An interviewer says - the executive team has committed to an ISO/IEC 42001 certification audit in nine months and you own the AI platform. The standard was published in December 2023 and certifies an artificial-intelligence management system rather than any individual model, through a two-stage audit with annual surveillance over a three-year cycle. Where do you take this?

What the interviewer is testing

Whether you can turn a management-system standard into platform features instead of a documentation project. The failure they are listening for is a candidate who starts writing policies. An auditor samples records, so the question is which records your platform emits without anyone being asked.

The clarifying questions that change the answer

• What is in the certification scope? Which legal entities, which systems, which uses. Scope sets both the cost and what the certificate tells a customer, and it is the only decision in the programme that is hard to reverse later.
• Does an inventory exist, and does it match what is serving? Every answer after this depends on it.
• Is this for a customer requirement or a board commitment? A customer requirement usually specifies the scope you need, and it is almost always smaller than "everything".
• What is already certified? If an information-security management system exists, the risk process, internal audit and management review already have owners, and this becomes an extension rather than a first build.

A strong answer's arc

1. Make the inventory a system of record, not a list. Mandatory fields at registration, lifecycle states from proposed to retired, and the registry written by the publishing pipeline rather than by a human after the fact.
2. Reconcile it against reality weekly. Every model-serving process registers at startup; the difference between the observed set and the registry is the report. Any model serving traffic and absent from the registry for more than 24 hours pages the platform team. That reconciliation is the evidence that the inventory is complete, and completeness is the first thing an auditor tests.
3. Emit the records the audit will sample. Per deployment: model version and artefact digest, the evaluation run it was approved against, the approver, the risk tier, the intended-use statement version. One record per deployment, produced by the deployment.
4. Reuse the tiering you already have as the control-selection input, because the standard asks for a risk-based system rather than a specific control set. A second parallel classification is pure cost.
5. Name what I will not build: a new committee, a bespoke documentation portal, and any claim that a model is certified.

What it costs

Nine months of scope freeze on the platform's control surface. Re-platforming the service that carries a control mid-period costs evidence continuity, and continuity is what the surveillance audits test. The annual surveillance cycle is the real design constraint: evidence has to be reproducible every year by people who were not there, not assembled once by the team that built it.

Common weak answers

• "Write the policies and map the controls." Produces a binder, and the first surveillance audit finds the controls were not operating.
• "Certify the models." The standard does not do that, and offering it to a customer as if it did is a misrepresentation they may later rely on.
• "Put every AI use in scope to be safe." The same mistake as over-scoping a security certification: the cost scales with scope and the certificate says no more.

What a strong answer adds

The scope statement is the part customers read and the part teams forget, so write it as though it will be quoted in a sales deck. State the honest limit out loud: a certificate says the organisation manages its AI deliberately. It does not say a model is accurate, fair or safe for your use case - that claim can only come from an evaluation of that use.