Blog

AI Implementation

Arabic and English AI in the UAE: What Bilingual Deployment Actually Requires

Most AI demos in the UAE run in English, then meet a document set that is half Arabic. A practical guide to the new UAE built Arabic models, the four places bilingual systems break, and how to test a vendor's Arabic claim on your own files.

Green road sign in the UAE showing place names in both Arabic and English, representing the bilingual document set AI systems have to handle
Photo by Ainur Iman on Unsplash Source

Almost every AI demo a UAE business sits through is run in English. The prompts are English, the sample documents are English, and the answers come back fluent. Then the system meets the real document set: a tenancy contract in Arabic, a ministry submission in Arabic, a scanned municipality approval with a stamp across half the text, and a supplier message that switches language mid sentence. Accuracy that looked like ninety percent in the demo becomes something much less comfortable in production.

This is not a translation problem, and it is not solved by choosing a model that claims to support Arabic. It is a deployment problem that touches how documents are ingested, how retrieval is indexed, how output is rendered, and how the whole thing is tested before anyone trusts it. This guide covers what changed in the Arabic model landscape over the past twelve months, the four places bilingual systems actually break, a build sequence that works, and the questions that separate a vendor who has done this from one who has read about it.

Bilingual Is the Default Here, Not an Edge Case

Arabic is the official language of the UAE, according to the government's own fact sheet. Business runs largely in English, which is why the bilingual character of the document set is easy to miss until an AI project exposes it. The split is not random either. It follows the counterparty.

  • Anything filed with, issued by, or litigated before a federal authority tends to be Arabic. Under Article 5 of Federal Decree-Law No. 42 of 2022, the Civil Procedure Law, Arabic is the language used in courts, with English permitted only where a judicial authority authorises it for specialised circuits.
  • Commercial documents between private parties are usually English, often with an Arabic version attached for registration or enforcement.
  • Operational records, site instructions, handover notes and messaging threads land in whichever language the person writing them was most comfortable with, which in practice means mixed.
  • Historical archives are the worst case: scanned, stamped, sometimes handwritten, and rarely searchable in either language.

So a UAE company of fifty to two hundred people does not have an English corpus with a bit of Arabic in it. It has two overlapping corpora with different formats, different quality, and different retrieval behaviour. Most AI proposals are scoped and priced as though only one of them exists.

The Arabic Model Landscape Changed in the Last Twelve Months

Two years ago, weak Arabic performance was a legitimate reason to delay a bilingual project. That has shifted quickly, and mostly from inside the UAE.

In December 2025, Inception (a G42 company), Cerebras and MBZUAI's Institute of Foundation Models released Jais 2, open weight Arabic models at 8 billion and 70 billion parameters, trained on what the developers describe as the largest Arabic first dataset assembled to date. MBZUAI's Professor Preslav Nakov stated the underlying problem plainly in that release: Arabic has long been underserved in AI development because of limited high quality data. Coverage of the launch by Middle East AI News notes that the models handle Modern Standard Arabic, regional dialects, and code switching between Arabic and English, which is the capability that matters most for real business text.

A month later, on 5 January 2026, Abu Dhabi's Technology Innovation Institute launched Falcon-H1 Arabic in 3B, 7B and 34B configurations. TII reported scores on the Open Arabic LLM Leaderboard of 61.87 percent for the 3B model, 71.47 percent for the 7B, and 75.36 percent for the 34B, with the largest model outperforming systems above 70 billion parameters, as reported at launch. For document heavy work the more useful figure is the 256,000 token context window, because it means a long contract or a full tender pack can be processed without being cut into fragments that lose their cross references. The earlier Falcon Arabic release, a 7 billion parameter model trained on 600 billion tokens with a 32,000 token context window, remains a reasonable baseline for lighter workloads.

Two practical conclusions follow. First, raw model capability in Arabic is no longer the binding constraint for most business use cases. Second, because several of these models are open weight, they can be hosted inside a UAE data boundary, which matters a great deal when the Arabic documents are also the sensitive ones. That intersects directly with the UAE data privacy rules governing where regulated data may be processed.

Where Bilingual Systems Actually Break

Retrieval, Not Generation

This is the failure almost nobody tests for. In a retrieval augmented system, the model can only answer from what the search layer hands it. If a user asks a question in English and the answer lives in an Arabic document, a naive keyword index or a single language embedding model will not return that document at all. The model then answers confidently from the English documents it did retrieve, which is worse than failing outright, because the answer looks complete. Cross lingual retrieval has to be built on purpose: a multilingual embedding model, and ideally the query expanded into both languages before the search runs.

Scanned Arabic Documents

Arabic optical character recognition is genuinely harder than Latin script, and not by a small margin. A 2023 survey of the field by Kasem, Mahmoud and Kang, Advancements and Challenges in Arabic Optical Character Recognition, sets out why: the script is cursive so characters connect and overlap, letters take different shapes depending on their position in a word, diacritical marks are usually omitted in everyday writing, ligatures change letterforms contextually, and labelled training data lags far behind Latin script equivalents. Reported character accuracies across the literature span roughly 83 to 99.8 percent depending on method and dataset, which is an enormous range when the extracted text is about to answer a contract question. Add a stamp, a signature crossing the text, or a third generation scan, and pipeline assumptions built on clean English PDFs stop holding.

Dialect and Code Switching

Modern Standard Arabic is what formal documents use. It is not what someone writes in a site group at two in the morning about a delivery that did not arrive. Gulf dialect, Levantine, Egyptian, Arabic transliterated into Latin characters, and sentences that change language halfway through are all normal in UAE operational text. A system evaluated only on Modern Standard Arabic will look fine in testing and then underperform on exactly the channel where response speed matters most.

Rendering and Text Direction

The unglamorous one. Right to left Arabic mixed with left to right numbers, English product codes and Latin punctuation breaks layouts, reverses digit order in rendered output, and produces citations that are technically correct and visually unreadable. It is a front end problem rather than an AI problem, which is precisely why it tends to be discovered after go live.

A Build Sequence That Works

The order matters more than the tooling. Each step below exists because skipping it invalidates the ones after it.

  • Inventory the corpus by language and format before anything else. Count how many documents are Arabic, how many are scanned rather than digital, and how many are mixed. One afternoon of counting reprices most bilingual projects honestly.
  • Fix ingestion before retrieval. Run a real sample of your worst Arabic scans through the text extraction layer and measure accuracy against a hand checked version. Do not accept a vendor's general OCR benchmark as a substitute.
  • Index bilingually. Use a multilingual embedding model, store the source language as metadata, and test cross lingual queries explicitly: ask in English and expect an Arabic source, then ask in Arabic and expect an English source.
  • Decide the answer language by policy rather than by accident. The usual rule is to answer in the language the question was asked in, cite the source document in its original language, and never silently translate a contractual clause without flagging that it is a translation.
  • Build the evaluation set in both languages, including dialect and code switched examples taken from your actual channels. Fifty real bilingual questions with known correct answers is worth more than any vendor accuracy claim.
  • Keep a person in the loop on anything legal, contractual, or filed with an authority. AI generated Arabic is a draft, not a certified translation, and the courts are explicit about which language governs.

How to Test an Arabic Claim in a Vendor Demo

The vendor evaluation framework that applies to any AI purchase still applies here, with four bilingual additions. Ask these in the room, before the contract.

  • Run this on my document, not yours. Bring one difficult Arabic scan, one bilingual contract, and one code switched message thread. A vendor who has solved this will not flinch at the request.
  • Ask it in English and make it answer from the Arabic file. This is the cross lingual retrieval test, and it is the single question that separates a genuine bilingual system from an English system with an Arabic label.
  • Which engine handles Arabic text extraction, and what accuracy did you measure on documents like mine? A named engine and a measured number are the right answer. A claim that the platform supports Arabic is not.
  • Which model, and where does it run? If the answer involves an open weight Arabic model hosted inside the UAE, that is a materially different compliance position from a foreign hosted API. The sovereign and local provider landscape gives you the alternatives to compare against.

What Getting This Wrong Actually Costs

A bilingual system that quietly fails in one language is not a partial success. It is a system that produces complete looking answers while ignoring part of the evidence, and the failures cluster in the highest stakes documents, because those are the ones most likely to be Arabic and most likely to be scanned. A tender response drafted without the Arabic addendum, a compliance answer that missed the Arabic ministerial circular, a contract summary built from the English version when the Arabic version governs: those are not accuracy percentages, they are commercial exposure.

The encouraging part is that this has become an engineering problem rather than a research problem. The models exist, several of them were built here, and several can run inside the country's data boundary. What is still commonly missing is the discipline of treating Arabic as a first class part of the pipeline instead of a feature checkbox, and of testing the system on the documents that actually sit in the filing room.

Where to Start

Count your Arabic documents. That is the whole first step. Most UAE businesses find they hold considerably more Arabic, and considerably more scanned Arabic, than the English shaped proposal on their desk assumes. Once that number exists, the questions above stop being theoretical and the vendor conversation changes character, because you are no longer asking whether a system supports Arabic. You are asking it to prove it on your own files.

Research Sources Used

FAQ

Common questions.

Can AI systems handle Arabic documents reliably in 2026?

For digital Arabic text, yes, for most business use cases. UAE built models such as Falcon-H1 Arabic and Jais 2 now score well on Arabic benchmarks and handle Modern Standard Arabic, regional dialects, and Arabic to English code switching. The weaker link is scanned Arabic. Optical character recognition on Arabic remains harder than on Latin script because the script is cursive, letterforms change by position and diacritics are often omitted, so reported accuracy varies widely with document quality. Test your own worst scans rather than trusting a general benchmark.

Do we need a separate Arabic AI system alongside our English one?

Usually no. One system with a multilingual retrieval layer is simpler to operate and avoids answers splitting across two knowledge bases. What you do need is a bilingual index, so a question asked in either language can find evidence in the other, plus an explicit policy on which language answers come back in and how translated text is flagged.

Which Arabic language models are built in the UAE?

Two main families. Falcon, from Abu Dhabi's Technology Innovation Institute, which launched Falcon-H1 Arabic in 3B, 7B and 34B sizes on 5 January 2026 with a 256,000 token context window. And Jais, from Inception (a G42 company) with Cerebras and MBZUAI, which released Jais 2 in 8 billion and 70 billion parameter open weight versions in December 2025. Because both are open weight, they can be hosted inside the UAE rather than accessed through a foreign API.

Does Arabic legally govern documents our AI produces?

For anything filed with a federal authority or heading toward litigation, Arabic governs. Article 5 of Federal Decree-Law No. 42 of 2022 makes Arabic the language used in the courts, with English allowed only where a judicial authority permits it for specialised circuits. In practice that means AI generated Arabic in a legal or regulatory context should be treated as a draft for a qualified person to review, and a certified legal translation is still required where one is required today.

What is the first thing to fix in a bilingual AI deployment?

Ingestion, specifically the quality of text extracted from Arabic scans. Everything downstream depends on it, and it is the stage most often assumed rather than measured. Run a sample of your genuinely difficult documents, the stamped, faxed and handwritten ones, through the pipeline and compare the extracted text against a hand checked version before judging the AI layer at all.