Arabic-first AI is more than translation.
Arabic-first product design is architecture, not translation. The decisions that determine whether an Arabic product feels native are made long before any copy is written.
Why the translated approach keeps failing
The standard route is to build in English, add a language toggle, mirror the layout and run the strings through a translator. It produces something that is technically bilingual and obviously secondary — and users notice within seconds, usually without being able to say why.
What they're reacting to is an accumulation of small wrongnesses. Line height tuned for Latin letterforms squashing Arabic ascenders. Numerals in the wrong system for the context. A date format nobody here uses. Icons that mean 'forward' pointing the wrong way. Search that returns nothing because the user typed the same word with a different hamza.
For AI products, each of those compounds. The model inherits the interface's assumptions, the retrieval layer inherits the indexing decisions, and the prompt inherits whatever language it was reasoned in. By the time it is obvious, it's a rebuild rather than a fix.
Where the decisions actually get made
The layers that determine whether an Arabic AI product feels native — roughly in the order they're locked in.
RTL UX
Direction as a first-class layout property, using logical rather than physical CSS, so mirroring is structural instead of a set of overrides.
Typography
Arabic typefaces chosen for the interface, with line height, weight and letter spacing set independently from the Latin scale.
Arabic content
Copy written in Arabic rather than translated into it. Tone, formality and sentence length behave differently and don't survive a direct pass.
Arabic prompting
Prompts authored in Arabic when the output is Arabic. Prompting in English and asking for Arabic output produces a recognisably translated register.
Dialect handling
Real users write in Gulf dialect, not Modern Standard Arabic. Systems tuned only on MSA degrade exactly where it matters — free-text input.
Arabic search
Normalisation for hamza and alef variants, taa marbuta, diacritics and elongation. Without it, plainly relevant documents simply don't return.
Agent behaviour
Politeness conventions, greetings and escalation phrasing differ. An agent that is merely accurate can still read as rude.
Bilingual data
Document sets that mix languages, sometimes mid-paragraph. Chunking and indexing have to handle that or half the corpus is invisible.
Language switching
Preserving state, context and conversation history across a switch — including inside an ongoing AI conversation.
Localisation
Numerals, dates, currency, name order, address structure and the plural rules Arabic needs that most i18n setups don't implement.
GCC context
Channel expectations, formality norms and the specific mix of Arabic and English a given audience actually uses day to day.
Trust
Getting the small things right is what earns an Arabic-speaking user's confidence. It compounds, and so does getting them wrong.
Checks I run on any bilingual AI build
A short list that catches most of what ships broken:
Type the same query with and without diacritics, and with each hamza variant. All three should return the same results.
Send a dialect message rather than an MSA one. This is how the product will actually be used.
Switch language mid-conversation and confirm the context survives.
Read the Arabic output aloud. Translated Arabic sounds wrong immediately in a way it doesn't look wrong on screen.
Check every mixed-direction string — a name in Latin script inside an Arabic sentence is where RTL layouts break first.
Confirm numerals, dates and currency match what the audience uses, not what the framework defaulted to.
Query a bilingual document set in each language and verify both halves are actually reachable.
None of these require specialist tooling. They require someone on the team who reads Arabic as a first language and is empowered to reject the build.
Built this way
AhlanAI
An Arabic-first directory for discovering AI tools — 500+ tools across 20+ categories, with the Arabic experience designed as the primary one rather than a mirrored translation.
faisalkarkoh.com
This site runs a full Arabic route tree with its own typography stack and RTL layout, including Arabic versions of the article archive.
Where this is delivered
Kuwait — Bilingual by default
Dubai — Multilingual customer bases
Abu Dhabi — Bilingual document sets
UAE — Across the Emirates
Related
AI Product Development — The full build process
AI Agents — Agents in Arabic and English
AI Automation — Bilingual content operations
Expertise — Background and evidence
Building for an Arabic-speaking audience?
Building for an Arabic-speaking audience?
The cheapest time to get this right is before the first screen is designed. The most expensive is after launch.