NIST’s AI Risk Management Framework 1.0 was published in 2023, AI 600-1 followed on 26 July 2024, and ISO/IEC 42001:2023 provides a certifiable AI management system. The three documents are often mentioned together as though they formed a complete package. They are indeed complementary, but they answer different questions: how to govern risk, which specific risks generative AI raises, and whether an organisation can manage them continuously. In 2024 companies began placing management system certification and model evaluation on the same roadmap, and that combination deserves to be taken apart.
Three documents, three questions
AI RMF 1.0 provides a general framework for risk governance, focused on how an organisation sets objectives, identifies risk, selects controls and reviews them over time. AI 600-1 adds specific risks and suggested actions for generative AI, isolating the new problems that generation brings. ISO/IEC 42001 starts from the management system and requires policies, roles, processes, internal audit and management review. The first two are methods; the third is an institutional system. Methods can be borrowed; systems have to be audited.
Certification proves process, not the model
Management system certification audits whether processes exist and operate, not whether a model is accurate or safe. That distinction is often blurred in procurement. A customer seeing a certificate may read it as an endorsement of product capability; if a company uses it that way in a tender, a later safety incident becomes harder to explain. A more defensible statement is that certification shows the organisation can identify and manage risk continuously, while the model’s specific performance is shown by test reports and evaluation records.
Model evaluations need comparable records
The difficulty with model evaluation is not running a test but making results comparable. Records should include model version, test scope, method, benchmark, environment and limitations. The risk categories in AI 600-1 can serve as a starting point for evaluation dimensions, while the AI RMF management cycle determines how results enter decisions. An evaluation without version and method cannot be compared across batches, nor can it explain the basis of a conclusion when a customer asks. Evaluation need not pursue a single score, but it should answer what result was obtained, under what conditions, for which category of risk.
Supply chains and overseas markets
Overseas customer questionnaires usually ask for both management system proof and model-level evidence: a certificate, a model description, test reports and an incident handling process. Submitting only one item invites a request for the next. For companies selling abroad, the same material also needs to correspond to technical documentation for EU high-risk systems and to US customer due diligence lists. Organising materials around one underlying set of facts and adding jurisdictional differences is more manageable than maintaining several unrelated sets.
Sequence for implementation
A sensible order is to define scope and roles, build a risk register, embed controls in research and release processes, and then conduct internal audit and certification. Model evaluation runs in parallel with system building, not after certification. Certification cycles are usually measured in months, while testing and remediation depend on product versions, so the two timetables need to be aligned at the start of the year. Treating certification as the finish line tends to end maintenance once the certificate is issued; treating it as one part of the management cycle is what gives it continuing value.
For security testing and incident response, see cybersecurity compliance; for aligning materials across jurisdictions, see multi-jurisdiction compliance; for EU high-risk system obligations, see EU AI Act services.
This article is based on NIST’s AI Risk Management Framework 1.0 (published in 2023), NIST AI 600-1 (published on 26 July 2024) and ISO/IEC 42001:2023, together with public information. It is for reference only and does not constitute legal advice.