Arabic AI

Arabic AI Chatbots: Why Saudi Businesses Need AI That Understands Local Customers

Almost every assistant sold in the Kingdom supports Arabic. Far fewer understand the Arabic customers actually type — which is not the Arabic of official documents, and is where most deployments quietly fail.

8 min read Jun 25, 2026

There is a specific way Arabic assistants fail in Saudi Arabia, and it is not the way anyone expects. They do not produce broken Arabic. The grammar is usually fine and the tone is usually polite. They fail earlier than that: they do not recognise what was asked, so they answer a question the customer did not put.

The customer sees a fluent, confident, irrelevant reply. They try once more, get another one, and leave. Nothing in the logs looks like an error. This is why Arabic support gets signed off in a demo and disappoints in production.

Two Arabics, and only one of them is in your documents

Modern Standard Arabic is the register of policy documents, contracts, news broadcasts and almost every corpus a model is likely to have seen a lot of. It is also, essentially, nobody’s conversational register. A customer asking about their order does not compose a sentence in MSA.

What arrives instead is everyday Saudi Arabic, with regional variation across the Kingdom, frequent English technical terms dropped in mid-sentence, Arabic typed in Latin characters, inconsistent hamza and taa marbuta, missing diacritics as a matter of course, and elongated letters for emphasis. Every one of these is normal writing. Every one of them moves the text away from the register the system was tuned on.

How an Arabic question arrives, versus how it is written down

Formal MSA — how the policy is written

Everyday Saudi Arabic — how it is asked

Arabic with English terms mixed in

Arabic typed in Latin letters

Illustrative, not measured. A system trained and tested only on the first row will look excellent in evaluation and struggle with the other three in production.

Structural comparison of registers, not a measured distribution of traffic. The point is the mismatch between the two columns, which is what retrieval has to bridge.

None of this is exotic. It is simply what people write. An assistant that handles only the first row of that figure has been tested against the documentation rather than against the customers.

Where Arabic actually breaks: retrieval, not generation

The instinct is to blame the language model. Usually the model is not the problem. The problem is one step earlier, in retrieval — the stage that decides which passages of your material the model is even allowed to see.

If a customer asks a question in everyday Arabic and retrieval returns passages about something else, the model is now being asked to answer a question using material that does not contain the answer. It will either refuse, which looks unhelpful, or produce something plausible from the wrong passage, which is worse. Both failures trace back to a retrieval step that did not recognise the phrasing.

Worth remembering

When an Arabic assistant gives a fluent answer to the wrong question, look at what was retrieved before you look at what was generated. The wrong passage almost always arrived before the wrong answer did.

Fixing this is unglamorous engineering: normalising the many ways the same Arabic word is legitimately typed, indexing the phrasings customers actually use alongside the formal wording in the source, and — critically — never letting the Arabic index fall back to English material because the English happened to score better on a distance metric. That last one is a real and common failure, and it renders an entire body of Arabic documentation invisible to Arabic-speaking customers.

Testing Arabic properly

The single most useful discipline here is to stop testing with the questions your documents already contain. Testing an assistant with a phrase lifted from the source passage measures nothing, because the match is guaranteed. It is the reason so many Arabic evaluations report near-perfect scores that production immediately contradicts.

A test set that tells you something

CollectReal questions from real channels

RephraseThree ways a customer would say each

Mix registersFormal, everyday, and mixed with English

Score retrievalWas the right passage found at all?

Then score answersOnly what retrieval could support

Score retrieval separately from answer quality. Conflating them hides which half is broken.

The rephrasing step is the whole exercise. Without it you are measuring whether text matches itself.

Run that test in Arabic and English as two separate exercises with two separate scores. A combined average is the most effective way to hide a language problem: strong English results will carry a weak Arabic set to a respectable-looking number, and the Arabic-speaking half of your customers will experience the weak half.

Arabic-first, versus Arabic added afterwards

	Arabic added later	Arabic-first
Its own retrieval index	No	Yes
Everyday phrasings indexed, not only MSA	No	Yes
Written-form variants normalised	No	Yes
Tested on rephrased customer questions	No	Yes
Scored separately from English	No	Yes
Right-to-left layout by design	No	Yes
Escalates to a person in Arabic	No	Yes

The fifth row is the one that hides all the others. A single blended score lets strong English results carry a weak Arabic set to a respectable number.

What distinguishes the two in practice. None of these are model choices — every row is a build decision made before a single question is answered.

2registers a customer may write in, often in one sentence

2indexes — Arabic material stays reachable to Arabic questions

3rephrasings per test question, minimum

0Arabic scores averaged into an English number

Tone, and the things a system should not say

Register matters beyond comprehension. Saudi customers expect a certain formality in a first exchange with an institution — a greeting, an appropriate level of directness, and courtesy that does not tip into the artificial. Assistants translated from English templates tend to land somewhere between abrupt and oddly casual, and both are noticed.

The other half of tone is knowing where to stop. An assistant handling a bereavement, a medical anxiety or a serious complaint should recognise the situation and reach a person quickly, in Arabic, without a menu in between. This is a design requirement, not a nicety, and it is worth writing into a specification explicitly.

What Arabic-first actually means in practice

Arabic-first is not a marketing position. It means the Arabic path is built and measured in its own right rather than inherited from the English one: its own index, its own normalisation, its own test set, its own reported score, and a layout that reads right-to-left because it was designed to rather than because a stylesheet was flipped.

This is the standard Elbi is built to. Arabic and English are maintained as separate, separately-measured paths over the same material, so an improvement to one is never quietly reported as an improvement to both. For an organisation whose customers write in both languages, that separation is the difference between a claim and a service.

Common questions

Do AI chatbots understand Saudi dialect?

Some do, and it depends far more on how the system was built than on which model it uses. The determining factor is usually retrieval — whether the everyday phrasings customers actually type are indexed alongside the formal wording in your documents. Ask a vendor to demonstrate with your own questions in everyday Arabic rather than with their script.

Why does our Arabic chatbot answer the wrong question?

Almost always because retrieval returned the wrong passage before generation began. The question was phrased in everyday Arabic, the material is written in formal Arabic, and nothing bridged the two. Check what was retrieved before judging what was generated.

Is Modern Standard Arabic enough for customer service?

For the assistant’s replies, generally yes — MSA reads as appropriately formal. For understanding incoming messages, no. Customers write in everyday Saudi Arabic, mix in English terms and sometimes type Arabic in Latin characters, and the system has to recognise all of it.

How should we test an Arabic AI assistant?

With rephrased questions, never with sentences copied from your own documents — that measures whether text matches itself. Collect real questions, write three customer phrasings of each, mix registers, and score Arabic and English as two separate exercises with two separate numbers.

Can one assistant serve Arabic and English customers equally well?

Yes, provided each language has its own index, its own normalisation and its own reported score. Problems appear when the Arabic path is treated as a translation layer over an English system, or when a single blended metric lets strong English results mask weak Arabic ones.