ISO/IEC 42001 is a management system standard for organizations that build or use AI. Its text does not prescribe SBOMs, but the models, datasets, libraries and services behind your AI systems are a supply chain, and a management system needs to know what is in it. Map your inventory and supplier reviews to the standard's themes.
This post explains the standard in general terms. We could not open ISO's own pages, so descriptions rely on secondary sources and the official listing from the Estonian standards body. Read the standard itself before you map clauses, and note that nothing here says any tool certifies you.
What is ISO 42001?
Published in December 2023 by ISO and IEC, it sets requirements for establishing, implementing, maintaining and continually improving an AI management system within an organization. It applies to organizations that provide or use products or services built on AI systems, regardless of size or sector. It is voluntary, and per secondary sources it is a standard you can be certified against, unlike several related AI documents such as the vocabulary standard ISO/IEC 22989, the risk guidance in ISO/IEC 23894 and the impact assessment guidance in ISO/IEC 42005.
Secondary sources also say it follows the high level structure shared by other management system standards, which is why it fits alongside ISO 27001. If you already run an information security management system, much of the machinery, such as scope, risk, internal audit and management review, is familiar.
What counts as supply chain for an AI system?
Walk through one AI feature and list what it depends on.
- Models. Third party foundation models, fine tuned variants, and open weights you downloaded.
- Data. Training, evaluation and retrieval sources, with their licenses and origin.
- Libraries and frameworks. The ordinary software packages around the model.
- Services. Hosted inference, vector stores and tool APIs.
- Tooling. Notebooks, pipelines and registries that touch all of the above.
NIST's AI Risk Management Framework takes a similar view of lifecycle thinking and gives four functions, Govern, Map, Measure and Manage, so teams often use it as a companion. NIST describes it as voluntary.
How to map
Rather than citing clause numbers we could not confirm, map by theme.
Inventory. A register of AI systems and what they use. A software bill of materials covers the libraries, and your own register covers models and datasets.
Supplier and third party management. Due diligence, contract terms and review for model and data providers, plus how you track changes to hosted models you do not control.
Risk assessment. Where a compromised library, a poisoned dataset or a swapped model would change behavior, and what you do about it.
Operation and change control. How a new model version or dependency is approved, tested and rolled back.
Monitoring and improvement. Tracking advisories for the components you use, and recording what you fixed.
Evidence. Dated records an auditor can sample.
Where tools help and where they stop
An SBOM and dependency scan give you the library layer of the inventory, and Safeguard produces both and scans AI related packages like any other dependency. Models, data sources and policy decisions belong in your own register. We do not claim Safeguard maps to ISO 42001 clauses. Check the framework library to see which frameworks are covered today, and note that its documentation says each framework page lists what a scan decides and what needs your evidence.
What to do
- List your AI systems and name an owner for each.
- For each, list models, data, libraries and services.
- Generate SBOMs for the software around the model and store them with releases.
- Add model and data providers to supplier review.
- Define how new model versions are approved and rolled back.
- Decide how advisories for AI libraries reach the owner.
- Ask your certification body, if you plan to certify, how they scope AI supply chains.
Quick check
Is ISO 42001 mandatory?
No. Secondary sources describe it as voluntary. Customers or contracts may still ask for it.
Does an SBOM cover models and datasets?
Not on its own. It lists software components. Keep a separate register for models and datasets, or use a format built for that.
Sources
- EVS: ISO/IEC 42001:2023 listing, found by search 18 October 2026
- LRQA: ISO 42001 overview guide, found by search 18 October 2026
- NIST: AI Risk Management Framework, fetched 18 October 2026
- Safeguard documentation: Compliance, fetched 18 October 2026
Last reviewed: 18 October 2026.
