AI governance · SDAIA aligned · Saudi Arabia

You cannot govern AI systems nobody has listed

Most AI governance programmes begin with a policy. They should begin with an inventory, because the estate is always larger than the organisation believes — and a policy written for an imagined estate governs nothing.

Seven principles SDAIA's AI Ethics Principles, published September 2023
Four risk tiers Classification determines the depth of control required
ISO/IEC 42001 A certifiable AI management system structure that maps onto them
Indirectly binding Enforced through PDPL, sectoral regulators and procurement
01 — The thing most organisations get wrong

"The principles aren't law, so we'll deal with it later"

Technically true, and commercially wrong. SDAIA's AI Ethics Principles are not directly enforceable in the way PDPL is. Their force arrives by three other routes, and each of them bites.

Through PDPL. The moment an AI system processes personal data — which is most systems worth governing — the binding data protection regime applies, and the ethics principles describe what good handling looks like under it.

Through sectoral regulators. A financial regulator, a health authority or a communications regulator can apply the principles within its own supervisory authority, where they carry that authority's weight rather than SDAIA's.

Through procurement. This is the one that arrives first in practice. Government entities and large enterprises increasingly expect suppliers to evidence responsible AI practice. Organisations discover the requirement when a tender asks for documentation they do not have, and the deadline is the submission date.

The practical conclusion: treat the principles as binding via that indirect pathway. The alternative is building the evidence under time pressure, badly, while a bid is open.

02 — Risk classification

Four tiers, and why classification comes first

SDAIA's framework applies a risk-based model across the AI lifecycle, from design and data selection through deployment, monitoring and decommissioning. The tier a system falls into determines how much control, documentation and oversight it needs — so classification is not paperwork that follows the work, it is the decision that scopes the work.

  • Unacceptable risk Not to be deployed

    Uses that cannot be made acceptable through controls, because the harm is inherent to the application rather than to its implementation.

    Requirement: identify and stop. The governance value is catching these at proposal stage, not after build.
  • High risk Full control set

    Systems materially affecting people's rights, safety, livelihood or access to services — credit decisions, hiring, clinical support, eligibility determination, biometric applications.

    Requirement: impact assessment, bias testing, explainability, meaningful human oversight, contestability, documented accountability and monitoring throughout life.
  • Limited risk Transparency obligations

    Systems where people should know they are interacting with AI, or where outputs shape a decision without determining it — chat assistants, recommendation and triage support.

    Requirement: disclosure that AI is in use, clear escalation to a person, output monitoring and defined scope boundaries.
  • Little or no risk Baseline hygiene

    Internal productivity and back-office uses with no material effect on individuals — document search, drafting assistance, classification of internal content.

    Requirement: inventory entry, data-handling rules for what may be sent to which tool, and periodic reclassification as use expands.

Tier definitions above summarise the risk-based model in general terms. Classification for a specific system should be made against SDAIA's published text and, where relevant, your sectoral regulator's position.

03 — The seven principles

What each principle means as an actual control

Principles are easy to agree with and hard to evidence. The right-hand column is what an assessor or a procurement reviewer will ask you to produce.

  • Fairness Defined population groups relevant to the decision, agreed fairness metrics, measured outcome differences across those groups, and a record of what was done where disparity was found.
  • Privacy and security Lawful basis for the data used in training and inference, minimisation, retention limits, protection of the pipeline, and an assessment of what leaves the organisation in a prompt or an API call.
  • Humanity Meaningful human oversight on consequential decisions — a person who can actually intervene, sees what the system saw, and is not merely rubber-stamping throughput.
  • Social and environmental benefits A stated purpose beyond efficiency, an assessment of effects on affected groups, and consideration of compute and resource cost for large-scale systems.
  • Reliability and safety Evaluation against a representative test set, defined failure behaviour, monitoring for drift in production, and a rollback path that has been exercised.
  • Transparency and explainability Layered documentation: technical model cards and evaluation reports for engineers and auditors, disclosure to affected individuals that AI is in use, and a route to contest a decision.
  • Accountability and responsibility Named owners for each system, a governance committee with real authority, third-party AI covered contractually, and audit documentation maintained rather than assembled retrospectively.
04 — Governance structure

The roles the framework expects you to have

This is where many programmes are thinnest. The framework anticipates named accountability inside the organisation, not a policy owned by whoever last attended a workshop.

  • Senior leadership engagement — AI risk owned at executive level, not delegated entirely to a technical function
  • An AI ethics governance committee — with authority to stop a deployment, which is the only test of whether it is real
  • Responsible AI Officers — accountable for turning policy into operating practice within their domain
  • AI System Assessors — able to evaluate a system against the principles independently of the team that built it
  • Chief Data Officer oversight — connecting AI practice to the wider data governance regime rather than running parallel to it
  • Annual reporting and KPI review — governance measured like anything else the board oversees
  • Contractual guarantees for third-party AI — most AI in an enterprise arrives inside someone else's product, and the obligation does not transfer with it
05 — Where it gets difficult

The parts of AI governance that are genuinely hard

Writing the policy takes a fortnight. These are the pieces that take real work, and where a programme either becomes credible or becomes a binder.

Finding the AI estate

Vendor features that quietly became AI, models embedded in purchased products, notebooks in a business unit, and staff using consumer tools with work data. The inventory is always longer than the organisation expects, and it is the foundation of everything else.

Discovery · Shadow AI

Bias testing in a Saudi context

Which groups matter depends on the decision. Nationality and residency status, region, gender, and Arabic dialect handling in language systems — where performance frequently varies in ways English-language testing never reveals.

Fairness metrics · Arabic dialect

Explainability that helps a person

SHAP values satisfy an auditor and mean nothing to a customer refused a facility. The requirement is a reason a person understands, in Arabic where the affected person is Arabic-speaking, plus a route to contest it.

Reason codes · Contestability

Oversight that is not theatre

A reviewer approving four hundred cases an hour is not oversight. Meaningful human review needs realistic volumes, visibility of what the system saw, and authority to overturn without career consequence.

Review capacity · Real authority

Generative AI usage

Different risk shape: outputs vary run to run, prompts leave the organisation, and staff adopt tools faster than governance approves them. SDAIA has issued guidance for government entities; enterprises face the same questions.

Approved use · Prompt handling

Keeping it current

Systems change, models are updated by vendors without notice, and use expands past the classified purpose. Governance that is accurate at go-live and never revisited is worse than none, because it is documented misinformation.

Reclassification · Drift
06 — Delivery

Inventory, classify, control the high-risk few

Most AI estates contain a small number of genuinely high-risk systems and a long tail of low-risk ones. Governing everything to the same depth exhausts the programme before it protects anything.

  1. AI inventory

    Discovery across business units, vendor products, embedded model features and staff tool usage. Produces the list the organisation did not have, which is usually the moment the programme gets sponsorship.

    3–4 weeks
  2. Risk classification

    Every system placed in a tier against defined criteria, reviewed with legal and the business owner. Establishes where the real exposure sits and where light-touch governance suffices.

    2 weeks
  3. Governance structure

    Committee, Responsible AI Officer and assessor roles defined with actual authority, decision rights documented, and third-party AI contract language drafted with legal.

    2–3 weeks
  4. High-risk controls

    Impact assessments, bias testing, explainability, human oversight design and model documentation for the systems that warrant it. Evidence built to be inspected.

    6–8 weeks
  5. Evidence and readiness

    Assembling what an assessor, a regulator or a procurement reviewer would ask for, and testing whether it can be produced on request rather than reconstructed.

    2 weeks
  6. Operate

    Periodic re-inventory, reclassification as use expands, monitoring for drift, and annual reporting. The estate changes faster than the documentation unless someone owns it.

    Ongoing

ISO/IEC 42001 runs alongside where certification is the goal. It is not mandated in the Kingdom, but it gives the programme a certifiable structure that maps onto SDAIA expectations and travels well in international procurement.

07 — Questions

Answered plainly

What are SDAIA's AI Ethics Principles?

Seven principles published in September 2023: Fairness, Privacy and Security, Humanity, Social and Environmental Benefits, Reliability and Safety, Transparency and Explainability, and Accountability and Responsibility. They apply to entities developing and using AI systems in the Kingdom, alongside a risk-based classification model applied across the AI lifecycle.

Are the principles legally binding?

Not directly, and this is widely misunderstood. Their force is indirect: through PDPL where AI processing involves personal data, through sectoral regulators applying them within their own authority, and through procurement, since government entities increasingly expect them of suppliers. Treating them as optional is a commercial risk even where it is not a legal one.

How does SDAIA classify AI risk?

Four categories: unacceptable risk, high risk, limited risk, and little or no risk. Classification determines the depth of controls, documentation and oversight required, which is why it is the first governance activity rather than a later formality.

What governance roles are expected?

Senior leadership engagement, an AI ethics governance committee, Responsible AI Officers and AI System Assessors, with Chief Data Officer oversight. The framework also anticipates annual reporting, KPI review, audit documentation and contractual guarantees covering third-party AI.

What is ISO/IEC 42001 and is it required?

The international AI management system standard, covering risk assessment, impact assessment, lifecycle controls and continual improvement. Not mandated in Saudi Arabia, but it provides a certifiable structure that maps onto SDAIA expectations and is increasingly useful evidence in procurement.

How do you test for bias in a Saudi context?

By defining which population groups matter for the specific decision, then measuring outcome differences across them using agreed fairness metrics. Saudi deployments raise particular considerations: nationality and residency status, region, gender, and Arabic dialect handling in language systems.

Where do governance programmes usually fail?

At inventory. Organisations cannot govern systems nobody has listed, and the list is always longer than expected once vendor features, embedded models and staff use of general-purpose tools are counted. Policy written before inventory governs an imagined estate.

On scope and currency: Crux builds the operational side of AI governance — inventory, classification, controls, testing and evidence — alongside your legal counsel, who provides interpretation and regulatory correspondence. SDAIA's framework and supporting guidance are periodically updated; confirm current requirements against SDAIA's published texts.

Start here

Start by finding out how many AI systems you actually run

A three to four week inventory and classification engagement finds the AI across your estate — including what arrived inside vendor products and what staff adopted without asking — and places each system in a risk tier with the controls it warrants.