News · 18 January 2024

ISO/IEC 42001: treating AI as a managed system, not a one-off project

In December 2023 ISO and IEC published ISO/IEC 42001, the first international standard for an AI management system. It did not make many headlines, but it is a useful marker: AI has moved from "something a data science team builds" to "something an organisation operates and is accountable for".

What the standard actually asks for

ISO/IEC 42001 follows the same high-level structure as ISO/IEC 27001 and ISO 9001, so organisations that already run a management system will recognise it. It asks you to define the scope and context of your AI use, assign roles, assess risks and impacts, put controls in place, monitor them, and improve. The Annex lists controls on topics such as data quality and provenance, documentation, human oversight and third-party suppliers.

The most interesting word is "impact". A conventional security risk assessment asks what could go wrong for the organisation. An AI impact assessment also asks what could go wrong for the people affected by a model's output: a patient, a loan applicant, a citizen.

Why this matters even if you never certify

Certification is optional, and for a small team it may not be the right first step. The value is in the questions the standard forces onto the table:

  • Where does the training data come from, and are we allowed to use it?
  • Who decides when a model is good enough to deploy, and on what evidence?
  • How will we notice when performance drifts after release?
  • Who is the human a decision-maker can ask when a result looks wrong?

Most failed AI projects we see do not fail on algorithms. They fail because no one owned these questions.

A practical starting point

You do not need a new bureaucracy. Start with a one-page register of every model in use or in development, with an owner, a purpose, the data it relies on, and the date of its last validation. Add an explicit "what this model must not be used for" line. That alone puts you ahead of most organisations, and it maps cleanly onto the standard if you later decide to pursue it.

Explainability fits naturally here. If a model cannot show what it relied on, then "human oversight" and "monitoring" become hard to carry out in practice. A management system makes that requirement visible early, which is cheaper than discovering it during an audit.

Written in January 2024; details reflect the situation at that time.

All news