CAIRNCYBER ADVISORY, homeContact

Insights

ISO 42001 readiness without building a second bureaucracy

· 5 min read · Gary Johnston

ISO/IEC 42001 asks for a management system for AI. With ISO 27001 in place most of the structure exists already; readiness is about the parts that are new.

ISO/IEC 42001 has moved quickly from a standard people had heard of to a line in due-diligence questionnaires. A client asks whether you are certified, a board asks whether you should be, and somebody is asked to find out what it would take.

The honest answer is usually less than feared, provided the organisation resists the instinct to build a new governance function around it. 42001 asks for a management system for AI: a way of setting direction, assessing risk, applying controls and checking that they work. Most regulated organisations already run one of those for information security. Readiness is mostly about extending it to cover what is genuinely different about AI.

What the standard actually asks for

42001 uses the same harmonised structure as ISO 27001. Clauses 4 to 10 cover context, leadership, planning, support, operation, performance evaluation and improvement, and anyone who has run an ISMS will recognise the shape: a scope, a policy, objectives, a risk assessment, a Statement of Applicability, internal audit and management review.

Three things are new in substance.

An AI risk assessment and treatment that covers the failure modes specific to AI systems, not only the confidentiality, integrity and availability of the data around them.

An AI system impact assessment. This is the genuinely new artefact. It asks what a system could do to individuals, groups and society, not just to the organisation. A DPIA covers part of that ground for personal data, and the two can share evidence, but they are not the same assessment.

Controls across the AI system life cycle, set out in Annex A: policy, internal organisation, resources, impact assessment, the life cycle itself, data, information for interested parties, use of AI systems, and third-party and customer relationships.

The standard also asks you to be clear about your role. Most regulated firms are chiefly users of AI systems that someone else built, with a growing amount of in-house development on top. The role shapes which controls apply, and how much of the evidence has to come from your suppliers.

What to have in place before anyone mentions an auditor

  1. A scope you can defend. Which AI systems, which parts of the organisation, which roles. Too broad and the programme never finishes; too narrow and the certificate reassures nobody. Starting with the systems a client or regulator is actually asking about is usually right.
  2. An inventory that includes the AI you did not build. The model inside a SaaS product counts. This is where most inventories fail, and the gap between the register and what the procurement pipeline and gateway logs show is itself a finding.
  3. An AI policy and objectives that leadership has actually signed, with named accountable roles. Short and specific beats comprehensive.
  4. The risk method you already use, extended with AI failure modes: prompt injection, data leakage through context, tool misuse, model change, over-reliance on output. A parallel risk scale just for AI will be ignored within a year.
  5. An impact assessment template people can finish in an afternoon for a typical system, with a route to escalate the ones that need more.
  6. Life-cycle gates that are enforced, not described. Approval before production, review on material change, monitoring in operation, and a retirement decision. A gate that lives in the deployment pipeline or the AI gateway produces its own evidence.
  7. Supplier management that covers model providers: what they commit to, how you would know if the model changed underneath you, and what you would do if they withdrew it.
  8. Evidence that comes from running systems rather than attestation, for the reasons set out in Evidence, not attestation. That is the difference between a management system that survives its second surveillance audit and one that is rebuilt every year.
  9. An internal audit and a management review that have actually happened, at least once, before an external auditor asks to see them.

If you already run ISO 27001

Treat 42001 as an extension of the same management system, not a second one. Document control, internal audit, management review, corrective action and competence records can all be shared. The AI-specific parts sit on top: the impact assessment, the AI risk criteria, the Annex A controls that have no 27001 equivalent, and the life-cycle gates.

Organisations that run the two separately end up with two policy sets, two risk registers and two audit calendars describing the same estate, and a permanent reconciliation job between them.

Certification or conformance

Not everyone who asks about 42001 needs the certificate. Some clients need you to answer their questions credibly and show the evidence behind the answers; for them, operating in line with the standard is enough, and considerably cheaper. Others, often in regulated supply chains, will want the certificate itself.

If you do go for certification, check that the certification body is accredited for ISO/IEC 42001 specifically, by UKAS in the UK. Accreditation for ISO 27001 does not carry over: bodies certifying AI management systems are assessed against additional requirements in ISO/IEC 42006, and UKAS has only been granting that accreditation since late 2025, so the list is short. UKAS publishes it.

Either way, the work is the same until the last step. Build the management system so it would pass, and the decision about certification becomes commercial rather than technical.

Written by Gary Johnston, Principal Security Architect at Cairn Cyber Advisory Ltd. CISSP, CCSP. Nothing here describes any client's environment.

More insights

Start a conversation

Send me an outline of the problem and your timescale. I normally reply within one working day and will tell you directly whether I think I can help.