This step finds the problems that could arise with each AI and sets the magnitude of risk (1 to 25 points) by multiplying the likelihood of occurrence by the damage if it occurs, in order to decide where to place controls first. You receive the risk assessment register and the treatment plan.
Using the Step 4 AI system inventory as input, all AI systems and tasks whose evaluation type is formal or simplified and which are included in the scope are assessed without omission. The evaluation type (formal, simplified, excluded) is determined, and risks in six categories (cause → event → consequence) are evaluated as likelihood × severity (1 to 25, with the severity axis noted).
Next, the treatment option, the controls that serve as the basis, the cross-check against the Statement of Applicability and the post-treatment risk level are determined. The results become input to 6 Impact Assessment and 7 Control Implementation and Statement of Applicability.
Output
- Reconciliation summary (input N = assessed M + excluded K) and consolidated registerFor each task: assessment unit · LEVEL · evaluation type · highest risk · risk level · grade · impact assessment · reassessment trigger
- Risk details and treatment plan (Excel)Cause → event → consequence / 23894 source and objective / severity axis / treatment option · basis controls · cross-check against the Statement of Applicability / post-treatment risk level
- Targets for impact assessment (to Step 6) · link to the Statement of Applicability (to Step 7) · self-check from the auditor's viewpoint (NC and OFI candidates)
What to do in this step, by scope of application
| Development | Supply | Use |
|---|---|---|
| Focused on lifecycle, data (bias, quality, provenance), robustness and security risks | Focused on misuse, customer harm, contractual liability and inadequate notification risks | Focused on incorrect output, personal information, input of trade secrets and provider dependency risks |
Assessment method: full assessment using the AI system inventory
- Based on ISO/IEC 23894:2023: the assessment follows 6.3 scope, context and risk criteria → 6.4.2 identification (assets · risk sources · events and consequences · current controls and their operating effectiveness · consequences for the organization, individuals and society) → 6.4.3 analysis (business, individual and societal impacts assessed separately, likelihood) → 6.4.4 evaluation → 6.5 treatment (risk-benefit analysis when the acceptance level is not met) → 6.6 and 6.7 monitoring and recording (traceability, items of the assessment record).
It is matched to the lifecycle stage of the project (Annex C). The text of the standard is not reproduced; grounds are cited by clause number. - Assessment target = AI system inventory: the "AI System Inventory" made in Step 4 (the Excel file or the Step 4 output in this browser) is the input, and every AI system or project whose assessment type is formal or simplified and which is within the scope is assessed. The AI number, purpose, AI role, supplier, model, input data, affected parties and LEVEL are taken over from the inventory as they are.
Under "Specify materials" (optional), specify current controls, incidents, error records, procedures and the previous risk assessment register, which are used as grounds for the assessment. Files are read only inside the browser and are not stored. - Full reconciliation: every AI project in the materials is assessed regardless of its status (planning, development, completed, in operation), and the count is reconciled as "inputs N = assessed M + excluded K". The project number is used as the key, and where an assessment differs from the previous one, the previous value, the changed value and the reason are recorded together. The assessment unit is divided into projects (before deployment) and AI systems in operation (periodic reassessment after completion or operation).
- Assessment type: Formal: any one of external customer exposure · mass processing of personal information or automated decisions about people such as recruitment or evaluation · LEVEL 3 or 4 · connection to live data of core systems / Simplified: other LEVEL 2 (converted to formal if 13 points or more) / Excluded: LEVEL 1 local (reason recorded). A possible high-impact AI under the AI Framework Act is marked as "high-impact review".
- Identification: for each project, six categories are checked (technology · data · ethics and society · laws and regulations · business and supply chain · lifecycle), and each risk is written as "cause → event → consequence", with the ISO/IEC 23894 Annex B sources (B.2 to B.8) and Annex A objectives (A.2 to A.12) attached.
- Treatment: for each treatment option (avoid · pursue an opportunity · remove the source · reduce likelihood · reduce consequences · transfer or share · accept (6 points or less)), the grounds control from ISO/IEC 42001 Annex A is attached and compared with the Statement of Applicability (applied · not applied · additional control · reflection needed). Next, the risk level after treatment and the approver of residual risk are determined. Information security controls refer to ISO/IEC 27001:2022 Annex A as a linked reference and are placed as "additional controls".
Scale: if your company has a risk assessment procedure, follow it
| Score | Likelihood (default values) |
|---|---|
| 5 | At least once a month · possible with every input |
| 4 | At least once a quarter · history of occurrence |
| 3 | 1 to 2 times a year |
| 2 | Less than once a year |
| 1 | Less than once in 3 years |
Severity (1 to 5) = the highest value of the three axes (business · individual · society), with the adopted axis recorded together (example: 4-individual). Grade: 1 to 6 low risk · 7 to 12 medium risk · 13 to 19 high risk · 20 to 25 critical; acceptance criterion is 6 points or less. For borderline cases a higher score is used, and a control that exists only on paper is not treated as a current control.
Common control codes: risks that recur across several projects are handled at once
| Code | Risk | Control |
|---|---|---|
| CC-01 | External data collection | Official APIs and RSS first, record of review of the target's terms and robots rules, citation centered on titles and links, source and collection date shown, internal distribution only (A.7.3, A.7.5) |
| CC-02 | Transmission to external generative AI | Comply with the prohibited input information in the generative AI use guideline, masking before transmission, secret management of API keys, confirmation of the supplier's data training and retention terms (A.7.6, A.10.3, ISO/IEC 27001 A.5.17 and A.5.23) |
| CC-03 | Generation and RAG | Show the source of the grounds, do not answer when there are no grounds, "AI-generated" notice, periodic sample checks of accuracy (A.6.2.4, A.8.2, A.6.2.6) |
| CC-04 | Automated calculation and aggregation | Verification with standard test cases, anomaly detection against reference values, ledger reconciliation, staff member approval before finalization (A.6.2.4, A.7.4, A.9.4) |
Output and next steps
- Reconciliation summary → Excel sheets "Consolidated register" (one row per project) · "Risk details and treatment plan" (one row per risk; risk level and grade are Excel formulas) · "Impact assessment targets" · "SoA linkage" → self-check from the auditor's viewpoint (NC and OFI candidates) · To be confirmed.
- Step 6 AI System Impact AssessmentProjects scoring 13 points or more are passed on (sensitive uses below 13 points are marked "score review recommended"). If you choose "Import risk assessment results" in Step 6, these results are entered as they are.
- Step 7 Control Implementation and Statement of ApplicabilityThe grounds controls cited in the treatment plan and the "SoA reflection needed" marks become the grounds for application in the Statement of Applicability. When you open Step 7, these results are imported automatically (results kept in this browser).
Grounds: ISO/IEC 42001:2023 6.1.1 to 6.1.4, 8.2 to 8.4, Annexes A and B; ISO/IEC 23894:2023 (identification, analysis, evaluation, treatment, monitoring, Annexes A, B and C); ISO/IEC 42005:2025; ISO/IEC TR 24028, TR 24027, 22989, 5338, 5259-3; ISO/IEC 27001:2022 Annex A (linked reference). Judgment and approval are made by the person responsible designated by your company.
Next steps
Previous steps: 4 AI System Inventory · 3 AI System Selection · 1 AI Context Analysis · 2 AI Policy, Roles and Objectives · 11 Monitoring, Change and Revision Control · 15 Nonconformity and Corrective Action · 13 Legal Requirements Identification and Monitoring · 14 Management Review · 16 Improvement · 10 Documented Information Control · The output of this step is carried forward in: 2 AI Policy, Roles and Objectives · 6 AI System Impact Assessment · 7 Control Implementation and Statement of Applicability (SoA) · 8 Verification and Use Approval · 9 Operational Control · 10 Documented Information Control · 11 Monitoring, Change and Revision Control · 12 Internal Audit · 14 Management Review · 15 Nonconformity and Corrective Action
