Certify your AI with ISO 42001

Building an artificial intelligence management system (AIMS).

AI frameworks are multiplying.

The latest to drop is ISO/IEC 42001. Published by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), it is a high-level framework for safe, secure, and effective use of artificial intelligence. It serves as the blueprint for building an “Artificial Intelligence Management System” (AIMS).

In general, it reads similarly to (and explicitly mentions) the National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF), although it is a bit more prescriptive. And as a result, ISO 42001 can get you something the NIST AI RMF cannot:

certification.

At least one company, AI Clearing, has already achieved this (Update, 25 March 2024: after speaking with an expert auditor, it looks like this certification was done against the Final Draft International Standard (FDIS) and is likely not formally accredited). Although I’ve expressed some skepticism about external audits in general (as they currently happen), I have no doubt ISO 42001 will take off in popularity. That is because people have a very strong and natural desire to get a “stamp of approval” for what they are doing. And these stamps are important for:

• Winning sales

• Pacifying regulators

• Deflecting bespoke questionnaires

So in this post, I’ll do a quick overview of the new standard.

Look, feel, and formatting

Something that always struck me was how not machine-readable the ISO 27001 and SOC 2 standards are. Unlike NIST ones where every line is numbered and can be directly referred to, in ISO standards it is extremely difficult to describe individual requirements in a structured way. They are full of sub-paragraphs, internal references, notes, and footnotes which make it difficult to analyze systematically or communicate with any more precision than “the second full sentence in section 6.1.3.”

Formatting aside, sections 4-10 of the document are the most substantive, and like ISO 27001, the 42001 standard has an Annex A with a list of controls. Organizations are required to develop a Statement of Applicability in which they justify the inclusion or exclusion of Annex A controls.

Content

4. Context

The first major section of the standard lays out how an organization should evaluate itself in terms of its AI use. Specifically, it requires analysis of external and internal issues like:

• Contractual obligations

• Competitive dynamics

• Regulatory decisions

• Legal requirements

• Incentives

• Culture

• Ethics

It also lays out the roles the organization plays when it comes to AI. It’s certainly possible - even likely - that your company would be several of these at once:

• Provider, developer, tester, and evaluator

• Data subject

• Regulator

• Customer

• Partner

• User

Section 4 wraps up by requiring establishment of the:

• Business requirements of the AI system

• Scope and boundaries

• Stakeholders

5. Leadership

This section requires establishment of a formal, documented AI policy. Without any overarching written guidance, it would be difficult to have a comprehensive governance program, so this makes sense. If you don’t have a policy yet, feel free to grab StackAware’s free template.

This section also lays out how an organizational management team should enforce the AI policy and pursue AI-related objectives (described in the next section), ensuring both of them:

• Align with the strategic direction of the organization

• Are properly resourced and communicated

• Integrate with existing business processes

• Undergo continuous improvement

• Are ultimately achieved

6. Planning

Most importantly, this section requires the establishment of AI objectives that are clearly documented, measurable, and aligned with the organizational AI policy. With those in place, an organization can effectively:

• Document business and security requirements

• Establish risk appetite and tolerance

• Conduct risk assessments

And once assessments are complete, the standard requires establishment and documentation of “risk treatment options,” i.e. the Four Horsemen of:

• Accept

• Avoid

• Mitigate

• Transfer

7. Support

This is primarily about internal implementation. A key requirement here is that the AIMS documentation itself must be protected from corruption or inadvertent disclosure. Especially if you are writing policy-as-code (which you probably should!), it will be exceptionally important to log access and prevent tampering.

8. Operation

Whereas section 6 is about planning, this is about execution. It lays out a series of guidelines for implementing the organization’s AI risk management plan (and documenting the results).

9. Performance evaluation

This section establishes auditing requirements and a management review process. Especially with non-deterministic systems, preventing repudiation (of both human and AI actions) and immutability of the audit logs will be key.

Similarly, having an engaged management team that reviews the AIMS on a regular basis is another requirement documented here.

10. Improvement

The final section requires continuous improvement and dealing with “nonconformities.” This mainly means having a standardized process for root cause analysis after every incidence of non-compliance with your AI policy. If you need a template for a governance procedure that includes this, check out the StackAware AI risk management standard operating procedure (SOP).

Certified artificial intelligence: the AIMS accelerator

The European Union AI Act is likely to be formalized this year, and will come into force gradually from now then until 2026. The Biden Administration is pushing (unwisely in my mind, but doing so nonetheless) to assign liability to software providers for security and other shortcomings unless they can achieve “safe harbor” by demonstrating adherence to certain principles.

These regulatory developments are likely to apply intense pressure to organizations to prove the soundness of their AI management programs. And where there is a need, StackAware will be there to help.

That’s why we are looking to soon launch the AIMS accelerator, a software platform to streamline your ISO 42001 certification process.

Interested?

---

This article is not a replacement for the actual ISO 42001 standard. I encourage you to buy access to the full document for complete information.

---

Related LinkedIn posts:

• ISO 42001 “ready”?