Blog Markets

AI development in Saudi Arabia: PDPL, data residency (2026).

What a Saudi company needs to settle before building an AI system: PDPL duties, where data may be processed, Arabic and WhatsApp, and a checklist for choosing a development partner.

Fig. 0Controls around data that has to stay put

For a Saudi company, the safest first AI system is a narrow one: a WhatsApp or web assistant, or document AI, built so it keeps personal data to a minimum, runs where your sector allows, and can swap its model later. The grace period for the Personal Data Protection Law (PDPL) ended on 14 September 2024, and fines can reach SAR 5 million.

This guide sets out, as of October 2026, what the rules ask of an AI system, how they shape hosting and model choice, what Arabic and WhatsApp change in practice, and how to choose a development partner.

What does the PDPL ask of an AI system?

  • Timeline. The law came into force on 14 September 2023, with a one-year grace period that ended on 14 September 2024.
  • Regulator. The Saudi Data and AI Authority (SDAIA) is the competent authority and publishes the law, the implementing regulations and the guidance on its laws and regulations page. Sector regulators such as the Saudi Central Bank (SAMA) add rules for the companies they supervise.
  • Penalties. Under the text of the law, a warning or a fine of up to SAR 5 million applies to violations of the law or its regulations, and the fine can double for a repeat violation. Disclosing or publishing sensitive data in breach of the law, with intent to harm the person or gain a personal benefit, carries up to two years in prison, a fine of up to SAR 3 million, or both.

The implementing regulation turns those principles into duties that land directly on how an AI system is built:

  • Consent and automated decisions. When you rely on consent, it must be specific to each purpose and documented. It must be explicit when the data is sensitive or credit data, or when a decision rests solely on automated processing.
  • Impact assessment. A written assessment is required before processing sensitive data, linking datasets from different sources, using new technologies or making decisions based on automated processing. AI projects often trigger at least one of these.
  • The processor contract. If a development partner or model host processes data for you, the agreement must state the purpose, data categories and duration. It must also commit the processor to tell you about a breach without undue delay, say whether foreign rules apply to it, and name its subcontractors. Sub-processors need your prior acceptance.
  • Breach notification. The controller notifies SDAIA within 72 hours of becoming aware of an incident that could harm the data or the person. Your partner’s reporting speed therefore decides whether you can meet that deadline.
  • Rights requests and deletion. Requests must be acted on within 30 days, and the controller can extend by up to another 30 when a request needs disproportionate effort or the person has sent several. Destruction must cover every copy in your systems, including backups, and you must ask anyone you disclosed the data to to destroy their copies too.

For an AI system, that last point is the one most often missed. Personal data lives in prompts, retrieved documents, vector indexes, caches, logs and evaluation sets, and a deletion request has to reach all of them. We design the data map first for this reason: if you can’t list where a customer’s data sits, you can’t delete it.

Can personal data leave the Kingdom?

Often yes, but not freely. SDAIA’s Regulation on Personal Data Transfer Outside the Kingdom (version 2.0, August 2024) works like this:

  • SDAIA publishes a list of countries and international organisations that provide an appropriate level of protection, and reviews it every four years.
  • Where a transfer needs a safeguard, the options are standard contractual clauses, binding common rules or a certificate of accreditation. SDAIA issued the standard contractual clauses in September 2024, with templates for controller to processor and other pairings.
  • A risk assessment comes first when you rely on those safeguards, and when you send sensitive data abroad on a continuous or widespread basis. It covers the purpose, the safeguards, how the data is limited to the minimum needed, and the likely effect on the people concerned.

An AI call to a model API in another country is a transfer of whatever personal data is in the prompt. That makes redaction before the call one of the simplest controls available: a customer’s name and national ID rarely need to reach the model for it to answer a question well.

Which hosting option fits which data?

Hosting is the choice that most shapes the architecture:

OptionFits whenTrade-off
International model API with transfer safeguardsNon-sensitive data, sectors without localisation rulesFastest to build; check whether the destination needs a safeguard and a transfer risk assessment
Model hosted in an in-Kingdom cloud regionData must stay local but you want a managed serviceFewer model choices; check what each region offers
Open-weight model on your own infrastructureStrictest requirements, or full control neededMost engineering and operations work

Build the system so the model behind it can change. Rules and regional availability are both moving, and a design tied to one provider makes every change a rebuild. In practice that means one internal interface to the model, and prompts and evaluation sets kept outside any provider’s tooling.

Which sector rules sit on top of the PDPL?

Government bodies, operators of critical national infrastructure and regulated financial institutions face the strictest hosting rules:

  • Banks, insurers and finance companies. SAMA’s Cyber Security Framework (section on cloud computing) expects approval from SAMA before cloud services are used, only Saudi-located cloud services unless SAMA explicitly approves otherwise, no secondary use of your data by the provider, logical segregation, audit rights, and return and irreversible deletion of data on exit. The framework applies to banks, insurers, financing companies and credit bureaus. Banks also fall under SAMA’s Rules on Outsourcing, which require SAMA approval for material outsourcing. Ask your compliance team how both apply to an AI vendor that hosts or processes your data. See our note on AI in financial services for the wider compliance picture.
  • Government bodies and critical infrastructure. The National Cybersecurity Authority’s Cloud Cybersecurity Controls (CCC-2:2024), an extension of its Essential Cybersecurity Controls, apply to government agencies and private entities that own, operate or host critical national infrastructure. The 2024 version removed its own in-Kingdom hosting subcontrols and, in its list of updates, tells entities to consult the National Data Management Office at SDAIA about localisation before acting.
  • Government use of generative AI. SDAIA’s guidelines for government say entities must not use generative AI in critical decisions affecting individuals or the Kingdom’s vital interests, and that results should be human-reviewed.

If one of these applies, get the regulator’s position in writing before you choose a model provider or a hosting region. An approval can add weeks that engineering alone can’t shorten.

What do Saudi companies build first?

SectorCommon first systemCheck first
Banking, insurance, finance companiesCustomer-service assistant, document intake and reviewSAMA approval for cloud use and where the data sits
Retail and e-commerceWhatsApp order, delivery and support agentWhatsApp opt-in and template rules, plus retention of chat logs
Healthcare and clinicsAppointment and follow-up messaging, summaries for cliniciansHealth data is sensitive under the PDPL, which means explicit consent (where consent is the basis) and an impact assessment
Contractors and suppliers to governmentTender and document search, report draftingWhat the tender asks on residency, Arabic capability and governance

Hotels and travel firms start in a similar way, as AI in travel and hospitality describes.

Do you need an Arabic-first model?

Not always, but you need to measure. Saudi Arabia has its own Arabic large language model, ALLaM, built on models from SDAIA’s National Center for AI and now the flagship model of HUMAIN, the AI company the Public Investment Fund launched in May 2025 across data centres, cloud and models. HUMAIN’s consumer app, HUMAIN Chat, runs on ALLaM 34B, hosted on HUMAIN infrastructure in the Kingdom. Confirm what is available for business use and on what terms. International models also handle Arabic, and the right choice depends on your hosting requirement and on scores from your own evaluation set.

That set should hold real messages and cover the cases that break Arabic systems:

  • Saudi dialects and Modern Standard Arabic in the same inbox.
  • Mixed Arabic and English in one message, and Arabic typed in Latin letters.
  • Different forms of the same letter, Arabic-Indic and Western digits, and Hijri and Gregorian dates in documents.
  • Right-to-left text inside templates and generated documents.

Score every model change against it. Our note on evals before features explains how.

Why WhatsApp is usually the channel

For customer-facing agents, WhatsApp is where the customers are. The Communications, Space and Technology Commission’s Saudi Internet 2025 report puts WhatsApp usage at 92.7%, the highest of any social app in the Kingdom.

The platform has rules your build must respect. Under the WhatsApp Business Messaging Policy, you may contact people only with their opt-in. You may reply without a template within 24 hours of their last message and use approved templates after that. You may use automation, but it must come with a prompt, clear route to a person. The Business Solution Terms also bar providers of large language models and other AI technologies from using the platform to offer those technologies when they are the primary thing on offer, so check that your use case is a business serving its own customers. You may still retain an AI provider as your solution provider, but the terms stop you using WhatsApp Business Platform data to train or improve AI models, other than fine-tuning a model for your exclusive use.

An agent here should answer in Arabic or English to match the customer, stay within limits set in code, and hand over with the full conversation. Our write-up of a production WhatsApp sales agent shows how this works. A WhatsApp sales agent we built handles 70% of a solar company’s monthly sales, with 99% uptime and a 4.8/5 customer rating. For how these agents work end to end, see our guide to WhatsApp AI sales agents.

Where should a person stay in charge?

Some decisions should not be left to a model, whatever the hosting:

  • Decisions about individuals. Credit, eligibility, hiring and complaints outcomes. The PDPL requires explicit consent for decisions based solely on automated processing, and the government guidelines above bar generative AI from critical decisions. Use AI to prepare the case and let a person decide.
  • Commitments. Prices, discounts, delivery promises and refunds belong in rules enforced in code, with a person approving anything outside them.
  • Regulated communications. Anything that reads as advice or a regulatory filing needs a qualified person to review it before it goes out.
  • Unclear answers. When the system isn’t sure, it should say so and route the case, not guess. Our note on designing AI that fails gracefully covers how.

How do you choose an AI development partner in Saudi Arabia?

Use these questions when you compare partners, and ask for the answers in writing. The first four map directly to the duties above.

  1. Where will our data be processed? Name the provider and region for each component, say whether any of it leaves the Kingdom, and say which transfer safeguard covers it.
  2. Will you sign a processor agreement that matches the implementing regulation? It should cover purpose, data categories, duration, breach notification without undue delay, foreign-law exposure and named subcontractors.
  3. How quickly will you tell us about an incident? You have 72 hours to notify SDAIA once you are aware.
  4. Can you show a data map? It should list every place personal data sits (prompts, indexes, logs, caches, evaluation sets) and how a deletion request reaches each one.
  5. What does the system log, and for how long? Logs contain personal data too, so retention should be set deliberately and not left as a default.
  6. How will you test Arabic? With a test set drawn from real messages, in dialect and Modern Standard Arabic, scored on every change, with results shared with you.
  7. Can the model be swapped for a locally hosted one if our requirements change, without a rebuild?
  8. What can the system do without a person? Ask to see the limits on what it can commit to, and the route to a human.
  9. What happens after launch? Look for monitoring, regression tests and a named owner, as described in our buyer’s guide to choosing a builder.
  10. How do we leave? Expect data return and deletion on exit, plus a handover of prompts, evaluation sets and documentation. SAMA’s framework expects termination rights and deletion on exit for the firms it supervises.
  11. Have you worked with your sector’s approvals before? Ask what the partner needs from you to support a SAMA, NCA or tender review.

A partner who answers these concretely has probably done it before. One who answers with slogans such as “fully compliant” has not shown you anything yet.

Vision 2030, the national AI strategy and government work

The National Strategy for Data and AI, approved in July 2020, aims to place the Kingdom among the top 15 countries in AI, attract around SAR 75 billion of data and AI investment, train more than 20,000 specialists and create over 300 startups. SDAIA has also published AI Ethics Principles, with defined roles for system developers, owners and assessors, plus an AI Adoption Framework and guidance on deepfakes.

Government tenders run through the Etimad platform. If you sell into government, read each tender for requirements on data residency, Arabic capability, explainability and governance. The engineering habits that answer them (evaluation sets, logging, human oversight and documentation) are worth building in anyway. Companies operating across the Gulf will find the UAE’s different rules in our guide to AI development in the UAE.

How to start, and what to measure

  1. Pick one flow with enough volume to measure, such as customer questions on WhatsApp or intake of a single document type.
  2. Classify the data and write the data map. Decide what is sensitive, where it can be processed and whether an impact assessment is needed.
  3. Choose the hosting option and get any sector approval started early.
  4. Build the evaluation set from real Arabic and English messages before building the system.
  5. Run it alongside your team, with a person approving output, then widen what it does alone as the numbers earn it.

Track a few numbers from the first week: response time, the share of conversations resolved without a person, accuracy on the evaluation set by dialect, the share of cases handed over and why, and the time taken to answer a deletion or access request.

Questions Saudi buyers ask

Can we use ChatGPT, Claude or Gemini for a customer assistant?

Often yes, through their business APIs with transfer safeguards in place, if your sector has no localisation requirement and the data isn’t sensitive. Regulated sectors should confirm with their regulator’s rules first.

Does the PDPL cover data that’s only processed by an AI model?

Yes. Personal data in prompts, retrieved documents and logs is processed like any other personal data, so the same duties on consent, security, breach notification and deletion apply.

Does AI have to be hosted inside Saudi Arabia?

Not under the PDPL alone, which allows transfers with conditions. Sector rules can require local hosting, for example for SAMA-supervised institutions and for government and critical infrastructure work, so check your regulator first.

What decides how long a project takes?

Mostly integration with your existing systems, building the Arabic and English evaluation set, and any approval your sector needs. In-Kingdom hosting or a regulator’s review adds time that engineering can’t shorten, so start those first.

Planning an AI system in the Kingdom?

Modulus Labs AI builds AI agents, assistants and the apps around them for companies in Saudi Arabia and across the Gulf. See what we build on our services page, or tell us about the project, and we’ll tell you what we’d build first.