You can run ISO/IEC 27701:2025 and ISO/IEC 42001:2023 as one integrated system for AI that uses personal data. In our reading, the privacy system owns risks to personal data and the people it describes, and the AI system owns its wider impacts. ISO/IEC 27701 has stood alone since 14 October 2025, so the pair no longer needs ISO/IEC 27001 underneath.
This guide is for DPOs, privacy leads and AI governance leads. Neither standard makes your AI lawful, so it also sets out what the EDPB, the EU AI Act and Gulf laws expect.
Do ISO/IEC 27701 and ISO/IEC 42001 fit together in 2026?
ISO/IEC 42001:2023 is Edition 1, and iso.org lists no revision. ISO's FAQ says ISO/IEC 27701:2025 can be used alone as an independent management system standard. Pages that still call it an extension of ISO/IEC 27001 describe the withdrawn 2019 edition. Our guides to ISO/IEC 27701:2025 and ISO/IEC 42001:2023 cover each one alone.
ISO's list of management system standards marks ISO/IEC 42001:2023 as Harmonized Structure but does not flag ISO/IEC 27701:2025. Accreditation bodies describe the 2025 edition as adopting the harmonized structure. ISO says shared text, terms and definitions help an organisation run one integrated system.
In our reading, choose one combined scope or two linked scopes. Then share one risk method covering risks to individuals and AI impacts, one internal audit programme and one management review.
Who covers what when AI uses personal data?
The standards columns are AGS's reading, not text from either standard.
| Risk or duty | ISO/IEC 27701:2025 PIMS | ISO/IEC 42001:2023 AIMS | Law and regulator |
|---|---|---|---|
| Lawful basis for training and use | Records the basis per processing activity | Records the AI system's intended purpose | GDPR Art. 6, EDPB 28/2024 three-step test |
| Is the model anonymous? | Owns the anonymity assessment as a privacy risk | Supplies model test results | EDPB 28/2024: extraction and query risk insignificant |
| Automated decisions | Notices, access replies, objection and review routes | Human oversight in system design | GDPR Arts 13 to 15 and 22, CJEU C-634/21 and C-203/22, UAE PDPL Art. 18 |
| Impact assessment | DPIA | AI system impact assessment (ISO/IEC 42005 guidance) | GDPR Art. 35, AI Act Arts 26(9) and 27(4), UAE PDPL Art. 21, Saudi Implementing Regulation Art. 25 |
| Bias testing with sensitive data | Record of processing entry and deletion evidence | Bias detection and correction method | AI Act Art. 4a, added by Regulation (EU) 2026/1744 |
| Bought-in models | Supplier privacy due diligence | Third-party AI supplier controls | EDPB 28/2024: deployer checks lawful development |
| Internal audit and management review | One programme and one review for both | GDPR Art. 5(2): you must show compliance | |
What does EDPB Opinion 28/2024 expect you to document?
The EDPB adopted Opinion 28/2024 on AI models on 17 December 2024. It is addressed to supervisory authorities, but it shows what they are asked to check.
A model trained on personal data cannot in all cases be considered anonymous. It is anonymous only if two likelihoods are insignificant, counting any means reasonably likely to be used by anyone: extracting training data from the model, and obtaining it through queries. Tests show resistance only to the attacks tested, such as membership inference, model inversion and regurgitation.
For an anonymity claim, an authority should ideally check DPIA information, including any decision that none was needed, and the DPO's advice. It also looks at design measures with a threat model, re-identification tests, and documentation given to deployers and data subjects. If anonymity cannot be confirmed, it may find an accountability failure.
Legitimate interest needs a three-step test: a clearly articulated lawful interest, necessity, and balancing against people's rights and reasonable expectations. Example safeguards include pseudonymisation, an unconditional opt-out, output filters against regurgitation and publishing the balancing test. A proposed Article 88c in the Digital Omnibus, COM(2025) 837 of 19 November 2025, would expressly allow legitimate interests for AI development and operation.
Buying a model that still holds personal data does not end the question. The deployer should assess, in proportion to the deployment risk, that it was not developed unlawfully, looking at data sources and any finding by an authority or court.
When is an AI decision "solely automated" under Article 22?
Article 22(1) bars decisions based solely on automated processing, including profiling, with legal or similarly significant effects. The exceptions are contract necessity, authorisation by law and explicit consent. For contract and consent, Article 22(3) requires at least human intervention and the right to express a view and contest.
In SCHUFA (C-634/21, 7 December 2023), the Court of Justice ruled on a credit agency's automated score. It was itself an Article 22 decision where a third party drew strongly on it to decide on a contract. In our reading, that puts the score provider in scope, not only its client.
In Dun & Bradstreet Austria (C-203/22, 27 February 2025), the Court held that an Article 15(1)(h) access reply must explain the procedure and principles actually applied, concisely and intelligibly. Claimed trade secrets go to the supervisory authority or court, which decides how much the person sees.
Changes to Article 22 are also proposed, not in force. The Digital Omnibus would let contract-necessity decisions be solely automated whether or not they could be taken otherwise. The European Parliament lists it as awaiting committee decision.
What did Regulation (EU) 2026/1744 change for privacy teams?
Regulation (EU) 2026/1744 amended the EU AI Act and has been in force since 27 July 2026. High-risk duties now start on 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems.
New Article 4a replaces the deleted Article 10(5). Providers of high-risk systems may exceptionally process special-category data, as far as strictly necessary to detect and correct bias. All of these conditions must be met.
- other data, including synthetic or anonymised data, would not do
- re-use is technically limited, with state-of-the-art security and pseudonymisation
- access is strictly controlled and documented, with no transmission to other parties
- data is deleted once bias is corrected or retention ends, whichever comes first
- the record of processing states why it was strictly necessary
Article 4a(2) extends this to other AI systems and models and to deployers of high-risk systems. It applies where bias is likely to harm health and safety or fundamental rights, or to cause unlawful discrimination, and the same conditions are met. It creates no duty to test for bias.
Under Article 26(9), deployers of high-risk systems use the provider's Article 13 information for their DPIA. From 2 December 2027, where Article 27 requires a fundamental rights impact assessment, the replaced Article 27(4) lets the deployer cross-reference DPIA sections that already meet those duties. In our reading, the DPIA becomes the base document.
What do UAE and Saudi rules add for automated decisions?
EU rules can also reach a Gulf company, through GDPR Article 3(2) and AI Act Article 2(1)(c).
Under the UAE PDPL, Federal Decree-Law 45 of 2021, Article 18 lets a person object to decisions from automated processing, including profiling, especially those with legal or adverse effects. The right falls away where the processing is agreed by contract, required by UAE law or based on prior consent. Even then, the controller must include the human element in reviewing a decision on request.
Article 21(2)(a) requires an impact assessment where automated processing, including profiling, systematically and fully assesses personal aspects with legal consequences or serious impact.
The federal law excludes free-zone companies with their own data protection law, such as those in the Dubai International Financial Centre (DIFC). DIFC announced Regulation 10 on 7 September 2023, covering autonomous and semi-autonomous systems. It requires clear notice to people whose data a system processes, and human-defined purposes must prevail over system-defined ones.
In Saudi Arabia, Article 25 of the PDPL Implementing Regulation requires a written impact assessment for, among other cases, large-scale, repetitive processing involving newly adopted technologies or automated decisions. SDAIA's AI Ethics Principles (2025) list Privacy and Security among seven principles and treat social profiling as an unacceptable risk.
Meeting one of these laws does not meet another. Our guide to Saudi PDPL enforcement covers the Saudi side.
How AGS can help
The AI Privacy Management System holds 168 documents: 80 Word, 87 Excel and 1 PDF. It includes a manual, 14 policies, 18 procedures and an AI privacy impact assessment set with a screening checklist, methodology, report template, register and action plan. Records include an AI decision privacy assessment and a human review privacy record. It has crosswalks to ISO/IEC 27701:2025 and the NIST Privacy Framework 1.0, pre-fills no legal content and is not an AI management system.
Frequently asked questions
Is ISO/IEC 27701 certification a GDPR certification?
No. France's CNIL says ISO/IEC 27701 is not GDPR-specific and is not, as such, a certification under GDPR Article 42. CNIL wrote this about the 2019 edition.
Can an authority order an AI model to be deleted?
Yes, where proportionate. The EDPB lists erasure of part or all of the dataset, or of the AI model itself, among corrective measures after unlawful processing.
Is there an ISO standard just for AI privacy?
Not a published one. ISO/IEC 27091, on privacy protection in AI, is a final draft (FDIS) registered for formal approval on 13 July 2026. It gives guidance on privacy risks across the AI life cycle.
AGS provides independent documentation toolkits. This guide and the referenced kit are not affiliated with, endorsed by or approved by ISO, IEC, the European Data Protection Board, the European Commission or any data protection authority. It is general information, not legal advice.
Reviewed October 2026. Facts checked against the sources current at publication.
