?
What to decideProduct names, technical terms and people's names may stay in Latin script or be transliterated. Keep a glossary and apply it consistently.
- 04
Which numerals and formats?
What to decideWestern or Eastern Arabic digits, date order, currency and units, matching your website and documents.
- 05
What tone?
What to decideGreetings and courtesy carry more weight in Arabic business writing. A direct translation of a brisk English style reads as curt.
Plan for content that exists in only one language
Most organizations hold some documents in Arabic, some in English and few in both. A user may ask in Arabic about a policy written in English. The assistant has to find the right passage across languages and then answer in the user's language, translating as it goes.
- 01
List content by language
For each source, note whether it exists in Arabic, English or both, and which version is authoritative where both exist.
- 02
Require cross-language search
The system must find an English passage from an Arabic question and the reverse. This depends on technical choices and must be tested, not assumed.
- 03
Decide how translated answers are marked
When the answer is translated from a source in the other language, say so and link the original, especially for legal or contractual text.
- 04
Fix the terminology
Supply the approved Arabic terms for your products, departments and processes, so the assistant does not invent its own translations.
- 05
Translate what matters most
Where a document is used constantly in the other language, a proper translation reviewed by a person is better than translating on every request.
For regulated or contractual content, be explicit about which language version governs. An assistant's translation is a convenience, and should never be presented as the official text.
Interface requirements for two directions
- Direction per message. Each message is laid out by its own language: Arabic from the right, English from the left, in the same conversation.
- Mixed text inside a message. English terms, numbers, links and codes inside Arabic sentences must display in the correct order. This is a known source of scrambled text and needs explicit testing.
- The input box. It should switch direction as the user types, and place the cursor sensibly in mixed text.
- Lists, tables and formatting. Bullets, numbering and table columns must follow the direction of the content.
- Buttons and labels. Every fixed element, including suggested replies, feedback buttons and error messages, exists in both languages.
- Typeface. An Arabic typeface chosen for screen reading, at a comfortable size and line height, matched with the Latin one.
- Voice, if used. Speech recognition and synthesized speech vary considerably by dialect. Treat voice as a separate requirement with its own tests.
Our guide to planning an Arabic and English website covers right-to-left design in more depth, and the same principles apply to an assistant's screens.
A worked example
A customer assistant for a telecom reseller
This example is invented to show requirements drawn from real messages. A telecom reseller in the Gulf wants an assistant on its website and messaging channel.
- EvidenceA sample of 500 past chat messages is sorted: roughly half in Gulf dialect, a quarter in English, the rest formal Arabic or mixed. About 1 in 12 Arabic messages is typed in Latin letters.
- InputRequirements name Gulf dialect, formal Arabic, English and mixed messages as supported. Arabic in Latin letters is supported for understanding, with replies in Arabic script.
- RepliesReply in the language of the latest message, in clear formal Arabic with a warm greeting. Plan names stay in English. Western digits throughout.
- ContentPlan details exist in both languages. Technical troubleshooting exists only in English. Answers drawn from it are given in Arabic with a line saying the source is in English, and a link.
- TestingTwo test sets of 80 questions each are built from the real messages, one per language, and scored by native speakers. The first run scores well in English and formal Arabic and noticeably lower on dialect questions about billing.
The dialect failures trace to billing terms that customers use and the documents do not. Adding those terms to the glossary and the content closes most of the gap, which a test in formal Arabic alone would never have revealed.
Test each language separately, with native speakers
Sound bilingual testing
- A separate test set for each language, from real messages
- Dialect and mixed-language questions included on purpose
- Native speakers score fluency and correctness
- Cross-language cases tested: question in one, source in the other
- Results reported per language, never averaged together
Testing that misleads
- English questions machine-translated into Arabic
- Only formal Arabic tested
- Scored by someone who cannot read the language
- One combined accuracy figure
- Interface checked with short phrases only
Check safety behaviour in both languages as well. Refusals, handoff to a person and handling of sensitive topics should work as reliably in Arabic and in dialect as in English. Our guide to evaluating AI answers before production describes the general method.
Bilingual assistant requirements checklist
Arabic and English assistant checklist
0 of 11 checkedWhat to do next
Pull two hundred real customer messages and sort them by language and style. The proportions will tell you which requirements matter for your audience, and the messages themselves become your first test set.
Our enterprise AI assistant development page explains how we build assistants, and you can tell us about your users and channels. A native Arabic review of any assistant before launch is a step we would always recommend.
Common questions
Do AI models understand Arabic dialects?
Modern models understand the major dialects to a useful degree, with quality that varies by dialect, by topic and by model. Formal Arabic is generally handled best. Because the variation is real and changes with each model release, the only dependable answer for your case comes from testing with messages written by your own users.
Should the assistant reply in dialect or formal Arabic?
For most businesses, clear and friendly Modern Standard Arabic is the safer choice. It is understood across countries and avoids sounding wrong to part of the audience. A consumer brand serving one market may prefer a light local flavour. Whichever you choose, write it into the requirements and have native speakers judge the result.
What if our documents are only in English?
The assistant can still answer Arabic questions, by finding the English passage and replying in Arabic. Require and test that cross-language search works, provide a glossary so that terms are translated consistently, and mark such answers as translated with a link to the source. For heavily used documents, a reviewed Arabic version is worth producing.
Is one bilingual assistant better than two separate ones?
One assistant is usually better. Users mix languages within a conversation, content often exists in only one language, and maintaining two systems doubles the upkeep. A single assistant with clear language rules, a shared glossary and separate tests per language gives a consistent experience and one place to make improvements.
How do we check quality if our team does not read Arabic?
You need native speakers in the evaluation, either from your own staff in the region or engaged for the purpose. They should write or select the test questions, confirm reference answers and score fluency as well as correctness. Automated scoring can help between reviews, but it cannot replace a reader who knows how your customers speak.
