ISO 42001 Clause 8: Put AI Controls into Operation
ISO 42001 clause 8 makes AI governance operational. It requires controlled processes, implementation of selected controls, repeated risk and impact assessment, and verification of risk treatment effectiveness. This is where an approved pilot becomes a managed service, with evidence showing that its conditions remain appropriate as technology and use change.
What clause 8 requires in practice
Clause 8 is where the plans made in clause 6 turn into daily operation. It has four parts. Clause 8.1 asks the organisation to plan, implement and control the processes it needs, set criteria for them, operate the controls chosen through risk treatment, monitor whether those controls work, keep enough information to be confident processes ran as planned, control changes and control relevant externally provided processes, products or services. Clause 8.2 repeats the AI risk assessment on a planned schedule, and also whenever significant changes are proposed or occur, and keeps every result. Clause 8.3 carries out the treatment plan, verifies it works, revisits it if it does not, and keeps results. Clause 8.4 repeats the AI system impact assessment on a planned schedule, and also when significant changes are proposed, and keeps results.
8.1: operate against defined criteria
The organisation plans, implements and controls processes needed to meet requirements and carry out planned actions. It establishes criteria, implements the operational controls determined through risk treatment and monitors their effectiveness, considering corrective action when intended results are not achieved. Documented information supports confidence that processes operated as planned.
Planned changes are controlled, the consequences of unintended changes are reviewed and adverse effects addressed as necessary. Externally provided processes, products and services relevant to the AIMS also require control. Supplier dependence therefore belongs in operational governance even when the organisation does not build the model itself.
The operational-control guide explores how to translate conditions such as “review before use” into workable responsibilities, limits and evidence without prescribing one universal lifecycle.
Why clause 8 matters to the business
This is the clause where governance either becomes real or stays on paper. Clients and regulators increasingly ask not only what an organisation intended, but whether its AI uses still operate inside the limits it set. A system that passed a careful approval and then drifted, through a new supplier behaviour or an extended use, exposes the business to complaints, rework and loss of confidence. Strong operation control gives managers early warning, keeps approvals tied to the use that was actually assessed, and produces the evidence needed for customer due diligence and for certification audits. It also stops good pilots from being switched off in a panic because nobody can show what they were allowed to do.
How to implement clause 8
Treat clause 8 as one operating loop rather than four separate tasks. One method Sterling teams use is to follow a live AI use through its change points and check that each of 8.1 to 8.4 is triggered.
- Pick the AI uses in scope and confirm each has operating criteria that a front-line employee could apply on a busy day.
- Link every risk treatment decision to a named operating control with an owner and a way to see whether it works.
- Define how proposed and unplanned changes become visible, including supplier updates, new data sources and extended purposes.
- Route significant changes into reassessment of risk and of impact before the new use goes live.
- Decide what evidence will show treatment is effective, and separate it from evidence that a task was merely completed.
- Set planned intervals for repeating assessments, and keep all results in a way that shows the history.
- Review the loop after the first real change and correct the points where it was missed or slow.
Example: a software assistant gains a new function
A hypothetical software company initially uses AI to suggest answers for its internal help desk. Staff check the suggestions before responding. A later integration allows the assistant to initiate account changes. The new function changes the consequences of a wrong output and the decisions that require control.
The original pilot approval should not be treated as approval for every subsequent function. The company needs to route the proposed change through the relevant operational and assessment processes, establish appropriate criteria and obtain necessary decisions. Evidence should reflect the new use, rather than rely solely on tests of the earlier drafting feature.
8.2: perform risk assessments at the required triggers
AI risk assessments take place at planned intervals and when significant changes are proposed or occur, with all results retained. The process uses the established planning method. A risk register that is never reconsidered after deployment cannot demonstrate this continuing application. The AI risk assessment guide explains the method and its repeated use.
Determine how significant changes become visible. Suggested triggers to examine include a different purpose, affected population, data source, supplier arrangement or connection to an operational decision. These are examples to consider, not an exhaustive ISO list. The established process should handle uncertainty about significance instead of leaving individual teams to guess without support.
8.3: implement treatment and verify whether it works
The risk treatment plan is implemented and its effectiveness verified. Newly identified risks require assessment of treatment options through clause 6.1.3. If treatment is ineffective, the organisation reviews and revalidates it and updates the plan. All treatment results are retained.
In a hypothetical manufacturer, an AI inspection system sends uncertain cases to an experienced inspector. Verification should examine whether the referral arrangement actually catches the intended cases and whether the inspector has enough information and time. Counting referrals proves activity; it does not alone demonstrate that the residual exposure is acceptable.
The treatment-effectiveness guide explains how to connect a control's intended result with suitable evidence and limitations.
8.4: perform system impact assessments
AI system impact assessments occur at planned intervals and when significant changes are proposed, with all results retained. Clause 8.2 also explicitly addresses significant changes that occur; the impact-assessment trigger here focuses on proposed significant changes. The planning process in clause 6.1.4 establishes what the impact assessment considers, including consequences for individuals, groups and society.
For the manufacturer, a proposed extension from product inspection to evaluating individual operator performance changes who may be affected and how. The impact assessment should address that proposed use before earlier conclusions are assumed to cover it. See the impact assessment guide for the combined planning and operating explanation.
Evidence and common operating failures
An auditor may select a change and follow its assessment, decisions, implementation and later results. They may compare supplier information with actual use and examine whether an ineffective treatment caused the plan to be reconsidered. Records should explain the conditions evaluated and the limits of conclusions.
Common mistakes include treating a purchase approval as complete operational approval, relying on a supplier's general performance claim for a different use, and retaining only the most favourable assessment. A global organisation should also check whether shared arrangements can be applied by local teams facing different languages, working hours or customer commitments.
Explore each numbered part
8.1 · Define the Conditions for Use
Make AI operating criteria usable where work happens. IT incident-routing and BPO document-extraction examples explain how approved purposes, supplier changes and informal workarounds affect control. Explore the evidence needed to understand actual operation, how control monitoring supports decisions, and why a purchase approval cannot cover every later use automatically.
8.2 · Repeat AI Risk Assessment When Conditions Change
Apply the established AI risk method during operation, including at planned intervals and when significant changes are proposed or occur. IT and construction examples show how changed consequences can matter even when the model is unchanged. Learn to recognise reassessment triggers, retain all results and connect findings with treatment decisions.
8.3 · Check the Risk Response
Check whether implemented treatment actually addresses the assessed AI risk. Software instructions and manufacturing inspection examples distinguish task completion from effectiveness and show why representative conditions matter. Learn how to retain meaningful results, recognise evidence limitations and revisit the treatment plan when controls fail or newly identified risks change the decision.
8.4 · Keep AI Impact Assessments Current
Keep impact assessment connected with current deployment, intended use and foreseeable misuse. Manufacturing and BPO examples show how a proposed extension changes who may be affected even when the tool stays the same. Learn to apply the defined process, retain all results and bring changed findings into risk and operating decisions.
Frequently asked questions
Is one pre-launch assessment enough?
No. The standard includes planned reassessment and change-related triggers. The operating system must keep those decisions current.
Does a supplier manage all risks of its service?
The supplier may manage parts of the technology, but your organisation still determines how it is used and controls relevant external dependencies.
Is installing a control proof of effectiveness?
No. Clause 8.3 requires verification of treatment effectiveness. Evidence should address the intended result, not just installation or task completion.
Connect AI plans with operating evidence
Discuss operational AIMS implementation with Sterling, including scoped process support and agreed Effivity ISO & GRC automation software. Explore AI Risk & Impact Assessment to develop understanding of the assessments that inform operation. Independent bodies decide management-system certification.
Previous: clause 7. Visit the hub or continue to clause 9: evaluation.