Skip to main content
    Back to Blog
    11 min readLast reviewed:
    ISO 27001
    Information Security
    AI Governance
    ISMS
    Compliance

    ISO 27001 for AI: Annex A Controls, ISO 42001 and Certified Vendors

    T

    Team Sonomos · Sonomos

    The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.

    Short answer, checked 18 November 2026: ISO/IEC 27001 never mentions AI, but it already covers your use of AI tools. Under the 2022 edition, an AI assistant is a supplier (Annex A controls 5.19 to 5.22), a cloud service (5.23), a route for transferring information (5.14) and a channel for leakage (8.12), and your ISMS has to treat it as all four. ISO/IEC 42001 is a different standard: it certifies an AI management system, and it matters when you build, provide or govern AI systems, not when staff simply use a chatbot inside a 27001 program.

    Most large AI vendors now hold both certificates for their business products. OpenAI, Anthropic, Google (for Gemini in Workspace), Microsoft (for its Copilots), Cursor and the legal AI company Harvey list ISO 27001 and ISO 42001 on their own trust pages; Perplexity and Cognition, the owner of Windsurf, list ISO 27001 only. None of those certificates covers your use of the tool, and most exclude consumer plans.

    This guide maps the Annex A controls that AI use touches, explains how 42001 differs from 27001, and lists each vendor's certificates with the page we checked.

    ISO 27001 and ISO 42001: which one covers AI

    ISO/IEC 27001:2022 was published on 25 October 2022. It sets requirements for "establishing, implementing, maintaining and continually improving" an information security management system, including the assessment and treatment of information security risks. Its Annex A reference list was cut from 114 controls to 93, grouped into organizational, people, physical and technological themes, and gained 11 new controls, among them cloud services, information deletion, data masking and data leakage prevention, as the certification body DQS summarised when the guidance standard was revised.

    The 2013 edition is gone. Accreditation bodies following IAF MD 26 required certification bodies to have completed transitioning their clients by 31 October 2025, so a vendor certificate you accept today should name the 2022 edition.

    ISO/IEC 42001:2023, published on 18 December 2023, specifies requirements for an AI management system, for organizations of any size that provide or use products or services built on AI systems. It adds what a security standard lacks, such as AI risk assessments, assessments of an AI system's impact on people, and controls over the AI system lifecycle.

    ISO/IEC 27001:2022ISO/IEC 42001:2023
    What it certifiesAn information security management systemAn AI management system
    Main questionIs information kept confidential, intact and available?Is AI developed, provided or used responsibly, with its risks and impacts managed?
    Typical holderAny organization handling sensitive informationAI vendors, and companies building AI into products
    Do you need it to use ChatGPT at work?If you are certified, your ISMS must cover that useNo; consider it if AI is part of what you sell

    The two share the same management-system structure, so a 42001 program can sit beside an existing ISMS. For AI risk outside both standards, our guides to the NIST AI RMF and the EU AI Act cover the frameworks auditors often map alongside them.

    The Annex A controls that AI tool use touches

    The numbering below is the 2022 Annex A. The right-hand column is the evidence an auditor can actually test, which is where most AI gaps show up.

    ControlWhat it means for AI toolsEvidence to keep
    5.9 Inventory of information and other associated assetsAI tools, plugins and agents are assetsAn AI tool register with owners
    5.10 Acceptable use of information and other associated assetsRules on what may go into which toolAn AI use policy, acknowledged by staff
    5.12 Classification of informationEach class mapped to the tools allowed to receive itClassification policy naming AI tools
    5.14 Information transferA prompt is a transfer to a third partyTransfer rules covering AI prompts and uploads
    5.19 to 5.22 Supplier relationships, agreements, ICT supply chain, monitoringAI vendors and their model sub-processorsRisk assessments, contracts, annual reviews
    5.23 Information security for use of cloud servicesEvery hosted AI serviceCloud service policy, including exit
    5.34 Privacy and protection of PIIPersonal data in promptsDPAs, transfer mechanisms, DPIAs
    6.3 Information security awareness, education and trainingWhat staff may pasteTraining records with AI content
    8.10 Information deletionChat retention and deletion settingsConfigured retention, deletion tests
    8.11 Data maskingPseudonymising data before it reaches a modelMasking procedure and tool configuration
    8.12 Data leakage preventionAI sites and apps as leakage channelsDLP rules covering AI destinations
    8.15 and 8.16 Logging, monitoring activitiesUsage and admin logs from AI toolsLog retention and review records
    8.28 Secure codingCode written with AI assistantsReview and testing of AI-generated code

    Three of these deserve more than a table row. Control 5.23 was new in 2022 and asks for processes covering the acquisition, use, management and exit of cloud services, which means an AI vendor review should cover how you would leave, including exporting and deleting chat history. Control 8.11 is where data masking and pseudonymization live, and masking sensitive values before a prompt is sent is the most direct way to keep personal data out of a vendor's systems. Control 8.12 is the data loss prevention control, and AI destinations need to be in its scope, whether you inspect traffic at the network or on the device; our comparison of cloud DLP with on-device masking sets out the trade-off.

    If your acceptable use rules for 5.10 do not yet mention AI, our AI use policy generator drafts the sections an auditor will look for.

    Which AI vendors hold ISO 27001 and ISO 42001

    We checked each vendor's own trust portal or documentation on 18 November 2026. Scope matters as much as the certificate, so the table says which products each page names.

    Vendor and products namedISO 27001ISO 42001Source
    OpenAI: ChatGPT Enterprise and Edu, APIYes (2022), with 27017, 27018 and 27701Listed for ChatGPT Enterprise and Edu; confirm the API's scopeOpenAI Trust Portal
    Anthropic: Claude API, Claude Enterprise, Claude TeamYesYesAnthropic Trust Center
    Google: Gemini in Workspace apps and the Gemini app for Workspace usersYes, with 27017, 27018 and 27701YesWorkspace generative AI privacy hub
    Microsoft: Microsoft Copilot, Copilot Chat, GitHub Copilot, FoundryYes, listed for CopilotYesMicrosoft ISO 42001 page
    Cursor (Anysphere)Yes (2022)YesCursor security
    Harvey (legal AI)Yes, with 27701YesHarvey Trust Center
    PerplexityYes (2022)Not listedPerplexity Trust Center
    Cognition (Devin, and Devin Desktop, formerly Windsurf)Yes (2022), on the Public Sector viewNot listedCognition Trust Center

    Notes on scope. Anthropic's scope table covers its commercial products and does not list the consumer Free, Pro or Max plans. Google's privacy hub says Gemini Notebook and Gemini in Chrome do not support its ISO, SOC or FedRAMP compliance. Microsoft's Copilot privacy page, updated 4 November 2026, lists GDPR, ISO 27001, HIPAA and ISO 42001 for Microsoft Copilot; GitHub added Copilot Business and Enterprise to its ISMS scope in May 2024. For Cognition, the ISO entry sits on a separate Public Sector view of its trust center, so ask which products and sites the certificate covers. Our Windsurf security guide covers that vendor in detail.

    What a vendor's certificate does and does not tell you

    A certificate says that an accredited auditor found the vendor's management system conformed to the standard, for a stated scope, on the audit dates. It does not say three things buyers often read into it.

    It does not cover the plan you use. Scope statements name products and sites, and consumer apps are usually outside them, so a team on personal ChatGPT, Claude or Gemini accounts gets nothing from the vendor's certificate. That is the shadow AI gap, and no vendor certificate closes it.

    It does not certify your use. Microsoft says it plainly on its 42001 page: you can use its certification in your own assessment, but "You're responsible, however, for engaging an assessor to evaluate the controls and processes within your own organization and your implementation." The same is true of 27001.

    It does not settle training or retention. Whether a vendor trains on your prompts or keeps them for 30 days is a contract and settings question, answered in its terms and data processing agreement, not in its certificate. Ask for the certificate, the scope statement and the Statement of Applicability, and read them next to the contract. For the parallel question under SOC 2, see our guide to SOC 2 and AI tools.

    Evidence to have ready before the audit

    When an auditor asks about AI, the questions are predictable: which AI tools are in use, how they were assessed, what the contracts say, what staff were told, and how you would know if confidential information reached an unapproved tool. A short AI addendum to the ISMS answers them without rewriting existing documents:

    • an AI tool register under 5.9, including tools found in use but not approved;
    • risk assessment entries for AI-specific risks, such as confidential data entered into consumer accounts, prompt injection through AI agents and inaccurate output acted on;
    • supplier files under 5.19 to 5.22 holding each vendor's certificate, scope and contract;
    • Statement of Applicability entries for 5.23, 8.10, 8.11 and 8.12 that name AI tools;
    • training records and policy acknowledgements under 6.3 and 5.10;
    • incident records for any accidental disclosure to an AI tool, with the response taken.

    For GDPR overlap on the same evidence, see our GDPR and AI compliance guide.

    Frequently asked questions

    Does an AI vendor need its own ISO 27001 certificate before we can use it?

    No. ISO 27001 requires you to assess supplier risk and put security requirements in supplier agreements; it does not require suppliers to be certified. A certificate with a matching scope, a SOC 2 Type II report or a reviewed security questionnaire can each support the assessment. A vendor with none of them is a higher risk to record and accept explicitly.

    Do we need ISO 42001 if we only use AI tools like ChatGPT?

    Usually not. If you are ISO 27001 certified, your existing ISMS must cover how staff use AI tools. ISO 42001 becomes relevant when AI is part of your product or service, or when customers ask you to show how you govern AI systems you build or deploy.

    Is ChatGPT ISO 27001 certified?

    OpenAI's trust portal lists ISO/IEC 27001:2022, along with 27017, 27018 and 27701, for ChatGPT Enterprise and Edu and for its API, as checked on 18 November 2026. The portal is about OpenAI's business services, so do not assume the certificate covers a personal Free or Plus account.

    Does ISO 27001 require us to ban consumer AI tools?

    No. The standard requires risk-based controls, not prohibitions. You can permit consumer tools for public information and require approved business accounts for anything confidential, provided the rule is in your policy, your risk assessment records the decision, and you can show how it is enforced.

    Put AI inside the ISMS you already have

    ISO 27001 compliance and everyday AI use fit together once the tools are treated as what the standard already calls them: suppliers, cloud services and channels for transferring information. Register them, assess them, contract with them, mask what should not leave, and collect evidence that the rules work. The vendor certificates help with the supplier controls; the masking, leakage and training controls remain yours. Our page for technology teams covers the engineering side of the same work.

    Locke, made by Sonomos, runs on your own computer and replaces names, account numbers, health details and other sensitive values with placeholders before a prompt reaches the AI tools it screens, so the provider never receives them. For the tools it covers, it is one way to put control 8.11 into practice at the point where staff type, though coverage depends on the app, platform and settings, and it does not certify anything on its own. You can read how Locke works.

    Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works

    This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.

    Reference pages for this post