SDAIA says what good AI governance looks like. This proves you did it.
Saudi Arabia has AI ethics principles, generative AI guidelines and an adoption framework — all authoritative, all non-binding, and none of them certifiable. ISO 42001 is the international standard that turns those expectations into something an auditor can examine.
This page reflects how we see ISO/IEC 42001 applied in practice, and how it sits alongside Saudi AI guidance. It is not an official reference. The authoritative sources are the standard as published by ISO and IEC, and the instruments published by SDAIA — both of which may change. Confirm what applies to your organisation against them and with your own advisers.
What is ISO 42001?
From the compliance programmes we build and run in Saudi Arabia: ISO/IEC 42001 is the international standard for an artificial intelligence management system — an AIMS. Published in December 2023, it was the first standard of its kind, and it specifies how an organisation governs the AI systems it develops, provides or uses, across their full lifecycle.
It is a management system standard rather than a technical one. It does not tell you how to build a model, what accuracy to reach, or which architecture to choose. It asks who is accountable, what risks you identified, what you did about them, and whether you can show it.
If you have implemented ISO 27001, the shape will be familiar — the same clause structure, the same Statement of Applicability, the same audit rhythm. What is new is the subject matter: the probabilistic behaviour of AI systems, the governance of training data, the difficulty of explaining automated decisions, and the impact those decisions have on the people subject to them.
Where this sits in Saudi Arabia
The Kingdom's AI governance position is unusual, and understanding it is what makes the decision about ISO 42001 straightforward.
That gap is the practical case for ISO 42001. It operationalises very largely what SDAIA's principles ask for — risk classification, impact assessment, human oversight, transparency, accountability, documentation — and it produces a certificate from an accredited body. SDAIA also designed its principles with international interoperability in mind, referencing the EU AI Act's risk-based approach and OECD guidance, which means the two do not pull in different directions.
SDAIA has issued a substantial body of AI governance material, and sectoral regulators fold AI into their existing mandates.
- AI Ethics Principles, issued 2023, with a risk-tiered approach covering fairness, accountability, transparency and safety
- Generative AI Guidelines for Government, issued January 2024
- Generative AI Guidelines for the public
- An AI Adoption Framework built around maturity levels
- Sector treatment through SAMA for finance, the SFDA for health, and the DGA for the public sector
SDAIA's instruments are non-binding in themselves — they become enforceable where they connect to another law, the PDPL being the clearest case. A dedicated AI law is under development but is not in force.
- No mandatory AI-specific statute currently in force
- No national AI certification to hold or display
- No way to demonstrate to a buyer or a board that you meet the published expectations
- Real enforceability today runs through the PDPL where AI touches personal data
What the standard contains
Clauses 4 to 10 carry the management system requirements. Annex A carries the AI-specific controls — 38 of them across nine areas.
A documented policy for developing or using AI systems, aligned with the organisation's objectives and reviewed as those change.
Roles, responsibilities and reporting lines for AI governance, including who decides and who escalates.
The data, tooling, computing resources and human competence the AI systems in scope depend on, documented rather than assumed.
Assessing the consequences for individuals and groups affected by an AI system. This is the control with no equivalent in ISO 27001, and usually the hardest for a security-led team.
Governance across design, development, verification, deployment, operation and retirement — not only at release.
Provenance, quality, preparation and governance of training and operational data, which is where most AI risk actually originates.
What you tell users, customers and affected individuals about the AI system, its limitations, and how to contest an outcome.
Responsible use within the organisation, including where staff use AI systems the organisation did not build.
Suppliers, model providers and customers — who is accountable for what when the AI system is not entirely yours.
As in ISO 27001, you record applicability and exclusions in a Statement of Applicability with justification. In practice most organisations implement the majority of the controls, because they are written at a governance level rather than prescriptively, and AI risk is broad enough that few exclusions survive scrutiny.
Annex B gives implementation guidance for each control, Annex C sets out organisational objectives and risk sources, and Annex D covers applying the AIMS across different domains and sectors. The wider family is developing alongside it, including ISO/IEC 42005 on AI system impact assessment and ISO/IEC 42006 on requirements for bodies certifying AI management systems.
Who ends up needing it
Note that using AI counts, not only building it — which catches organisations that do not think of themselves as AI companies.
SDAIA's generative AI guidance is directed at government entities, and the DGA folds AI into digital government policy. A certified AIMS is how an entity evidences that the expectations were met rather than noted.
Models used in credit decisions, fraud detection or customer servicing sit inside an existing supervisory relationship. Governance evidence for those models is asked for as part of that relationship, not separately.
Where the SFDA's mandate applies, the impact assessment requirement in A.5 is not a formality — it is the substance of what a regulator will want to see.
AI governance questions have arrived in security questionnaires faster than any topic in recent memory, and most vendors have no structured answer. A certificate is a short answer to a long section.
A certified AIMS is evidence of structured AI governance and is widely used to support readiness — though certification alone does not establish conformity with the Act, and should not be presented as if it does.
Control A.9 covers responsible use of AI systems within the organisation. Staff using third-party AI tools on company data is a governance question whether or not anyone has framed it as one.
Running ISO 42001 in Govrly
An AIMS shares more with your existing programmes than the novelty of the subject suggests — and where it does not, the gap is worth seeing clearly.
Scope, leadership, competence, internal audit and management review carry across from ISO 27001 or ISO 27701 from the same records. If you hold either, clauses 4 to 10 are largely built.
Which AI systems you develop, provide or use, with their purpose, their data, their owner and their risk classification. Most organisations discover during this step that they have more AI in production than they thought.
Impact assessments held as records with owners, dates and outcomes — structured the same way as a privacy impact assessment, and often triggered by the same change.
Where training data contains personal data, the processing records, lawful bases and retention decisions supporting A.7 are the same records carrying your PDPL obligations.
A.10 asks who is accountable when the model is someone else's. Those relationships sit alongside your other suppliers rather than in a separate list nobody maintains.
All 38 controls with applicability, justification, status and evidence — updated as the AI estate changes, which for most organisations means frequently.
ISO 42001 alongside 27001 and 27701
Three management systems, one architecture. The overlap is larger than the subject matter suggests.
Because all three follow the harmonised structure, an organisation holding one has already built the management system layer for the others. The marginal work is the annex and the subject-specific requirements — which is why sequencing them onto one control library costs far less than running three programmes.
| ISO 42001 | ISO 27001 | ISO 27701 | |
|---|---|---|---|
| Subject | AI systems and their impact | Information security | Personal data governance |
| Published | December 2023 | 2022 edition | 2025 edition |
| Clause structure | 4 to 10 | 4 to 10 | 4 to 10 |
| Annex A | 38 controls, nine areas | 93 controls, four themes | Consolidated controller and processor controls |
| Distinctive requirement | AI system impact assessment | Risk treatment against Annex A | Controller and processor role determination |
| Prerequisite | None | None | None, since 2025 |
A practical path to an AIMS
The first two steps are where programmes succeed or stall, and neither is technical.
- 01/08Find the AI you already have
Models in production, AI features in purchased software, and tools your teams adopted without procurement. This inventory is almost always longer than expected, and you cannot scope an AIMS without it.
- 02/08Decide your role for each system
Developer, provider, or user. The standard treats these differently, and an organisation is usually several at once — building one model, embedding another, and using a third through a vendor's product.
- 03/08Classify by risk and impact
Not every AI system warrants the same governance. A risk-tiered approach concentrates effort where consequences fall on people, which is also how SDAIA's principles are structured.
- 04/08Extend the management system you have
If you hold ISO 27001 or ISO 27701, clauses 4 to 10 largely exist. Scope, leadership, competence, internal audit and management review extend rather than restart.
- 05/08Build the impact assessment process
This is the genuinely new work. It asks what an AI system does to the people subject to its outputs, which is a different question from what it does to your data — and one most security teams have never had to answer.
- 06/08Govern the data, not just the model
Provenance, quality and preparation of training data under A.7. Where that data contains personal data, this is shared with your PDPL work rather than separate from it.
- 07/08Complete the Statement of Applicability
All 38 controls, with justification for anything excluded. Expect exclusions to be challenged more than in a security certification, because AI risk cuts across functions.
- 08/08Run it, then engage a certification body
Confirm they are accredited for ISO 42001 specifically. Because the standard is comparatively new, the pool is smaller than for ISO 27001 and worth checking early.
Your AI system register, impact assessments and Statement of Applicability sit on the same control library as your NCA ECC and ISO 27001 programmes — while the data governance behind A.7 draws on the processing records already carrying your PDPL obligations, rather than a second inventory maintained in parallel.
ISO 42001 — FAQs
The international standard for an artificial intelligence management system, or AIMS. Published in December 2023, it was the first standard of its kind, and it specifies how an organisation governs the AI systems it develops, provides or uses across their full lifecycle. It is certifiable by an accredited certification body.
No. SDAIA has issued AI Ethics Principles, generative AI guidelines and an AI adoption framework, but these are non-binding in themselves — they become enforceable where they connect to another law, most clearly the PDPL where AI touches personal data. A dedicated AI law is under development but is not in force. ISO 42001 is voluntary, and it is currently the practical way to evidence that the published expectations have been met.
It evidences structured AI governance; it does not discharge anything. SDAIA's instruments remain what they are, the PDPL applies regardless where personal data is involved, and a forthcoming AI law will apply on its own terms. What a certificate does is demonstrate to a regulator, a buyer or a board that the governance exists and has been independently examined.
Annex A contains 38 controls grouped into nine areas, running from A.2 to A.10 — AI policy, internal organisation, resources, impact assessment, lifecycle, data, information for interested parties, use of AI systems, and third-party relationships. Applicability is recorded in a Statement of Applicability with justification for exclusions, as in ISO 27001.
No — ISO 42001 can be implemented and certified independently. That said, if you hold ISO 27001 or ISO 27701 the work is considerably shorter, because all three follow the same harmonised clause structure. The management system layer transfers; the AI-specific annex and the impact assessment process are the new work.
Yes. The standard addresses organisations that develop, provide or use AI systems, and control A.9 covers responsible use within the organisation specifically. Staff using third-party AI tools on company data is a governance question whether or not it has been framed as one, and it is frequently where the largest unmanaged exposure sits.
The requirement in A.5 to assess the consequences of an AI system for the individuals and groups affected by its outputs. It is the requirement with no equivalent in ISO 27001, and usually the hardest for a security-led team, because it asks what the system does to people rather than what it does to data.
A certified AIMS is widely used as evidence of structured AI governance and supports readiness for the Act's obligations. Certification alone does not establish conformity with the Act, and presenting it that way is a claim that will not survive scrutiny. Notably, SDAIA designed its own principles with international interoperability in mind, referencing the Act's risk-based approach, so the two do not pull in opposite directions.
It depends far more on the size of your AI estate and the maturity of your existing management system than on the standard. Organisations already certified to ISO 27001 or ISO 27701 move faster, because the clause requirements are shared. The inventory and impact assessment work is usually what sets the timeline.
Yes. Controls, evidence and records are shared objects, so the management system requirements carry across from your ISO 27001 or ISO 27701 programme, the data governance behind A.7 draws on the processing records supporting your PDPL obligations, and each framework keeps its own scope, scoring and reporting.
Start by finding the AI you already run
Most organisations have more AI in production than their inventory shows. See what an AI system register looks like against your existing control library.
