Risk, Ethics & Compliance ยท RSK-02
AI risk assessment method
The organisation uses one consistent, documented method to assess the risks of each use of AI, and the results feed its corporate risk management.
Why it matters
Different teams assessing AI risk in different ways produce results that cannot be compared or added up. A single method, built into the existing risk process, lets the organisation see its overall AI risk and makes sure the risks particular to AI are not missed.
Levels
| 0 Absent | Confirmed absent: the evidence shows nothing is in place. Not known is recorded as insufficient evidence, not 0. |
|---|---|
| 1 Initial | AI risk is assessed, if at all, by individual judgement with no common method. |
| 2 Developing | A method or checklist exists, but it is used in some areas only, does not cover risks particular to AI, or sits outside corporate risk management. |
| 3 Defined | A documented, proportionate method assesses every use of AI before approval and in live use, records treatment and residual risk, and feeds corporate risk registers. |
| 4 Managed | Assessments are quality checked, and the AI risk profile, including trends in residual risk, is reported to the oversight forum. |
| 5 Optimising | The method is refined from incidents, near misses and assessment reviews, and its results are shown to predict where problems occur. |
Level 3 criteria
Level 3 is met when every applicable criterion is met. Criteria marked with an overlay apply only to organisations with that characteristic (overlays).
- RSK-02.1 A documented AI risk assessment method, proportionate to risk tier, covers risks particular to AI as well as general risks.
- RSK-02.2 Every use of AI is assessed with the method before approval.
- RSK-02.3 Each use of AI is assessed again during live use and on material change.
- RSK-02.4 Assessment results are recorded in the corporate risk registers.
- RSK-02.5 Each assessment records the treatment chosen for each risk and the controls applied, with the reasons for choosing them.
- RSK-02.6 Each assessment records the residual risk.
- RSK-02.7 The method covers the recognised generative AI risk categories. GEN
Assessment questions
- How is the risk of a proposed use of AI assessed, and is the method the same across the organisation?
- Does the method cover risks particular to AI, such as wrong outputs, drift or misuse, alongside general ones?
- Where do assessment results go, and when is a use reassessed?
- The documented risk assessment method and guidance
- A sample of completed assessments for recent uses of AI
- Corporate or departmental risk registers showing AI risks
- Risk treatment records showing the controls selected for each use and why
- Reassessment records for live uses, with residual risk recorded
How to reach Level 3
- Write a risk assessment method with depth scaled to risk tier, covering AI-specific risks (such as error, drift, misuse and opacity) as well as general ones, and the recognised risk categories for generative AI.
- Require each assessment to record the treatment chosen for each risk, the controls applied with the reasons for choosing them, and the residual risk.
- Apply the method to every use before approval, and again during live use and on material change.
- Record results in the corporate risk registers using the organisation's standard scales.
Common pitfalls
- Assessing once at approval and never again as the use or its data changes.
- Keeping AI risks in a separate register that corporate risk management never sees.
Starter artefacts
- AI risk assessment method and template by tier
- AI risk catalogue including generative AI categories
- Treatment and controls record with reasons
Ownership and sequence
- Accountable
- risk lead
- Contributors
- ai governance lead, business owner, security lead
- Depends on
- RSK-01
- Roadmap step
- 3
- Lifecycle
- design, acquire_build, deploy, operate
Cross-references
| AI Cyber Security Code of Practice | 2.1 partial; 3.1 partial; 3.4 partial; 4.3 partial; 7.2 partial |
|---|---|
| Algorithmic Transparency Recording Standard | Tier 2 Risks, mitigations and impact assessments direct |
| EU AI Act | Art 9 partial; Art 17 partial |
| GDS AI Playbook | P10 partial |
| ISO/IEC 42001 | 6.1.1 partial; 6.1.2 direct; 6.1.3 direct; 8.2 direct; 8.3 direct |
| JSP 936 Part 1 | para 6 partial; para 18 partial; para 49d partial; para 81 partial; para 82 partial; para 86 partial; para 87 partial; para 88 partial; para 92 direct; para 141 partial; para 142 partial; para 181 partial; para 191 partial; para 192 partial; para 194 direct; para 196 partial |
| NIST AI 600-1 | R1 partial; R3 partial; R5 partial; R11 partial |
| NIST AI RMF 1.0 | GOVERN 1.3 direct; GOVERN 1.4 direct; GOVERN 4.2 direct; MAP 3.2 partial; MAP 4.1 partial; MAP 4.2 partial; MAP 5.1 partial; MEASURE 2.6 partial; MEASURE 3.1 direct; MANAGE 1.2 direct; MANAGE 1.3 direct; MANAGE 1.4 partial |
| UK Data and AI Ethics Framework | Transparency 8 partial; SAT 2.11 partial |
| UNESCO AI Maturity Framework | 5.3 direct |
direct: this control at Level 3 meets the item. partial: it covers part of it.