Talk to a specialist

ISO 42001 · Practical implementation guide

ISO 42001 AI risk treatment and control selection

AI risk treatment turns an assessment into changes that reduce exposure, improve detection or limit the consequences of failure. A useful treatment plan explains what will change, why it should work, who will deliver it and what evidence will show the result. Selecting a control name is only the start; the release decision depends on whether the chosen safeguards work in the actual service.

In this guide

Begin with the scenario and the decision

Take one assessed scenario and restate its causal path. Identify where intervention is possible: purpose, inputs, permissions, model behaviour, output handling or downstream action. This keeps treatment focused on the mechanism of failure. A general policy or training action may help, but it should not be used to conceal a technical boundary that remains open.

Clause 6.1.3 requires a treatment process, selection of necessary controls, comparison with Annex A, consideration of Annex B guidance and documented applicability decisions. It also addresses the treatment plan and designated management approval of residual risk. Clause 8.3 covers implementation. These are requirements references; the design and testing approach below is recommended practice. Reference: ISO/IEC 42001:2023.

Decide whether the proposed use itself should change. A feature may be acceptable as a draft generator but not as an autonomous decision-maker. A service may work for a restricted data set or audience while broader use needs more evidence. Narrowing the use can be a deliberate treatment that preserves value and makes the remaining controls feasible.

Make tradeoffs explicit. A safeguard can reduce one risk while increasing another. Requiring human review may reduce incorrect action but create delay and workload. Adding logs can improve investigation while increasing sensitive data retention. Consider the combined effect and give owners enough information to choose a balanced design.

Compare treatment options before committing

For each material risk, identify plausible options rather than jumping to the first familiar control. Options can include avoiding the use, reducing exposure, changing architecture, improving data, adding review, strengthening detection or accepting a defined remainder. External agreements may allocate responsibilities, but they do not necessarily prevent the event or remove its consequences for users.

Assess effectiveness, feasibility, dependency and operational burden. A technically strong safeguard may fail if the team cannot maintain it. A supplier commitment may be useful but depend on information the supplier cannot provide quickly. A manual check may be suitable at low volume and become unreliable when demand grows. State these assumptions in the plan.

Prefer controls close to the failure path. If the risk is cross-customer retrieval, an authorization check before retrieval directly addresses the boundary. A reminder to staff not to share customer information is further from the mechanism and should not be the only treatment. Layering can be useful, but explain what each layer adds rather than counting controls as if more always means better.

Consider failure of the control itself. What happens if a reviewer is absent, a filter is unavailable or a supplier changes an interface? Define safe fallback behaviour where practical. A control that silently fails open may provide less assurance than the design document suggests. Test the degraded condition as well as normal operation.

Select controls from the risk context

Use Annex A as a structured completeness check, not as the sole source of possible safeguards. Necessary controls may come from the organization’s architecture, sector practice, legal obligations or other frameworks. The standard’s treatment process requires attention to controls beyond the reference set when needed. Keep the reason for each choice linked to the risk and the scoped activity.

Read applicable Annex B guidance when designing the selected controls. Its role is part of the treatment process; it should not be ignored because a spreadsheet has already marked the Annex A row as included. At the same time, distinguish the standard’s requirements from your chosen implementation details. A tool, form or review frequency is not automatically mandated merely because it appears in your procedure.

Use the Statement of Applicability to explain inclusion and exclusion. A control may be relevant to one deployment and not another, so avoid a vague organization-wide answer that hides important differences. Where shared controls support several use cases, link the common implementation and the deployment-specific conditions. This reduces duplication without erasing context.

Do not equate outsourcing with exclusion. If a supplier develops the model, your organization may still need evidence about the service, local evaluation, restrictions and incident coordination. Document who performs each activity and how you gain confidence in it. An exclusion justified only by vendor responsibility often leaves the interface unmanaged.

Build a plan that can be delivered

Give each action a clear owner, deliverable and completion test. Replace improve AI governance with a specific result, such as enforcing user-level access before retrieval and testing the boundary with defined cases. Identify dependencies and the resources required. A due date without capacity or authority can create an overdue item that nobody can resolve.

Separate design, implementation and effectiveness evidence. A design explains the intended mechanism. Implementation evidence shows that it was deployed. Effectiveness evidence shows what happened when it was tested or operated. All three can matter, and none should be confused with a manager marking the task complete.

Record interim conditions when treatment cannot be completed before use. These may include a limited audience, reduced permissions, extra review or suspension of a particular feature. State who approved the conditions and when they expire. Avoid treating a future remediation plan as if it already reduces current risk.

Plan for ongoing ownership. Controls need maintenance when data, models, users or suppliers change. Include the team responsible after the project closes and the signals that trigger review. A one-time implementation can become ineffective if nobody owns updates or monitors whether its assumptions still hold.

Verify effectiveness with decision-focused tests

Design a test that could show the control is inadequate. If the treatment is an access boundary, include attempts that should be denied. If it is human review, examine whether reviewers catch the relevant errors under realistic workload. If it is a supplier notification route, test whether a message reaches the correct owner and produces the expected assessment.

Use realistic conditions without exposing unnecessary sensitive data. Synthetic cases can test some boundaries, while carefully governed representative data may be needed for performance questions. Explain the limits of the test. Passing a narrow evaluation should not be presented as proof of effectiveness in every language, environment or use pattern.

Check for unintended effects. A filter may block legitimate work, a warning may be ignored after repeated exposure, or a manual review step may create an unmanageable queue. These effects can undermine the treatment or introduce new risks. Include operational feedback before expanding the control across the organization.

Compare the observed result with the intended risk reduction. Update residual risk based on the evidence rather than automatically lowering the score because a task is complete. If the result is uncertain, document the uncertainty and the next step. Management should see the actual position when deciding whether use can continue.

Example: a software assistant with powerful connectors

Imagine a software company building an assistant that can read project information and create deployment requests. The risk assessment identifies a path in which hostile text inside a project document causes an unintended tool call. The initial treatment proposal is to add stronger instructions to the assistant. Security review finds that the application still holds broad permissions.

The team compares options. It can remove the connector, restrict the service account, require a separate approval for action or redesign the tool interface so that only narrow, validated requests are possible. It chooses limited permissions and an approval step for changes that affect production, while retaining read-only assistance for routine questions. The system’s allowed behaviour is enforced outside the model as well as described in instructions.

Tests include hostile documents, unauthorized users and malformed tool requests. The record shows which attempts were blocked, which reached the approval screen and which conditions remain untested. Operations confirms that the fallback can disable action without making the entire knowledge service unavailable. The treatment plan distinguishes implemented safeguards from a later improvement to input handling.

Management approves the restricted release with a review date and monitoring conditions. The residual risk is accepted for that configuration, not for every future connector. Adding a new action later triggers a change assessment and new evidence. This preserves the link between the control design and the business decision.

Example: AI-supported document processing in a BPO

Consider a BPO team using AI to extract fields from customer documents. The service feeds an operational system, and incorrect values can cause rework or delayed service. The first proposed treatment is to have staff review every output. A workload analysis shows that this would consume most of the expected efficiency gain and may lead to rushed approval.

The team separates fields by consequence and checks how errors arise. Some fields can be validated against a known format or a second source. Others require judgment and have more serious downstream effects. The treatment combines deterministic validation, targeted human review and a rule that uncertain cases remain out of the automatic path. The operating procedure makes the exception route visible.

Evaluation uses document types and quality levels representative of the service. It measures both missed errors and unnecessary referrals, because too many referrals can overwhelm reviewers. Supervisors test the process during a busy period and identify an escalation bottleneck. Management funds additional review capacity for the initial deployment rather than assuming staff will absorb the work.

The plan records ownership of validation rules, review instructions and monitoring. A change in document format becomes a trigger to reassess the treatment. The organization can show why the controls were chosen, how they were tested and what conditions support continued use.

Avoid treatment plans that only change paperwork

Look closely at repeated actions such as update policy, provide training or remind staff. These can be appropriate when unclear instructions or competence are causal factors. They are weak responses when the system permits an action that should be technically impossible or when workload makes the expected review unrealistic. Match the response to the cause.

Challenge unsupported risk reductions. Ask what evidence justifies a lower residual rating and whether the control has operated long enough to reveal relevant failures. A supplier certificate, signed policy or completed course may support part of the story but may not establish the service-specific outcome. Keep the limitation visible.

Escalate stalled treatments with options. Management needs to know whether to add resources, change suppliers, narrow scope or accept a permitted remainder. Reporting only that an action is overdue creates little basis for a decision. Explain the consequence of delay and the current operating conditions.

Before seeking external support, assemble a few material scenarios, the options considered and the current evidence. A useful implementation engagement should improve the control design and the ability to demonstrate its operation. The number of templates delivered is a poor substitute for that result.

Keep the treatment decision attached to the service record so that later owners can understand its limits and review conditions.

Frequently asked questions

Must every Annex A control be implemented identically?

No single implementation fits every use. Apply the standard’s selection process, justify the decisions and consider additional necessary controls. Deployment context and organizational roles matter; document differences rather than hiding them in broad answers.

Can we accept a risk instead of treating it?

Acceptance must follow the organization’s criteria, authority and applicable obligations. Describe the remaining risk, evidence and conditions clearly. Acceptance cannot make a prohibited activity permissible or remove an obligation that the organization must meet.

What makes a treatment action complete?

Completion should include the agreed deliverable and evidence that it works as intended. A closed ticket can show administration, but the risk decision needs implementation and effectiveness evidence appropriate to the scenario.

Is a supplier contract a sufficient control?

It may allocate obligations and support recourse, but assess whether it changes the event’s likelihood or consequences. Local restrictions, evaluation, monitoring and response arrangements may still be needed to manage the service.