ISO/IEC 42001
Clause 9.1 — Monitoring, measurement, analysis and evaluation
Clause 9.1 is flat — no sub-clauses — and it asks four things before it asks for a number: what needs monitoring and measuring, by which methods, when that happens, and when the results are analysed and evaluated. Its subject is the performance and effectiveness of the management system, which is why a pack of model metrics rarely answers it.
§ 1Who it binds
What this reaches
Audited organisation
The whole of the management system, at intervals the organisation determines for itself. Clause 9.1 binds the organisation operating the AIMS and is flat: it carries no sub-clauses, so all four determinations sit in one place. It reaches AI systems the organisation neither built nor hosts, where the question becomes what can be measured directly and what has to be obtained from a supplier.
§ 2In practice
Valid results, and whose performance
The requirement doing the work here is the one about methods. The organisation has to determine methods for monitoring, measurement, analysis and evaluation such that what comes out of them is valid — and validity, on a probabilistic system, is not a property a figure acquires by being computed correctly. A benchmark score produced under development conditions is valid evidence about development conditions. The NIST AI Risk Management Framework states the same requirement in a form that can be tested, its MEASURE 2.3 subcategory asking for performance or assurance criteria to be demonstrated for conditions similar to the deployment setting; the distance between those two sentences is where most AIMS measurement sits, and testing a probabilistic control sets out what closes it.
The second question is what needs measuring, and the answer the clause is looking for concerns the management system rather than the models inside it. Organisations answer with model metrics. Accuracy, latency, drift and incident counts are properties of AI systems; whether the AIMS is working is a separate question — are impact assessments happening at the interval set for them, are approved treatments being implemented, is a competence gap closing, are nonconformities recurring. Neither set substitutes for the other, and an assessor handed a pack of model dashboards asks what any of it establishes about the system managing them. That is also the line against internal audit: this clause measures how the AIMS performs, 9.2 examines whether it conforms.
Timing is two determinations rather than one, and they get collapsed. When measurement happens is a question about instrumentation; when results are analysed and evaluated is a question about somebody sitting down with them, so continuous telemetry with no scheduled evaluation satisfies the first and not the second. The evaluation is also what the rest of clause 9 consumes: its output goes to management review, and its findings shape the next round of operational risk assessment. An arrangement whose output reaches no decision is instrumentation rather than performance evaluation, which is the practical reading of the clause’s final limb.
§ 3What discharges it
What the measurement file must show
The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.
01
A measurement plan covering the AIMS, not only the models
What is monitored, by which method, how often, and who evaluates it — including measures whose subject is the management system itself: interval adherence, treatment implementation, recurrence of nonconformities.
02
A validity argument attached to each method
Why the method yields results holding for the deployment population rather than for a test set: the conditions a figure was produced under, and the respects in which live operation departs from them.
03
Analysis and evaluation records, dated and attributed
Not the telemetry. The record of a named person examining it, what they concluded and when — the artefact separating a dashboard that exists from performance that was actually evaluated.
04
The route from each evaluation to a decision
Where the output went and what it changed: an entry in the management review pack, a treatment reopened, a threshold moved. Evaluation with no destination is the commonest defect against this clause.
§ 4Worked example
Worked example — an hourly dashboard and no evaluation
A logistics company’s AIMS covers a routing model and a demand forecast. Its weekly monitoring pack carries model accuracy against a held-out set, endpoint latency, a drift score, and a count of dispatcher overrides. The AI governance lead reads the pack and files it. Across the year two impact assessments fell due and one was done; four risk treatments were approved and two implemented. None of that appears anywhere in the pack.
The company instruments more than most. What does clause 9.1 still find?
That nothing in the pack measures the management system. Every figure is a property of a model, and three of them are produced against a held-out set rather than against live traffic, so they carry no argument about deployment conditions either. The two facts that would say whether the AIMS is working — the impact assessment that never happened and the two treatments that were never implemented — are known to somebody, and are neither measured, nor evaluated, nor on their way to anyone able to act on them. Filing is also not evaluation: the pack has a reader but no dated conclusion, which leaves the third and fourth determinations unanswered. Four lines added to the existing pack would move the company from good instrumentation to performance evaluation without adding a system: interval adherence for impact assessments, treatment implementation rate, open nonconformities by age, and one measure of whether the people the AIMS governs can find what it asks of them.
§ 5What a weak answer looks like
Model metrics answering a management-system question
A model performance dashboard offered as evaluation of the management system. It is real, it refreshes hourly, and everybody looks at it — and every metric on it is a property of a model. Nothing measures whether the AIMS is doing its job, no method carries an argument that its figures hold outside the conditions they were produced under, and no record shows anyone reaching a conclusion from them on a date. It is a strong answer about the models and no answer at all to this clause, which is why organisations with excellent instrumentation fail it more often than organisations with poor instrumentation.
§ 6Elsewhere
Monitoring duties in Union law
Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.
Article 72 requires a provider to run a post-market monitoring system collecting and analysing data on how a high-risk system performs in service, which is the same measurement problem attached to a statutory duty.
Article 15 requires appropriate accuracy, robustness and cybersecurity, with the accuracy metrics declared. A declared figure is a claim about conditions, and the validity question is whether those conditions still hold.
A correspondence indicates that two instruments address the same underlying obligation. It is not a mapping endorsed by either body, not a statement that one satisfies the other, and not a judgement about which is more demanding.
§ 7When it applies
Which edition applies
31 December 2023
Publication of BS ISO/IEC 42001:2023, BSI’s adoption of the first edition. The date identifies an edition; ISO/IEC 42001 is voluntary and no measurement duty commences on anyone here.
31 March 2026
Publication of BS EN ISO/IEC 42001:2026, the European and UK adoption, which BSI records as carrying no technical differences from the 2023 adoption. Again an edition date rather than a deadline.
§ 8provenance
The provision itself
This page sets out what the instrument requires and what discharges it. The official text is the authority — these go straight to it.
- Official contents page of ISO/IEC 42001:2023 — clause 9.1 Monitoring, measurement, analysis and evaluation, listed with no sub-clauses.
- BSI Knowledge record for BS ISO/IEC 42001:2023 — published 31 December 2023, status Current, identical to ISO/IEC 42001.
- BSI Knowledge record for BS EN ISO/IEC 42001:2026 — published 31 March 2026; BSI states no technical differences from the 2023 adoption.
- The NIST-hosted crosswalk, attributed to Microsoft, maps AI RMF subcategories onto clause 9.1 and corroborates that 9.1 carries no sub-clauses while 9.2 and 9.3 do.
§ 9Also read
Also read
Testing controls on a probabilistic AI system
NIST AI RMF: a practical walkthrough
Certification
Assessed on the same standard of evidence
Every Council credential is examined on applied judgement against a published anchor, set out the way the obligations on this page are. The Certified Applied AI Associate, the Council’s flagship, is open to any adult today, its fee waived by a coupon code from the AI-Readiness Check, and the register lists what is open for enrolment.