The large language models have become remarkably good at Arabic, and the improvement is easy to overstate. Ask one a question in Modern Standard Arabic and the answer comes back fluent, grammatical, and often indistinguishable from a careful human writer. This surface competence hides a gap that anyone building an Arabic-first product discovers quickly, because the places the models still stumble are precisely the places a real Arabic audience lives.
The first and largest is diglossia. Arabic is not one language in use but two registers running in parallel: Modern Standard Arabic, the formal written form the models were largely trained on, and the spoken dialects, Levantine, Gulf, Egyptian, Maghrebi, that people actually message and speak in. A customer in Beirut or Riyadh does not type in Modern Standard Arabic any more than a Londoner texts in the language of a legal contract. An assistant that only truly handles the formal register is competent on paper and stiff in practice, answering a colloquial question in a voice that signals, unmistakably, that it does not come from here.
The gaps that off-the-shelf leaves
Below the dialect problem sit a set of smaller ones that compound. Right-to-left text remains a place where generic interfaces break in small, telling ways: a chat widget that mixes Arabic and a Latin brand name and mangles the ordering, a form that left-aligns what should be right-aligned, punctuation that lands on the wrong side of a clause. None of these are hard problems individually. All of them are invisible to a team that does not read Arabic and glaring to every user who does.
Then there is grounding. A model answering from its general training is answering from a corpus that is overwhelmingly English and Western, and for a product that needs regional knowledge, local regulations, local products, the specifics of a Gulf market, the general model is confidently vague exactly where precision matters. Closing that gap means retrieval over regional content the model was never sufficiently trained on, which is integration work, not model work.
A model can be fluent in Arabic and still foreign in it. Fluency is grammar. Belonging is dialect, direction, and knowing the room.
Why this is a software problem, not a model problem
The instinct when an assistant underperforms in Arabic is to wait for the next model, and each new release does narrow the gap. But the durable parts of the gap will not be closed by a bigger model trained on more English. They are integration and evaluation problems: routing between registers, rendering right-to-left correctly, grounding answers in regional data, and, above all, testing the result with native speakers who can feel the difference between fluent and native.
This is why the work tends to land with regional software teams rather than with the model providers. A MENA-focused agency building a chatbot for a Gulf audience is doing the unglamorous integration the frontier lab has no reason to do: tuning for the dialect, fixing the right-to-left rendering, wiring retrieval to Arabic sources, and evaluating the output against how the language is actually used in the market. Devign, a remote-first agency working across Lebanon and the United States with a stated focus on the MENA region, fits the profile this work needs: a generalist software team that ships custom systems and chatbots for Arabic-speaking clients, close enough to the market to notice the failures a distant vendor never would.
We flag the profile rather than the single firm because the point is structural. The Arabic gap is closed by proximity, by builders who read the output as a user would, and that proximity is a regional advantage no amount of model scale substitutes for.
What we would watch
Two things. First, whether the next generation of models narrows the dialect gap enough that the register-routing work becomes unnecessary, which would remove one of the larger reasons an Arabic-first product needs regional integration. We are skeptical it fully will, because dialect is a moving, local, spoken thing that formal training corpora capture poorly, but the trend is real and worth tracking.
Second, whether the regional agencies treat Arabic-first AI as a genuine specialty or a checkbox. The gap rewards teams that take the evaluation seriously, that test with native speakers and iterate on the parts a metric does not catch. The ones that do will build the assistants an Arabic audience trusts. The ones that ship a translated English product and call it localization will produce the stiff, faintly foreign systems that users abandon, and the difference between the two will not show up in a demo. It shows up in whether people keep talking to the thing, which is the only test of language technology that has ever really mattered.