Measure & Improve ยท MON-05
Independent assurance reviews
Uses of AI are reviewed by people independent of their delivery, with the depth and frequency of review proportionate to risk.
Why it matters
Teams that build or run a use of AI are not well placed to judge it objectively. Independent review gives leaders confidence that controls work in practice, and directs review effort to the uses that carry most risk.
Levels
| 0 Absent | Confirmed absent: the evidence shows nothing is in place. Not known is recorded as insufficient evidence, not 0. |
|---|---|
| 1 Initial | No independent review; delivery teams assure their own work. |
| 2 Developing | Independent reviews happen for occasional high-profile uses, without a defined basis for which uses are reviewed or how. |
| 3 Defined | Every use of AI gets independent review set by risk tier and triggered by change, findings are owned and reported, and AI governance as a whole is audited periodically. |
| 4 Managed | Assurance coverage against the plan is reported, findings are tracked to closure, and overdue actions are escalated. |
| 5 Optimising | The assurance approach is refined from findings, incidents and external developments, and uses methods that test real behaviour as well as documents. |
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).
- MON-05.1 A defined assurance approach sets the depth and frequency of independent review for each risk tier.
- MON-05.2 Independent reviews are carried out on all uses of AI in line with the approach.
- MON-05.3 Reviews are also triggered by material change or emerging risk.
- MON-05.4 Reviews check compliance against the obligations register (RSK-04).
- MON-05.5 Every finding is recorded, with an owner and a remediation date set by risk.
- MON-05.6 Findings are tracked, and their status is reported to leadership (MON-08).
- MON-05.7 The AI governance arrangements as a whole are independently audited on a set cycle.
Assessment questions
- Which uses of AI have had an independent review, and how was the reviewer independent?
- How is the depth and frequency of review decided?
- What happened to the findings of the last review?
- Finding log with owners, remediation dates and status, and the record of the periodic audit of AI governance as a whole
- The assurance plan or schedule by risk tier
- Recent review reports with findings and action tracking
How to reach Level 3
- Write an assurance approach that sets the depth and frequency of independent review for each risk tier (RSK-02), covering all uses of AI in the inventory (GOV-06).
- Add triggers for extra reviews on material change or emerging risk, and include a check of compliance against the obligations register (RSK-04) in every review.
- Record each finding with an owner and a remediation date set by risk, and report status to leadership (MON-08).
- Arrange an independent audit of the AI governance arrangements as a whole on a set cycle, such as through internal audit or an external reviewer.
Common pitfalls
- Using reviewers from the delivery team's own line, who are not truly independent.
- Reviewing only high-profile uses and leaving lower-risk uses with no review at all.
- Recording findings without owners or dates, so they never close.
Starter artefacts
- Assurance plan by risk tier
- Findings log with owners and dates
- Governance audit schedule
Cross-references
| AI Cyber Security Code of Practice | 9.2 partial |
|---|---|
| Algorithmic Transparency Recording Standard | Tier 2 Deployment context partial |
| EU AI Act | Art 9 partial |
| GDS AI Playbook | P10 direct |
| ISO/IEC 42001 | 8.2 partial; 9.2 direct |
| JSP 936 Part 1 | para 6 partial; para 76 partial; para 92 partial; para 99 partial; para 101 direct; para 105 partial; para 131 partial; para 188 partial; para 189 partial; para 198 partial; para 206 partial |
| MITRE AI Maturity Model | P2.3 direct; P4.2 partial; P5.3 partial |
| NIST AI RMF 1.0 | GOVERN 1.5 partial; MEASURE 1.2 partial; MEASURE 1.3 direct; MEASURE 2.13 partial; MEASURE 3.1 partial; MEASURE 3.2 partial; MEASURE 4.2 partial |
| UK Data and AI Ethics Framework | Accountability 3 direct |
direct: this control at Level 3 meets the item. partial: it covers part of it.