Skip to content

Which ISO 27001:2022 controls cover your use of AI

By Sam Rivera, Founder, SentinelPanda · November 7, 2026 · 4 min read · ISO 27001

An auditor will not ask whether you have an AI control. They will ask whether the AI tools in use are inside your risk assessment and your Statement of Applicability, which is a harder question to answer late.

Start with the management system, not the controls

Clause 4.1 asks you to determine the internal and external issues relevant to your ISMS. Staff use of AI tools, AI features inside existing SaaS and AI in your own product are all such issues. Clause 6.1.2 then requires a repeatable information security risk assessment, and 6.1.3 a risk treatment process whose output includes the Statement of Applicability.

If AI is missing from those three places, a new policy will not help much at audit. If it is present, the Annex A controls below are mostly a matter of applying what you already run to a new class of asset and supplier.

Know what is in use

The asset inventory (A.5.9) should list AI tools and AI-enabled services, each with an owner. Acceptable use (A.5.10) is where the rule for staff lives: which tools are approved, and what information may and may not be entered into them. Classification (A.5.12) gives that rule its vocabulary. "Nothing confidential goes into an external model" is only enforceable once confidential is defined.

Treat AI vendors as suppliers and cloud services

Most AI is consumed as a hosted service, so the supplier controls apply in full: security in supplier relationships (A.5.19), security within supplier agreements (A.5.20), the ICT supply chain (A.5.21) and monitoring and change management of supplier services (A.5.22). So does the cloud services control (A.5.23), which expects defined processes for acquiring, using, managing and leaving a cloud service.

For AI, the agreement terms that matter are retention of prompts and outputs, use of your data for model training, data location, subprocessors and what happens to your data when you leave. A vendor that changes those terms, or adds AI features to a service you already use, is a supplier change under A.5.22.

Control what AI systems can reach

Access control (A.5.15), access rights (A.5.18), privileged access rights (A.8.2) and information access restriction (A.8.3) apply to AI integrations as they do to people. An assistant connected to a document store or a ticketing system can read whatever its service account can read. Distinct identities and least-privilege grants for integrations are what an auditor will look for; one shared API key with broad scope is the finding.

Keep data where it belongs

Data leakage prevention (A.8.12) is the control most directly relevant to generative AI, because the leakage path, pasting into a prompt, looks like ordinary work. Measures range from blocking unapproved AI domains and inspecting uploads to enterprise tenancy with retention switched off. Information transfer (A.5.14), data masking (A.8.11) and information deletion (A.8.10) support it, and privacy and protection of PII (A.5.34) applies wherever personal data reaches a model.

If you build with AI

Organisations putting AI into their own products also rely on the secure development life cycle (A.8.25), application security requirements (A.8.26), secure coding (A.8.28) and security testing in development and acceptance (A.8.29). Prompt injection and unsafe handling of model output are application security requirements to specify and test. Logging (A.8.15) and monitoring (A.8.16) should capture what AI-driven features do, attributable to an identity.

Where ISO/IEC 42001 comes in

ISO/IEC 27001 protects information. It does not ask whether an AI system is fit for purpose, whether its outputs are fair, or whether the people affected by it are informed. ISO/IEC 42001 sets up an AI management system for those questions using the same high-level structure as 27001, so the two can share one management system: one risk process, one internal audit programme, one management review.

For organisations that use AI rather than build it, 27001 with AI properly in scope is usually the right first step. For those whose products are AI systems, 42001 alongside it is increasingly what customers ask for.

The short version

  • Put AI use in the ISMS context, risk assessment and Statement of Applicability.
  • List AI tools and AI-enabled services as assets with owners.
  • Write the acceptable-use rule in terms of your classification scheme.
  • Apply the supplier and cloud controls, with retention and training terms in the agreement.
  • Give integrations their own identities and least-privilege access.
  • Point data leakage prevention at the prompt path, not only email and USB.
ISO 27001 Annex A technological controls ISO 27001 Statement of Applicability ISO 42001 vs NIST AI RMF AI security risks across frameworks

Run your compliance program in one workspace.