Buyer Guide

Arabic AI chatbots for customer support: dialects, testing and vendors

How to choose an Arabic AI chatbot for support: Gulf, Egyptian and Levantine dialects, Arabizi, WhatsApp, a test plan and 7 vendors compared.

Author
By the Open Team
|Updated November 17, 2026|9 min read

An Arabic AI chatbot for customer support is an AI agent that reads and answers customers in Arabic: Modern Standard Arabic, the dialects people type (Gulf, Egyptian, Levantine), Arabizi written in Latin letters, and messages that switch to English halfway through. Most vendors say they support Arabic, but very few publish accuracy by dialect, so the only reliable way to choose one is to test it on your own conversations, per dialect, on the channel your customers use. In Saudi Arabia and the UAE that channel is usually WhatsApp.

TL;DR

  • "Supports Arabic" can mean anything from Modern Standard Arabic only to Gulf, Egyptian and Levantine dialects. Ask which, and ask for evidence.
  • Customers rarely write formal Arabic. Expect dialect, Arabizi ("wen talabi?"), English product names and voice notes in the same thread.
  • None of the vendors we checked publish dialect accuracy figures. Build a test set of 30 to 50 real messages per dialect and score each vendor on it.
  • WhatsApp is the main channel. Saudi Arabia's Saudi Internet Report 2025 puts WhatsApp usage at 92.7%, the most used social platform in the country.
  • Data residency is a legal question, not a feature. Saudi Arabia and the UAE allow transfers abroad under conditions. Ask vendors where data is stored and processed, and get it in writing.
  • Open answers in Arabic and English for customers such as Foodics, Qoyod and Banqy, and charges $0.75–$0.90 per resolution. Handoffs to a human aren't billed.

What is an Arabic AI chatbot?

An Arabic AI chatbot is software that understands customer messages written or spoken in Arabic and replies in Arabic, usually on top of a large language model (LLM). For support, the useful kind does more than answer questions: it looks up orders, changes bookings and hands the conversation to a person with the context attached.

Older Arabic chatbots ran on decision trees and keywords, which broke on any phrasing nobody predicted. LLM-based agents cope with free text far better, but Arabic still exposes their weak spots.

Why Arabic is harder than another language on the list

Dialects vs Modern Standard Arabic

Modern Standard Arabic (MSA, or fusha) is the Arabic of news, contracts and official websites. Customers mostly write in their own dialect instead. The differences show up in the most common support words:

VarietyWhere you'll see it"I want""How?""What?"
Modern Standard ArabicFormal writing across the regionأريدكيفماذا
Gulf (Khaleeji)Saudi Arabia, UAE, Kuwait, Qatar, Bahrain, Omanأبي / أبغىشلون / كيفوش / شنو
EgyptianEgypt, and widely understood elsewhereعايزإزايإيه
LevantineJordan, Lebanon, Syria, Palestineبديكيفشو

A model trained mostly on MSA can read "وش صار على طلبي" (Gulf for "what happened to my order") and still miss the intent, or answer correctly but in stiff formal Arabic that feels like a government letter. Research backs this up. AraDiCE, a benchmark from QCRI published at COLING 2025, found that Arabic-specific models outperform multilingual ones on dialect tasks, and that models are better at understanding dialects than at generating them.

Arabizi

Many customers, younger ones especially, type Arabic in Latin letters with numbers standing in for sounds English doesn't have: 3 for ع, 7 for ح, 2 for ء. "3ayez a3raf wen el order" is Egyptian Arabizi for "I want to know where the order is." A bot that only detects language by script will treat this as broken English.

Code-switching with English

Gulf customers in particular mix English into Arabic sentences: "أبي أسوي refund للـ order حقي". Product names, app screens and payment terms are often in English too. Some platforms detect the language once at the start of a conversation and stick with it. Intercom's own documentation says Fin "only detects the language once per conversation," which is worth knowing if your customers switch partway through.

Right-to-left text

Arabic is written right to left, but order numbers, phone numbers, emails and English product names inside it run left to right. Check that the chat widget, email templates and any buttons render mixed-direction text correctly on mobile. Ask for screenshots of real Arabic conversations, not a demo in English.

Tokenisation and model quality

LLMs split text into tokens, and most tokenizers were built around English. A NeurIPS 2023 paper by Petrov and colleagues found that the same text can need up to 15 times more tokens in some languages than in others. Arabic sits on the expensive side of that gap. More tokens means more cost per message for the vendor and less room in the model's context window, which can show up as shorter memory in long conversations. It's one reason Arabic quality varies between vendors running different models.

WhatsApp is where Gulf customers expect support

The Saudi Internet Report 2025 from the Communications, Space and Technology Commission found WhatsApp was the most widely used social media platform in Saudi Arabia, at 92.7%.

That changes what you should test. On WhatsApp, customers send voice notes in dialect, screenshots of error messages and photos of receipts, and they come back days later expecting the thread to remember them. A bot that handles typed MSA well but can't transcribe a Khaleeji voice note will hand most of those conversations to your team.

For the channel itself, see our WhatsApp AI customer support guide, how to set up a WhatsApp chatbot and what WhatsApp Business API replies cost.

How to test an Arabic AI chatbot before you buy

None of the vendors in the table below publish accuracy by dialect. A two-week test on your own data tells you more than any demo.

  1. Pull real messages. Export 30 to 50 recent customer messages per dialect you serve, plus a set of Arabizi and mixed Arabic-English messages. Remove personal data first.
  2. Label the right answer. For each message, write down the intent and what a good resolution looks like, including when a person should take over.
  3. Add the hard cases. Include voice notes in dialect, screenshots, typos, and conversations that switch language halfway.
  4. Run every vendor on the same set. Connect each one to the same knowledge base and, where you can, the same order or booking system.
  5. Score more than accuracy. Mark whether the intent was understood, whether the answer was correct, whether the reply matched the customer's dialect and tone, and whether escalation happened at the right moment.
  6. Read the Arabic yourself. Have a native speaker from each market review a sample. Fluent-sounding MSA that answers the wrong question still fails.

A simple scorecard keeps the comparison honest:

TestWhat a pass looks like
Gulf, Egyptian, Levantine messagesCorrect intent and answer, reply in matching or neutral Arabic
MSA messagesCorrect answer, natural tone
ArabiziUnderstood as Arabic, answered in Arabic or the customer's choice
Arabic-English mixNo language flip-flopping, English terms kept where customers use them
Voice notesTranscribed and answered correctly
HandoffPerson gets the full thread in the helpdesk, customer stays in the same chat

Arabic AI chatbot vendors compared

These are vendors that came up when we searched for Arabic chatbots for customer service in the GCC. Every claim is from the vendor's own page, checked on 17 November 2026. "Not stated" means we couldn't find it on the pages listed in Sources.

VendorWhat it says about ArabicWhatsAppPublished pricingLikely better fit when
MaqsamAI agent "understands different Arabic dialects"; says its LLM was "built specifically for Arabic"WhatsApp messages billed by usagePrices not published; AI agent via salesYou want a Saudi-based contact center with phone and AI in one product
LucidyaAI engine "understands 15 different Arabic dialects, from Khaliji to Maghrebi"Yes, in its OmniServe inbox$125 or $250 per seat per month; AI agent add-on $100 per 500 conversationsYou also want Arabic social listening and surveys from the same vendor
YourGPTGulf, Egyptian and Levantine dialects, MSA, Arabizi and RTLYes, WhatsApp Business APIFrom $39 per month (billed annually), credit-basedYou're a smaller team that wants a self-serve builder
SleekFlow"Native Arabic and English support"; no dialect detail foundYes, WhatsApp Business APIFrom US$49 per month; AI credits US$20 per 5,000WhatsApp marketing and sales broadcasts are as important as support
Intercom Fin"Arabic (Modern)" listed; detects language once per conversationYes$0.99 per outcome (50 a month minimum on another helpdesk)You already run Intercom and most Arabic traffic is close to MSA
Zendesk AI agentsArabic fully supported, including right-to-left text; transliteration not supportedYesAutomated resolutions $1.50 committed or $2.00 pay-as-you-goYou're on Zendesk Suite and want AI from the same vendor
OpenAnswers in Arabic and English for customers in Saudi Arabia, the UAE and Jordan; no dialect claim publishedYes, including voice notes and images$0.75–$0.90 per resolution; handoffs not billedYou want one agent across WhatsApp, chat, email and voice on your existing helpdesk

Lucidya and YourGPT make the most specific dialect claims, and if dialect breadth beyond the Gulf matters to you, they belong on your shortlist. Maqsam's strength is phone: if most of your Arabic support happens on calls and you want a local provider, start there. And if your team already lives in Intercom or Zendesk and customers mostly write clear MSA, the built-in agent may be good enough without adding a vendor. Zendesk's note that transliteration isn't supported is worth testing against your Arabizi share.

Data residency: what the law says and what to ask

This section describes the law in general terms. It isn't legal advice, and the rules change, so check with counsel.

Saudi Arabia. The Personal Data Protection Law (PDPL) became enforceable when its grace period ended on 14 September 2024. Its Transfer Regulation, issued by SDAIA, allows personal data to leave the Kingdom for permitted purposes under conditions: the transfer can't prejudice national security, the destination should offer adequate protection, and data has to be minimised. Where a transfer relies on a safeguard such as Saudi standard contractual clauses, the controller must run a risk assessment first. King & Spalding notes that despite a common perception, the regime is built around conditional transfers rather than a blanket localisation rule. Some other categories of data carry separate local requirements.

UAE. Federal Decree-Law No. 45 of 2021 entered into force on 2 January 2022. Articles 22 and 23 allow transfers to countries with adequate protection and, in listed cases, to countries without it. The law doesn't apply to health and financial data covered by their own legislation, and the DIFC and ADGM free zones run their own data protection laws.

Questions to put to any vendor, in writing:

  • Where is conversation data stored, and where is it processed, including by the LLM provider?
  • Which sub-processors touch the data, and in which countries?
  • Can data be kept in-region, and does that change the price or the models available?
  • Which transfer mechanism do you rely on for Saudi and UAE customer data?
  • How long are transcripts and voice notes kept, and can we set our own retention?
  • Is our data used to train shared models?

Open publishes its security and privacy documentation at trust.open.cx.

Where Open fits

Agent 5, Open's AI agent, runs one agent across WhatsApp, chat, email, voice and social, and connects to the helpdesk you already use. On WhatsApp it transcribes voice notes, reads screenshots and photos, and keeps the same thread when a person takes over.

Arabic-market customers on Open include:

  • Syarah, a Saudi used-car marketplace where about 95% of customers prefer Arabic. Syarah says it chose a vendor based on which agent held up in Arabic. Read the Syarah story.
  • Qoyod, an accounting platform serving 25,000+ businesses, automates 85% of support across email and chat, answering in Arabic and English. Read the Qoyod story.
  • Foodics, a Riyadh-founded restaurant platform in 35+ countries, answers point-of-sale questions in Arabic or English during service hours. Read the Foodics story.
  • Banqy, a UAE mortgage platform, runs the agent on WhatsApp and web chat in Arabic and English, in whichever language the applicant opens with. Read the Banqy story.

Open doesn't publish dialect-by-dialect accuracy figures either, so run it through the same test set as everyone else.

Open charges $0.75–$0.90 per resolution, depending on volume, and handoffs to a human aren't billed. Carrier costs on voice are passed through at cost. The rate card is on the pricing page.

Sources

Frequently Asked Questions

It depends on your channels and dialects. None of the vendors we checked publish accuracy by dialect, so shortlist three, run each on 30 to 50 real messages per dialect you serve, and compare correct resolutions. Pick one that also works on WhatsApp if your customers are in Saudi Arabia or the UAE.

Most LLM-based chatbots understand common Gulf, Egyptian and Levantine phrasing, but quality varies. The AraDiCE benchmark (COLING 2025) found models are better at understanding dialects than generating them, and that Arabic-specific models outperform multilingual ones on dialect tasks. Test replies in each dialect, not only understanding.

Some do and some don't. YourGPT says it handles Arabizi, while Zendesk's AI agent documentation says transliteration isn't supported and text must be in the native script. Include Arabizi and mixed Arabic-English messages in your test set.

Not always. Saudi Arabia's PDPL Transfer Regulation allows transfers abroad for permitted purposes under conditions such as adequacy, safeguards and a risk assessment, and some data categories have separate rules. Ask each vendor in writing where data is stored and processed, and check with counsel.