Skip to content
NEWWhich kind of AI user are you?
Take the two-minute check
AIAC AI ASSURANCE COUNCIL™

ISO/IEC 42001

Clause 8.4 — AI system impact assessment

ISO/IEC 42001 puts AI system impact assessment in two places: 6.1.4, inside Planning, and 8.4, inside Operation. The second is the one that recurs. Read against where it sits, 8.4 is the running of the process rather than the writing of it — assessments performed, results retained, and the work done again when the system or its setting moves.

§ 1 — Who it binds

Which systems, and for how long

Audited organisation

Every AI system inside the scope of the AI management system, for as long as it operates. Clause 8.4 binds the organisation running the AIMS rather than the supplier of any particular system: where a model is bought or hosted, the assessment of its impacts on individuals, groups and society is still owed by the organisation putting it to work, and a vendor’s own assessment is an input to that rather than a discharge of it.

§ 2 — In practice

Two clauses, one title, and the one that recurs

Two clauses carry the identical title AI system impact assessment: 6.1.4, within Actions to address risks and opportunities, and 8.4, within Operation. Both appear on the standard’s own contents page, so this is neither a mis-citation nor a duplication anybody needs to reconcile. It is the ordinary planning-and-operation split running through the whole standard, and its consequence is practical: an organisation can hold a documented impact assessment methodology and still be short of 8.4, or hold a stack of completed assessments and still be short of 6.1.4. Practitioners meet it first as a search problem, because a query for the phrase returns advice about both without saying which, and meet it again as an audit problem when the evidence offered belongs to the other clause. ISO/IEC 42001 explained for practitioners sets out how clauses 6 and 8 mirror each other across the management system.

An AI system impact assessment is not a data protection impact assessment and it is not a fundamental rights impact assessment. The difference is the subject rather than the depth. A DPIA reasons about the processing of personal data. A fundamental rights impact assessment, where Article 27 requires one, reasons about fundamental rights and reaches only certain Annex III systems in the hands of certain deployers. The 42001 assessment reaches individuals, groups of individuals and society, and the third of those has no counterpart in either instrument. Societal impact is where an organisation has to say something about displaced labour, degraded information environments, or the distributional effect of one decision rule applied at scale — none of which is a question about a data subject, and all of which is why the section comes back empty so often.

Where the clause sits decides how often the work happens. 8.4 is in Operation, the part of the management system that executes rather than plans, so an assessment performed once at go-live and filed with the launch approval has cleared a gate rather than met a clause. Two triggers do the work in practice: the interval the organisation sets for itself, and material change to the system or to the setting it runs in. Both have to be defined, because neither is self-evident, and a trigger nobody defined cannot be shown to have fired. The same test separates a lifecycle process from a pre-market exercise under Article 9, where what distinguishes the two is whether anything ever causes the process to run a second time.

For method there is a dedicated standard. ISO/IEC 42005, Information technology — Artificial intelligence — AI system impact assessment, exists and carries that title; its edition, publication status and internal numbering are not asserted here and no clause of it is cited. The control side is thinner still. A NIST-hosted crosswalk, attributed to Microsoft and written against the FDIS rather than the published standard, lists Annex B controls numbered B.5.2 to B.5.5 covering the impact assessment process, the documentation of assessments, and the assessment of impacts on individuals and on society. Those numbers are a map of what exists in the annex, not citable titles, and the crosswalk’s text layer carries scanning artefacts. An organisation building a method should treat them as a checklist of the ground to cover and take the wording from the standard it has bought.

The output has to go somewhere. An assessment that terminates in its own document has told the organisation something and changed nothing, and because 8.4 sits alongside operational risk treatment the useful test is whether any treatment decision can be traced back to an impact finding. Where an assessment concluded that a group is disproportionately affected and no control, no monitoring measure and no recorded acceptance followed, what exists is a description. This linkage is also the first thing a certification body follows, because it is the cheapest available way to tell an assessment that was performed from one that was produced.

§ 3 — When it applies

Editions, not deadlines

  1. 31 December 2023

    Publication date of the BSI adoption of the first edition; an edition date, not a compliance deadline.

  2. 31 March 2026

    European and UK adoption, BS EN ISO/IEC 42001:2026 — BSI states no technical differences in the requirements or guidance.

§ 4 — What discharges it

What the assessment file must produce

The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.

  1. 01

    Completed assessments, one per system and configuration

    Each carrying the version of the system it examined, the date, and a named approver. An assessment tied to a product name rather than to a configuration cannot tell a reader which of several retrained models it was actually about.

  2. 02

    The reassessment trigger definition, approved and dated

    Which events oblige the work to be redone, written before they happen and specific enough to argue about — retraining, threshold change, a new user population, a new jurisdiction. An undefined trigger can never be shown to have fired.

  3. 03

    The societal-impact section, actually completed

    What the system does to people who are neither its users nor its subjects, and to the market or public good it operates inside. This is the section that separates the clause from a privacy exercise and the section most often left blank.

  4. 04

    A reassessment log against the interval the organisation set

    Dates the work was due, dates it was done, and the reasons for any gap. A log recording only completions cannot demonstrate that the cadence the organisation chose for itself was honoured.

  5. 05

    The linkage record from findings to treatment decisions

    Each material finding followed to the control, monitoring measure or acceptance record it produced, with the owner. Where a finding leads nowhere, the record should say that it was considered and closed, and on whose authority.

  6. 06

    The approval of record for each assessment result

    A named individual with the authority to accept the conclusion, and the date they did. Approval by a working group or a distribution list leaves nobody who can be asked afterwards why the result was accepted.

§ 5 — Worked example

Worked example — one assessment, fourteen months ago

A national employment agency deploys a model that ranks jobseekers for referral to caseworkers. An impact assessment was completed before launch in March: accuracy by age band and gender, the data protection position, and the appeals route. The head of digital approved it and it was filed. Fourteen months on, the model has been retrained twice, the referral thresholds have been tuned three times, and the tool has been extended to a second benefit stream.

The agency says it has an AI system impact assessment. What does clause 8.4 still want?

Evidence that the exercise happened more than once. The March document is the artefact of a launch gate; two retrainings, three threshold changes and an extension to a new population are precisely the events an operational clause exists to catch, and not one of them appears anywhere in the file. Three things are missing and each is small. A stated trigger definition saying which changes force the work to be redone — a retraining does, a hotfix probably does not, an extension to a new benefit stream certainly does. A log showing the interval the agency set for itself was met, or recording why it was not. And the societal section, which the original never attempted: a ranking model inside a public employment service allocates a scarce good, and the question of which cohorts see fewer referrals in aggregate is not answered by parity across two demographic fields. The remedy is not a longer document. It is a shorter one that runs again.

§ 6 — What a weak answer looks like

A DPIA wearing a new title

The DPIA with a fresh cover sheet. A data protection impact assessment is reused, an "AI" heading is added, and the organisation reports impact assessment as covered. It is usually a competent document answering a narrower question — lawful basis, necessity, proportionality, the rights of a data subject — and it stops at the boundary of the individual. Nothing in it reaches groups as groups and nothing reaches society. The reuse also tends to be one-off, because the instrument it was borrowed from is a pre-processing exercise, so the recurrence this clause is built around never starts at all.

§ 7 — Elsewhere

The other impact assessments, and what each asks

Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.

  • EU AI Act

    Article 27 obliges certain deployers of certain Annex III systems to assess the impact on fundamental rights before use. Both instruments ask about effects on people; each fixes its own population, its own trigger and its own subject matter.

  • EU AI Act

    Article 9 requires a continuous, iterative risk management process across a high-risk system’s lifecycle. It meets clause 8.4 at the point where a recurring assessment has to feed a decision about treatment rather than sit in a file.

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.

§ 8 — provenance

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.

§ 9 — Also read

Also read

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.

Flagship · AIAC-CAAACertified Applied AI Associate30 minutes · 18 or over · on the register
Open now →The credentials open for enrolmentThe register →