Two words, and the sale is already gone
A customer in Erbil opens your Instagram DM and types:
شكد سعرها؟
How much is it? Two words. If your automated reply comes back as كم يبلغ سعر هذا المنتج؟ يسعدنا خدمتكم, the reader can tell in about a second that nobody is home.
The reply is not wrong. It is correct Arabic. It is the Arabic of a news bulletin, a school textbook and a government form. It is not the Arabic of a person asking about a dress at ten at night.
The customer does not think “this is a robot”. The customer thinks “this is not for me”, and closes the app. That happens before your price, your stock or your delivery time ever enters the conversation.
For scale: Instagram’s potential ad reach in Iraq was 21.5 million people, 45.4 percent of the population, according to DataReportal Digital 2026 Iraq, data snapshot October 2025. In the inboxes we have worked in, those messages arrive in dialect. We have not yet seen a customer open a DM in Modern Standard Arabic.
What is the difference between Iraqi Arabic and Modern Standard Arabic?
Modern Standard Arabic, الفصحى, is a written register. Nobody speaks it at home. Nobody types it into a DM.
Iraqi Arabic is a spoken variety with its own vocabulary, its own verb forms and its own prepositions. The distance between the two is not an accent. It is a different set of words in exactly the sentences a customer uses when they are about to buy something.
Here is the same handful of shopping questions in four Arabics.
| What the customer means | Iraqi | Gulf | Egyptian | Modern Standard |
|---|---|---|---|---|
| How much is it? | شكد سعرها؟ | بكم؟ | بكام؟ | كم يبلغ سعرها؟ |
| Do you have it? | عدكم؟ | عندكم؟ | عندكو؟ | هل يتوفر لديكم؟ |
| Is there a discount? | اكو تخفيض؟ | فيه تخفيض؟ | في خصم؟ | هل هناك خصم؟ |
| When does it arrive? | شوكت يوصل؟ | متى يوصل؟ | امتى يوصل؟ | متى سيصل؟ |
| Send me a photo | دزلي صورة | رسلي صورة | ابعتلي صورة | أرسل لي صورة |
| a lot | هواي | وايد | كتير | كثير |
| now | هسه | الحين | دلوقتي | الآن |
| mine (possessive) | مالي | حقي | بتاعي | no equivalent particle |
Read the last row again. Four regions, four completely different words for something as basic as “mine”. A bot that says حقي in Erbil is not slightly off. It is from somewhere else, and readers notice.
These forms also vary between Baghdad, Basra, Mosul and Erbil, and between older and younger speakers. There is no single official Iraqi. There is only Iraqi, and not-Iraqi.
Why a wrong preposition is worse than a wrong word
Vocabulary mistakes make a bot sound foreign. A customer reads وايد and assumes the shop is Gulf-owned. Annoying, survivable.
Prepositions are a different kind of failure, because no native speaker gets them wrong. Three that break Arabic systems in Iraq:
بعد. In Iraqi, بعده ما جاوبتوني means you still have not answered me. In Modern Standard Arabic, بعد means after. A system trained mostly on standard text can read it as “after him” and produce an answer about sequence. The customer asked why they were being ignored and got a sentence about timing.
من. In Iraqi, من اجي المحل means when I come to the shop. In standard Arabic, من means from. The bot answers about location. The customer was arranging a visit.
The attached object pronoun. Iraqi collapses “book me an appointment” into one word: احجزلي موعد. Standard Arabic splits it: احجز لي موعدا. A system that always splits reads like a government form, not a reply.
A wrong word makes your business sound foreign. A wrong preposition makes your business answer a question the customer did not ask, and nobody writes back to complain about that one.
Why do most vendors ship Gulf or Egyptian Arabic?
Because that is where the training data and the benchmarks are.
AraDiCE (arXiv 2409.11404, September 2024, published at COLING 2025), one of the benchmarks Arabic models are tuned against, builds its cultural evaluation on Gulf, Egypt and the Levant. The survey “Evaluating Arabic Large Language Models” (arXiv 2510.13430, October 2025) states directly that countries such as Egypt and the UAE appear better represented than others such as Iraq, Mauritania and Yemen.
We looked for a published Iraqi-dialect performance figure for any commercial chatbot platform and did not find one. That is not proof that none exists. It is what our search returned, checked July 2026, and we would rather say so than round it up into a claim.
We also read the published dialect lists of the best-resourced regional platforms. The lists we found name Gulf, Egyptian, Levantine and Moroccan. Iraqi was on none of them. Neither was Kurdish, in any variety. That gap is the reason chatbots built in Iraqi Arabic and Kurdish Sorani are worth paying for rather than downloading.
Kurdish Sorani is a different problem, not a harder version of the same one
Sorani is not an Arabic dialect. It is an Iranian language written in a modified Perso-Arabic script. Sharing letter shapes with Arabic makes it look close on screen while behaving completely differently underneath.
| Feature | Arabic | Kurdish Sorani | Why it matters in a reply |
|---|---|---|---|
| Letters | 28 | 33 | Several have no Arabic counterpart |
| Short vowels | Diacritics, in practice almost never typed | Written as letters, not marks | Arabic-style output comes out skeletal and hard to read |
| Word order | Verb early | Verb last | Word-for-word translation from Arabic reads broken |
| Grammatical gender | Yes | No | Arabic-trained output marks gender that does not exist |
| Definite marker | Prefix ال on the front | Suffix ەکە on the end | Attached in the wrong place, every noun looks wrong |
Letters with no counterpart in the Arabic alphabet include ڕ ڵ ۆ ێ ڤ, alongside the Persian-origin پ چ ژ گ.
- ڕ is a trilled r, and it is a separate letter from ر
- ڵ is a velarized l, separate from ل
- ۆ is the vowel o
- ێ is the vowel ê
- ڤ is v
Substituting ر for ڕ is not a typo. It produces a different word, the same way swapping ص for س produces a different word in Arabic. A system that quietly normalises Sorani into the Arabic alphabet outputs text a Sorani reader has to decode rather than read, which amounts to not replying.
The vowel point is the one most vendors miss entirely. Arabic is an abjad: short vowels are diacritics, and in practice nobody types them. Sorani writes its vowels as letters on the line instead. Text generated with Arabic habits is missing half of itself.
See what this looks like as a working reply flow in our worked example of Kurdish Sorani and Iraqi Arabic handling.
The honest part: image models and Sorani text
We have not found an image generation model that renders Kurdish Sorani correctly, and Arabic letter joining has been unreliable in every model we have tested. As of July 2026 we found no published benchmark for Sorani text rendering in image models at all, which means nobody can currently prove otherwise either.
So the rule on every build is simple. Images are generated with no text in them, with an explicit instruction against text, lettering and watermarks. The Arabic and Kurdish typography is set afterwards in a design tool. It costs two to three minutes per image.
If a vendor shows you an AI-generated graphic with Sorani text inside it, look closely at the letters. They are usually shapes that are not words.
Duhok is not Sorani
Erbil and Sulaimaniyah speak Sorani. Duhok speaks Badini, which is Kurmanji: a different variety with a different written tradition.
Any vendor who tells you “Kurdish is Kurdish” is telling you they have not built for Kurdish. Badini should be quoted and tested as its own language, never bundled in as a free variant of Sorani.
Five questions to ask any vendor before you pay
| Ask this | A bad answer | What good looks like |
|---|---|---|
| Which Arabic dialects do you support? | “Arabic” | A named list with Iraqi on it |
| Reply to اكو تخفيض؟ right now, on my phone | A screenshot from last month | Their live system, your device, this minute |
| Who checked your Kurdish? | “The model handles it” | A named native Sorani speaker who reviewed the output |
| What happens with a voice note in dialect? | Silence, or a fallback message | A transcription and a correct answer |
| Do you handle Badini as well? | “Kurdish is Kurdish” | Badini quoted and tested separately |
The second row is the one that carries the most weight. Dialect quality is not something you can verify from a feature list. Hand over your phone and watch.
How we handle it
Every build is tested with 10 to 15 real customer messages in each supported language before it goes live, pulled from the client’s own inbox rather than written by us. Sentences we invent are always sentences the system already handles. Sentences the client’s customers actually sent are the real test.
Kurdish Sorani is included from the Growth tier upward, which is 599,000 IQD (USD 399) to build and 99,000 IQD (USD 66) a month. Every conversation opens by disclosing that the customer is speaking to an AI, and that line is never removed. The point was never to pass as human. The point is to be understood by somebody who types شكد سعرها؟ at ten at night.
What dialect work costs is published in full, in dinars and in dollars, alongside every other price on the rate card.
Test the dialect before you discuss anything else. Send us a message in Iraqi Arabic or in Sorani and read what comes back.