ISO Compliance Insights & Best Practices
CART
ISO Compliance Insights & Best Practices
AI impact assessment mitigation is the part of the assessment that changes what the system does to people. Identifying impacts is only useful if something is then done about them, and a list of harms with vague promises to “monitor” or “review” leaves the affected people exactly where they started. Good mitigation is specific, owned, dated and evidenced, and it is proportionate to the severity of the impact.
This guide explains how to plan AI impact assessment mitigation: the types of measure available, how to match them to impacts, how to assign owners and evidence, how to judge residual impact and how to keep mitigations working. It is general guidance that follows the spirit of ISO/IEC 42005 and should be adapted to your systems.
An impact assessment that stops at description is a report, not a control. The value comes from reducing harm before and after deployment. Regulators, auditors and affected people will judge the assessment by what changed, not by how well the impacts were listed.
ISO/IEC 42005, available at ISO/IEC 42005 on AI system impact assessment, describes impact assessment as including the identification of measures to address impacts and the consideration of residual impact. Treat AI impact assessment mitigation as its own workstream, with a plan, owners and evidence, rather than a final paragraph in the report.
List each impact with its severity and likelihood, using the scale in AI impact assessment severity rating. Prioritise the most serious and the most likely, and the irreversible ones above all. For each, ask what would prevent it, what would detect it early and what would repair the harm if it occurred.
Think in layers: prevent, detect, respond and repair. A single measure rarely covers all four. For example, testing may prevent bias, monitoring may detect drift, human review may catch individual errors and an appeal route may repair harm to a person.
The most effective mitigation often changes the system’s design or scope. Narrow the purpose, exclude high-risk uses, reduce the data collected, remove problematic features, add a simpler and more transparent model or limit the population it applies to. These choices reduce both likelihood and severity, because there is less that can go wrong.
Decide early. Design changes are far cheaper before launch, so run impact assessments at the design stage, not the day before go-live. Record scope limits clearly and enforce them technically and in policy, since a system built for one purpose tends to drift into others.
| Type | Examples | Impacts addressed |
|---|---|---|
| Design and scope | Narrow the use case, exclude sensitive uses, reduce data | Harm from overreach and misuse |
| Testing and validation | Fairness testing, robustness tests, local validation | Bias, errors, failure in new settings |
| Human oversight | Review of adverse decisions, escalation, override powers | Wrong or unfair individual outcomes |
| Transparency | Notices, explanations, documentation | Loss of autonomy and inability to challenge |
| Redress and monitoring | Appeal routes, correction, monitoring, incident response | Continuing harm and unnoticed drift |
Testing shows whether the system behaves as intended for the people it will affect. Include accuracy testing across relevant groups, robustness testing against unusual inputs, adversarial testing for misuse, security testing and, for generative systems, testing of harmful outputs. Validate locally on your own population and data.
Document the tests, the results, the thresholds and the remediation. Use approaches described in AI bias testing and AI model drift. Re-test after significant changes. A mitigation that says “we tested for bias” without saying how, on what and with what result is not evidence.
Human oversight can prevent or correct individual harms: review of adverse decisions, escalation of uncertain cases, authority to override and the ability to stop the system. It must be designed to work, with trained reviewers, enough time and useful information; see human oversight of AI.
Process measures include approval gates, change control, training, clear roles, complaint handling and incident response. Connect them to existing governance so they are not forgotten. For example, link changes to the model to a mandatory review of the assessment, as described in AI impact assessment monitoring.
People affected by AI systems need to know that they are, to understand the main factors in decisions that affect them and to have a way to challenge and correct them. Provide plain-language notices, explanations and accessible appeal routes, and make sure appeals reach a person with authority to change the outcome.
Redress also means putting things right: correcting errors, reversing wrongful decisions, compensating where appropriate and fixing the cause. Plan the process before harm occurs, including who decides, how quickly and how people are informed.
Each measure needs a named owner, a completion date and evidence of effectiveness. Evidence might be a test report, a signed procedure, a training record, a configuration screenshot or an appeal log. Without evidence, a measure is an intention and should not reduce the residual rating.
Track measures in a register with status and review dates, and report overdue items to the governance forum. Link the register to your AI risk treatment plan so risks and impacts are treated together, and to the AI risk register.
Free AI impact assessment (ISO 42005)
Who could this AI system affect, and how?
Screen the system against sensitive and prohibited uses, describe it, check the safeguards for fairness, transparency and oversight, and rate its impacts on people and society from 26 scenarios with ISO 42001 Annex A measures. Free, with findings.
Start the free AI impact assessment → or View premium report sample
After mitigation, re-rate each impact to show what remains. Decide whether the residual impact is acceptable, with a named approver who has authority. If it is not, options include further mitigation, narrower scope, delayed launch, or not proceeding. If you accept a significant residual impact, record why and set a review date.
Consult the people affected or their representatives where feasible, especially for serious residual impacts. Their view may reveal that a mitigation which looks adequate on paper does not work in practice.
Frequent errors include vague measures such as “monitor bias”, no owners or dates, crediting planned measures as if they were in place, relying on a single control, using human review as a catch-all without designing it, ignoring redress, failing to test effectiveness and never revisiting measures after changes. Another is copying a generic list from another assessment without linking each measure to a specific impact.
Avoid these by tying every measure to an impact, using verbs that can be checked, collecting evidence and reviewing measures regularly.
Not every measure can be delivered at once. Start with those that address the most severe and irreversible impacts, and with quick wins such as narrowing scope or changing a default. Then plan larger items, such as building an appeal process or retraining a model, with clear milestones. Explain the order in the assessment so a reviewer can see that the sequence was deliberate.
Consider a phased release: start with a smaller population, a more limited use or extra human review, and expand as evidence builds that the measures work. That approach reduces exposure while generating real-world data.
An employer uses a tool to screen job applications. The assessment identifies that candidates with career gaps, more common among carers, are ranked lower. Mitigation measures: remove gap-related features; test selection rates by sex and age; ensure a recruiter reviews all rejections of candidates who meet minimum criteria; inform candidates about the tool and how to request review; and monitor outcomes monthly.
Each measure has an owner and date, and evidence is stored: test reports, procedure documents and monitoring dashboards. Residual impact drops from serious to limited, and the head of HR approves it with a six-month review. The mitigation plan is easy to explain to candidates and to regulators.
Controls decay. Reviewers change, thresholds drift, staff forget procedures and models are updated. Include periodic checks: a quarterly confirmation from owners that the measure is in place, spot tests by an independent team and re-testing after major changes.
Feed lessons from incidents and complaints back into the measures. If a harm occurs despite mitigation, ask why the measure failed and whether other systems share the weakness. Update the assessment and share the improvement across the portfolio.
If you want a report and workbook that link impacts, mitigation measures, owners, evidence and residual ratings in one place, the AI Impact Assessment Report and Workbook provides a structured layout built around ISO/IEC 42005. Whatever tool you use, effective AI impact assessment mitigation is specific, owned, evidenced and proportionate to the harm it is meant to prevent.
The measures chosen to prevent, detect, respond to and repair the impacts of an AI system on people, together with owners, evidence and a judgement of the residual impact.
Design and scope changes that remove the risk often work best, combined with testing, well-designed human oversight, transparency and accessible redress.
Not usually. Human review helps but can fail through fatigue, bias or lack of information, so it should be combined with other measures and designed carefully.
Collect evidence such as test results, logs, training records, appeal outcomes and monitoring data, and check it on a schedule.
A named senior person with authority over the system and its risks, taking advice from privacy, legal, risk and affected stakeholders.
September 30, 2026
