ISO 42001 was published in December 2023. It is the first international standard specifically for AI management systems — a certifiable framework that defines how organisations should govern, monitor, and improve their use of AI. Two and a half years on, most teams deploying AI agents still haven't looked at it.
That is starting to create a problem. Enterprise procurement teams are adding AI governance questions to vendor questionnaires. Regulated industries are treating documented AI management as a baseline expectation. And the EU AI Act, which will be fully enforceable for high-risk systems from August 2026, explicitly references structured risk management processes that map closely to what ISO 42001 describes.
This article explains what the standard actually requires — not at a marketing level, but at the level of what your team would need to do differently.
What ISO 42001 is (and what it isn't)
ISO 42001 is a management system standard, not a technical specification. It doesn't tell you which AI model to use, how to train it, or what your architecture should look like. What it does is define how an organisation should govern its AI activities — the policies, processes, roles, risk assessments, and monitoring practices that should exist around any AI system you deploy or develop.
If your organisation has ever implemented ISO 27001 (information security) or ISO 9001 (quality management), the structure will be immediately familiar. It follows the same high-level structure — clauses 4 through 10, built around the Plan-Do-Check-Act cycle — with AI-specific requirements layered on top. Organisations already certified to ISO 27001 can often implement 42001 with significantly less effort because the management system foundations are already in place.
The seven clauses that matter in practice
The standard's mandatory requirements sit in clauses 4 to 10. Here is what each one asks of your team in plain terms:
| Clause | What it requires | What that looks like |
|---|---|---|
| 4 — Context | Understand your organisation's AI landscape and stakeholder expectations | Map every AI system in use, who uses it, and what decisions it influences |
| 5 — Leadership | Senior leadership must own AI governance and establish a documented AI policy | A written policy, signed off at exec level, covering responsible AI use across the org |
| 6 — Planning | Risk assessments for each AI system, with documented objectives and treatment plans | Formal assessment of what each agent can do wrong and how you'll prevent or catch it |
| 7 — Support | Competence, awareness, and documentation controls | Training records, maintained documentation, communication plans |
| 8 — Operation | Implement and control the processes that manage your AI risks | Active governance: audit logging, access controls, human oversight mechanisms |
| 9 — Evaluation | Monitor and measure whether your controls are actually working | Regular reviews, internal audits, metrics on AI system behaviour |
| 10 — Improvement | Act on what you find — address failures, update controls, close gaps | Documented corrective actions when incidents occur or audits surface problems |
The standard also includes Annex A, which contains 38 specific controls grouped into areas including AI system impact assessment, data management, human oversight, transparency, and system lifecycle management. Annex A controls are not all mandatory — you select the ones applicable to your context and document why you've excluded any that aren't relevant.
How it compares to other AI compliance frameworks
ISO 42001 is often mentioned alongside the EU AI Act and the NIST AI Risk Management Framework. They are not alternatives. They approach the same problem from different angles:
| Framework | Type | What it gives you |
|---|---|---|
| EU AI Act | Legal regulation | Mandatory requirements for high-risk AI systems in the EU market. Non-compliance carries fines up to €35M or 7% of global revenue. |
| NIST AI RMF | Voluntary framework | A structured way of thinking about AI risk across four functions: Govern, Map, Measure, Manage. No certification. |
| ISO 42001 | Certifiable standard | A documented, auditable management system. Third-party certification available. Increasingly requested in enterprise procurement. |
The practical relationship: if you implement ISO 42001 properly, you will have satisfied a large portion of the EU AI Act's requirements for high-risk systems — because both require structured risk management, documented human oversight, and audit trails. ISO 42001 doesn't make you EU AI Act compliant by itself, but it makes compliance significantly easier to demonstrate.
Do you actually need certification?
Certification is not the only reason to implement ISO 42001. Many organisations use the standard as an internal framework without pursuing third-party certification. That is entirely legitimate and often the right first step.
Certification makes sense when:
- Enterprise customers are asking for proof of AI governance in vendor questionnaires — a certificate is a faster answer than explaining your internal processes
- You are selling AI tools or services into regulated industries (financial services, healthcare, critical infrastructure) where third-party assurance carries weight
- You want external validation that your governance posture is sound, not just internally assessed
- You are preparing for EU AI Act compliance and need to demonstrate a documented management system to a regulator
What you need to build before you can pass an audit
If you are starting from scratch, here is the practical sequence of what needs to exist before you could credibly claim ISO 42001 alignment:
-
An AI system inventory. A documented list of every AI system your organisation uses or develops — what it does, what data it processes, what decisions it influences, and who is accountable for it.
-
A written AI policy. Approved at senior leadership level. Covers your organisation's principles for responsible AI use, the scope of your AI management system, and how you will handle AI-related risks.
-
Risk assessments for each AI system. Not a generic AI risk register — individual assessments that identify what could go wrong with each specific system, how likely it is, and what you're doing about it.
-
Operational controls you can demonstrate, not just describe. Audit logs, access controls, human oversight mechanisms, and incident response processes that are technically active — not procedures that live in a document but aren't enforced.
-
An evaluation and review process. Evidence that you are regularly reviewing whether your controls are working — internal audits, management reviews, metrics on AI system behaviour over time.
The gap most organisations have is not in the documentation. It's in item four. Policies and risk assessments are relatively straightforward to produce. The harder question is whether the controls those documents describe actually exist and are running every day — and whether you have evidence to prove it.
The difference between documented governance and real governance
ISO 42001 auditors look for evidence, not paperwork. A policy document that says "all AI tool calls are logged" is worthless if the logs don't exist, are incomplete, or can't be produced on request. A risk assessment that says "human oversight is maintained" is meaningless if there's no mechanism by which a human can actually intervene in real time.
This is the distinction that separates organisations that will pass an ISO 42001 audit from organisations that will be surprised when they don't. Technical governance — the systems and tools that enforce your policies automatically — is what makes documentation credible. Without it, you have a compliance theatre problem: the appearance of governance without the substance.
The bottom line
ISO 42001 is young as standards go, and the certification market around it is still developing. But the underlying requirement — that organisations deploying AI should be able to demonstrate structured, documented, continuously monitored governance — is not going away. It is the direction every major regulatory framework is heading.
The teams that will find ISO 42001 certification straightforward are not the ones that study the standard hardest. They are the ones that already have their AI systems under governance: logged, monitored, with clear policies that are technically enforced rather than just written down. For them, certification is mostly an audit of what's already true. For everyone else, it's a gap analysis followed by an implementation project — and the longer you wait, the more agents you have to retrofit.
The technical controls ISO 42001 requires — already built.
TrustLoop gives you tamper-evident audit logs, documented human oversight, real-time policy enforcement, and PII masking across all your AI agents — the operational layer that makes governance documentation credible.
Start free — no credit card