Home / Why Arabic AI

Google Translate doesn't speak Gulf. Neither do most AI chatbots.

Most AI tools default to Modern Standard Arabic (MSA) or worse, Egyptian or Levantine dialects. To a Saudi customer, that sounds as weird as a British butler answering a Texas BBQ joint's phone. Gulf Arabic is different — Khaleeji dialect, code-switching with English, specific cultural references, reading tone not just words. I build for that. This page explains why that matters, which LLMs actually work, and what I do differently.

The dialect problem

"Arabic" isn't one language. It's a family.

When an AI vendor says they "support Arabic," ask which one. There are at least four dialect clusters used in customer conversations, and mixing them sounds as jarring as a Scottish brogue in a Mississippi call center.

📖

Modern Standard Arabic (MSA)

Formal. Newscasters, government documents, school textbooks. Nobody chats in MSA. "How are you" = كيف حالك. Sounds like a press release.

🇸🇦

Gulf (Khaleeji) Arabic

Saudi, UAE, Qatar, Kuwait, Bahrain, Oman. "How are you" = شلونك or كيفك. This is what your customers actually type on WhatsApp.

🇪🇬

Egyptian Arabic

"How are you" = ازيك. Common in media (Egyptian films are pan-Arab) but sounds out-of-place in a Gulf customer service context. AI defaults here often.

🇱🇧

Levantine Arabic

Lebanon, Syria, Jordan, Palestine. "How are you" = كيفك (same spelling as Gulf but different prosody). Close to Gulf but not native. Gulf users clock it instantly.

A Gulf customer asking "شلون اقدر احجز موعد؟" (how do I book an appointment?) and getting back "يمكنك حجز موعدك عبر الرابط التالي" (MSA formal) will instantly feel like they're talking to a machine — or worse, a machine trained for the wrong country. The fix isn't "more Arabic data." It's the right Arabic data, at the dialect level.

The tech stack

Which Arabic LLMs actually work for Gulf.

There are more Arabic LLMs every month. Most are academic. A few are production-ready. Here's my honest assessment from deploying them in real Gulf businesses.

🌴

Jais (G42, Abu Dhabi)

UAE-built, 13B + 30B + 70B sizes. Strong on MSA and Gulf context. Best open model for Gulf deployments today. Cultural awareness is noticeably better than Western LLMs.

🦅

Falcon (TII, Abu Dhabi)

TII's flagship. Multilingual with solid Arabic. Better for general reasoning than conversational Gulf Arabic — I pair it with Jais for dialogue heads.

🧠

GPT-4 / GPT-4o (OpenAI)

Solid Arabic but MSA-leaning. Needs prompt engineering + Gulf example shots to stay in Khaleeji. Expensive at scale. Best for edge cases, worst for defaults.

🔮

Claude (Anthropic)

Excellent MSA comprehension. Improving on Gulf dialect output. Strong at following system-prompt dialect constraints. My go-to for nuanced customer conversations.

🌐

Cohere Aya

101 languages, Arabic included. Good multilingual baseline, but not Gulf-specialized. I use it when I need a cheap multilingual fallback, not as a primary.

🎯

Fine-tuned custom heads

For enterprise clients I fine-tune on client-specific Khaleeji transcripts (with consent). That's how you go from "speaks Arabic" to "speaks like your best agent on their best day."

The WhatsApp advantage

90% of Gulf customer convos are on WhatsApp.

In the US, customer service happens over email, web chat, and (grudgingly) the phone. In the Gulf, it happens on WhatsApp. Dental appointments, restaurant bookings, real estate inquiries, e-commerce orders — WhatsApp Business number. Not email. Not a form on your website. WhatsApp.

AI tools built for US markets don't deploy cleanly on WhatsApp Business API. They're web-widget first, with "Twilio integrations" bolted on. What you actually need: session-aware message handling, template messages approved by Meta, Arabic-aware interactive buttons, media (voice notes! so common in the Gulf) and the compliance layer to keep your number from getting banned.

I build WhatsApp-first. See the WhatsApp Bot product for the self-serve version, or book a call for a custom deployment.

Compliance angle

SAMA. UAE NESA. Islamic finance.

Western AI platforms rarely think about Gulf compliance. If you're in regulated industries — banking, healthcare, takaful, government — this is where most "cheap" AI deployments fall over. Here's what I actually account for:

🏦

SAMA (Saudi Central Bank)

Data residency in KSA, cybersecurity framework (CSF) alignment, customer data handling rules. Most US AI SaaS runs in Virginia — not acceptable for Saudi financial deployments.

🛡️

UAE NESA / TDRA

UAE Information Assurance standards for government and critical-sector deployments. Local hosting + audit trails required. Jais and Falcon help because they can run on Abu Dhabi infra.

☪️

Islamic finance rules

Sharia-compliant advisory needs careful prompt engineering — no recommending riba-based products for Islamic-banking clients, no cross-selling conventional products to halal-only customers.

📋

PDPL & local data laws

Saudi PDPL, UAE PDPL, Qatar Law No. 13. Customer consent, data minimization, right to erasure. Western AI defaults break these regularly — I bake the controls in from the start.

What I do differently

Four practices that matter.

🗣️

Dialect-aware training

Every bot gets Khaleeji-shot prompts, Gulf greetings, code-switch examples ("احجز لي appointment بكرا"), and fallback handling for MSA or Egyptian inputs. Tested with native speakers before launch.

🎭

Gulf-relevant prompts

Not "Hey there!" in Arabic. Real Gulf greetings — حياك الله, السلام عليكم, masha'Allah framing where appropriate. Cultural references that don't sound like a tourist.

💬

WhatsApp-first deployment

Meta Business verification done right, template approvals in Arabic + English, voice-note handling (huge in Saudi), broadcast compliance — not just "we have a webhook."

⚖️

Local compliance review

Every enterprise project gets a compliance checklist against SAMA / UAE NESA / local PDPL before go-live. No mystery about data residency or audit trails.

FAQ

Common questions about Arabic AI.

Can't I just use Google Translate to make any AI Arabic?
No — Google Translate is a translation layer, not a dialogue model. It translates text literally and defaults to MSA. The moment a customer uses slang, code-switches, or types fragments (which is how WhatsApp works), Google Translate breaks. You also get a translator-shaped conversation: stilted, formal, and obviously machine-generated. A native Arabic LLM generates dialogue directly in Arabic — that's a fundamentally different system.
Why not just use ChatGPT / GPT-4 alone? It speaks Arabic.
GPT-4 speaks good MSA and decent Arabic generally, but it leans formal and occasionally drifts into Egyptian or Levantine patterns. Without careful prompting and few-shot examples, it won't stay in Gulf dialect reliably. For a single-user assistant that's fine. For a customer-facing bot taking 1,000+ conversations a month, those drift moments cost you real customers. I layer Gulf-specific prompting, dialect enforcement, and model routing (sometimes Jais for dialogue, Claude for reasoning, GPT for edge cases).
How is your Arabic different from other agencies?
Two things. First, I'm Arabic-fluent myself — I write the prompts, I test the outputs, I can tell when a bot sounds off. Most agencies outsource Arabic to a translator who isn't a native Gulf speaker. Second, I specialize in Gulf (Khaleeji) dialect specifically. Not "Arabic in general." That focus means my deployments feel native, not generic.
Do you support specific dialects or just "Gulf"?
Default is pan-Gulf Khaleeji, which covers Saudi, UAE, Qatar, Kuwait, Bahrain, Oman. For enterprise clients I tune to specific country patterns — Saudi Najdi vs Hijazi, Emirati vs Omani — based on your customer base. Multi-branch clients get a base Khaleeji model with regional tuning per location.
What about Maghreb Arabic (Morocco, Tunisia, Algeria)?
Honest answer: not my specialty. Darija (Moroccan Arabic) and Maghrebi dialects are genuinely different languages from Gulf Arabic — closer to each other and to French-influenced vocabulary than to Khaleeji. If you're a Maghreb business, I'd recommend a Darija-focused agency. I can help if you're Maghreb-headquartered but serving Gulf customers, in which case I'd build the Gulf-facing layer.

Ready for AI that speaks your customers' language?

Self-serve subscriptions start at $49/mo. Custom enterprise builds start at $25K. Or just book a call and we'll figure out what you actually need.

See Products → Book a Strategy Call