AI in Saudi Arabia

Where the Kingdom actually is

Artificial intelligence in Saudi Arabia is not a speculative conversation. It is national policy, with an institution behind it: the Saudi Data and Artificial Intelligence Authority, SDAIA, which coordinates the National Strategy for Data and AI under Vision 2030. Around it sits a growing apparatus — the National Center for AI, the annual Global AI Summit in Riyadh, sovereign investment vehicles building compute and model capability, and a regulatory layer in the Personal Data Protection Law that determines what you may actually do with customer data.

That matters for a reason that has nothing to do with national pride. It means the buyers you sell to — ministries, universities, hospitals, banks, large family groups — are being asked about their AI posture by their own boards. The question has moved from should we to what have we done, and it is now normal for a tender to include questions your competitors cannot answer either.

The Arabic problem, and why it is the whole game

Here is the thing most vendors will not tell you: general-purpose AI models are considerably better in English than in Arabic, and they are better in Modern Standard Arabic than in the Gulf dialect your customers actually type. A system that scores well in an English demo can degrade badly the moment a real customer writes in Saudi dialect, switches to English mid-sentence, or drops the diacritics — all of which happen constantly.

The Kingdom has taken this seriously. ALLaM, developed domestically, is an Arabic-first large language model built precisely because Arabic capability could not be left to models trained overwhelmingly on English text. But at the level of an individual business, the lesson is more practical: if nobody has evaluated your AI system in Arabic separately, you do not know whether it works.

This is the single most common thing we fix. A team buys or builds an assistant, tests it in English because that is what the demo was in, ships it, and then discovers that Arabic accuracy is materially worse — usually from a customer complaint rather than a dashboard. We evaluate Arabic and English separately, report them separately, and treat the Arabic number as the one that decides whether the thing is ready.

What "code-switching" does to accuracy

Saudi users routinely write sentences that begin in Arabic and end in English, or that use English technical terms inside Arabic grammar. Models handle this unevenly. Retrieval systems handle it worse, because the embedding for a mixed-script query often lands nowhere near the documents that would answer it. If your customers do this — and in Riyadh, Jeddah and the Eastern Province they certainly do — it needs to be in the test set, not discovered in production.

What is genuinely working right now

Stripped of the hype, four categories are delivering reliable value for Saudi organisations today.

1. Arabic-capable customer assistants

Not a scripted chatbot with buttons. An assistant that answers from your actual policies, prices and documents, in the language the customer wrote in, and hands off to a human when it is not confident. The value is not headcount reduction — it is answering at 11pm on a Thursday, which is when a meaningful share of enquiries arrive.

2. Retrieval over your own knowledge

Every organisation has knowledge trapped in PDFs, shared drives and the heads of three people. Retrieval-augmented systems let staff ask questions and get answers with citations back to the source document, which is what makes the answer checkable and therefore usable. This is the most consistently valuable internal deployment we see, and it is unglamorous enough that it rarely gets a press release.

3. Document processing

Invoices, contracts, government forms, delivery notes, medical records — extraction and classification across Arabic and English documents. Saudi organisations handle an enormous volume of bilingual paperwork, and this is the category with the clearest, fastest payback because the manual cost is already on a spreadsheet somewhere.

4. Being found by AI, not just by Google

A quieter shift, and a consequential one. People increasingly ask ChatGPT, Perplexity or Google's AI Mode for a supplier recommendation instead of running a search and comparing results. If the model does not know your business exists, you are not in the shortlist at all — and there is no page two to recover from. We cover this as generative engine optimisation, and it is currently far cheaper to win in Arabic than in English.

Why AI projects fail in Saudi Arabia

We have been asked to rescue enough of these to see the pattern. The failures are almost never technical.

• No costed problem
• No Arabic evaluation
• Data not ready
• No human fallback
• Ignored PDPL
• Nobody owns it after launch

No costed problem. The project begins with "we should do something with AI" instead of "this process costs us X hours a month". Without the second sentence there is no way to tell whether the result was worth it, so the pilot runs forever and quietly dies.

Data not ready. Retrieval systems inherit the quality of what they retrieve from. If your policy documents contradict each other, the AI will confidently surface the contradiction. Cleaning the source material is usually the majority of the work and it is always underestimated.

No human fallback. Any system making consequential decisions needs a defined path to a person. Not as a courtesy — as a control. When the model is uncertain, the correct behaviour is to say so and escalate, and that behaviour has to be designed in rather than hoped for.

PDPL treated as an afterthought. Saudi Arabia's Personal Data Protection Law governs how personal data may be processed and transferred. Deciding where your data lives, who can see it and how long it is retained is a design input, not a compliance review you schedule after launch.

How to start in ninety days without wasting a year

The sequence that works is unromantic and it is the same every time.

A realistic first ninety days

Weeks 1–2	Pick one process with a number attached to it. Write down what it costs today and what "better" would mean numerically.
Weeks 3–4	Assemble a test set — fifty to two hundred real examples, in Arabic and English, including the messy ones. This is the asset that outlives the project.
Weeks 5–8	Build a narrow prototype. Narrow is the point. One process, one language pair, one clearly defined success threshold.
Weeks 9–10	Evaluate against the test set. Arabic and English scored separately. If it misses the threshold, you have spent ten weeks, not a year.
Weeks 11–13	Integrate into the system your team already uses, with monitoring, a cost-per-query figure and a human escalation path.

Notice what is absent: choosing a model, selecting a platform, comparing vendors. Those decisions are downstream and they are far easier once the test set exists, because the test set turns a subjective argument into a measurement.

AI has also changed how customers find you

Worth separating out, because it affects every business in the Kingdom whether or not they ever deploy a model. Search itself now generates answers rather than only listing links. Google's AI Overviews summarise, ChatGPT and Perplexity recommend, and a growing share of queries are resolved without a click to anybody's website.

Two disciplines respond to this. Answer engine optimisation structures your content so it can be lifted correctly into an answer. Generative engine optimisation builds the entity signals and citable material that make a model name you when someone asks for a supplier. Both are cheaper in Arabic right now than in English, because far fewer Saudi businesses have done the work.