An AI governance programme can pass its launch review or certification audit and still lose track of the AI systems in use within a year, because new use cases, suppliers and model changes arrive faster than the register is updated. This health check is a recurring audit for an AI management system: ten KPIs with formulas, data source types and thresholds, a cadence with named approvers, a findings log and a remediation map. It is run quarterly by the AI governance lead, re-performed yearly by internal audit, and reported to the risk committee or board. The operating model, policies, risk tiers and oversight design are set out in the AI Governance Guide; this page measures whether they are working and does not repeat them.
What this health check covers
It is also different from the ISO 42001 Readiness Checklist, which prepares an organisation for a one-time certification audit clause by clause. This page is the recurring health check that keeps the system in shape between audits, and it works whether you follow ISO/IEC 42001, the NIST AI RMF or both. If you are looking for AI governance KPIs, AI risk management metrics or an AI governance audit checklist with targets, start here.
Key takeaways
- The register is the denominator: nine of the ten KPIs depend on the AI system inventory, so measure its completeness against discovery first and treat a weak inventory as a red finding in itself.
- Coverage means current, not ever done: a risk classification or impact assessment counts only if it post-dates the system's last material change.
- Suppliers are in scope: most AI in use arrives inside third-party products, so the share of third-party AI assessed is a core KPI, not an annex.
- Exceptions and incidents show the truth: the age of policy exceptions and how incidents are logged and closed reveal whether governance is followed in practice.
- Thresholds are recommendations: the values below are NHI Mgmt Group starting points; record your own risk appetite and the reason for any change.
How to use this template
Copy the metrics table and the findings log into your own tracker, then run this AI Governance Programme Health Check at the cadence set out below. Collect each figure from the data sources described, score it against the green, amber and red thresholds, and record every amber or red result as a finding with an owner and a due date. Keep each completed review so you can compare results quarter on quarter: the trend matters as much as any single score.
Cadence and sign-off
| Review | Frequency | Owner | Approver | Evidence kept |
|---|---|---|---|---|
| Inventory reconciliation (G1) | Monthly | AI governance office | AI governance lead | Discovery exports, register snapshot, reconciliation sheet |
| Full KPI run (G1 to G10) | Quarterly | AI governance lead | Accountable executive (chair of the AI risk or review committee) | KPI workbook, findings log, committee minutes |
| Third-party AI review (G5) | Quarterly, and at contract renewal | Procurement and third-party risk | AI governance lead | Supplier assessments, contract clauses, exceptions |
| Board or risk committee report (G10) | At least half-yearly | AI governance lead | Board or delegated committee | Report pack, minutes showing discussion and decisions |
| Independent re-performance | Yearly | Internal audit | Audit committee | Re-performed samples, variance notes |
Metrics
"AI system" follows the definition in your AI policy and includes AI inside purchased products and AI agents. "High-risk" means the top tiers of your own risk tiering scheme. NHI Mgmt Group recommended thresholds; adjust to your risk appetite.
| ID | Metric | Formula (numerator / denominator) | Data source type | Green | Amber | Red |
|---|---|---|---|---|---|---|
| G1 | AI system inventory completeness | AI systems found by discovery that were already in the register / AI systems found by discovery | AI register; procurement and expense records; SaaS and OAuth grant inventories; network or AI gateway logs; code and model repositories | ≥ 95% | 85% to below 95% | < 85% |
| G2 | Risk classification coverage | Registered AI systems with a risk tier assigned after their last material change / registered AI systems | AI register; change records | ≥ 98% | 90% to below 98% | < 90% |
| G3 | Impact assessment coverage | High-risk AI systems with an approved impact assessment less than 12 months old and post-dating the last material change / high-risk AI systems | Impact assessment records; AI register | 100% | 90% to below 100% | < 90% |
| G4 | Model and system card coverage | Production AI systems with a current model or system card (purpose, data, limitations, evaluation results, owner) / production AI systems | Documentation repository; AI register | ≥ 95%; 100% for high-risk | 80% to below 95% | < 80%, or any high-risk system without one |
| G5 | Third-party AI assessed | Third-party AI products and services with a completed AI-specific supplier assessment / third-party AI products and services in the register | Third-party risk records; procurement contracts | ≥ 95% | 80% to below 95% | < 80% |
| G6 | AI incident log health | AI incidents and near misses closed within the target time / AI incidents and near misses closed in the period; plus count logged per quarter (trend) | Incident management records tagged as AI | All within target; log in use | One overdue | Two or more overdue, or zero logged across a large estate |
| G7 | Monitoring coverage | Production AI systems with a defined post-deployment monitoring plan and evidence of a review in the last quarter / production AI systems | Monitoring dashboards and alert records; monitoring plan register | ≥ 95%; 100% for high-risk | 80% to below 95% | < 80% |
| G8 | AI literacy training completion | In-scope staff who completed role-appropriate AI training in the last 12 months / in-scope staff (users, builders, approvers) | Learning management system records; HR role data | ≥ 95% | 80% to below 95% | < 80% |
| G9 | Policy exception age | Open AI policy exceptions past their expiry date / open AI policy exceptions | Exception register | 0% | Any past expiry under 30 days | Any past expiry 30 days or more, or exceptions with no expiry |
| G10 | Board reporting cadence | Board or committee AI governance reports delivered on schedule in the last 12 months / reports scheduled | Board and committee minutes | 100%, with decisions minuted | One late | Missed, or reported without minuted discussion |
Notes on the formulas
- G1 is measured against discovery, not self-declaration: asking teams to declare AI use yields the register you already have. Use spend, OAuth grants, gateway logs and repositories. The Shadow AI Discovery Guide lists sources.
- "Material change" needs a definition: new model or model version, new data source, new user group, new tool access or new purpose. Write it into the policy once so G2 and G3 are comparable between quarters.
- G6 low counts are a warning: a large AI estate with no logged incidents or near misses usually means people do not know what to report. NIST AI RMF MANAGE 4.3 expects incidents and errors to be communicated, tracked and responded to.
- G7 counts evidence of review: a dashboard nobody looks at is not monitoring. Accept a dated review note, alert triage record or monitoring meeting minute.
- G8 is role-based: builders, approvers and general users need different content. NIST AI RMF GOVERN 2.2 calls for AI risk management training for personnel and partners; organisations in scope of the EU AI Act should also read the EU AI Act AI Literacy Obligations (Article 4) Guide.
How to collect the data
- Snapshot the AI register: system ID, owner, risk tier and date assigned, last material change date, links to impact assessment, model or system card and monitoring plan, supplier reference.
- Run discovery: procurement and expense lines with AI products, SaaS admin and OAuth grant inventories, network and AI gateway logs of model API calls, code and model repositories, and cloud AI service inventories. Reconcile with the register for G1.
- Join change and assessment records: compare each system's last material change date with the date of its risk classification and impact assessment for G2 and G3.
- Check the documentation store: for G4, confirm each card exists, names an owner and was updated after the last change.
- Pull supplier and incident records: third-party risk assessments for G5; incident tickets tagged as AI, including near misses, for G6.
- Collect monitoring evidence: the monitoring plan register and a dated review artefact per system for G7.
- Pull training, exception and governance records: learning records joined to HR roles for G8, the exception register for G9, and committee minutes for G10.
- Sample and verify: trace five random systems end to end through every KPI. A KPI that cannot be traced is reported amber at best.
Red flags that need action now
- A high-risk AI system in production with no impact assessment, or one that predates a model or purpose change (G3).
- AI systems found processing personal or regulated data that are not in the register (G1).
- Third-party AI handling sensitive data without an AI-specific supplier assessment or contract terms on data use (G5).
- No AI incidents or near misses logged in a year across a large estate (G6).
- Production high-risk systems with no monitoring plan or no reviewed alerts (G7).
- Policy exceptions with no expiry date, or expired exceptions still in use (G9).
- The board or committee has not received an AI governance report in the last six months (G10).
- Systems whose owner has left the organisation and no successor is named.
Findings log
| Finding ID | Finding | Metric | Severity | Owner | Due date | Status |
|---|---|---|---|---|---|---|
| HC-AIG-01 | [Number] AI systems found via OAuth grants and spend not in the register | G1 | [High / Medium / Low] | [Role] | [Date] | [Open / In progress / Closed / Risk accepted] |
| HC-AIG-02 | [Number] high-risk systems with impact assessments older than their last change | G3 | [Severity] | [Role] | [Date] | [Status] |
| HC-AIG-03 | [Number] third-party AI services without AI supplier assessment | G5 | [Severity] | [Role] | [Date] | [Status] |
| HC-AIG-04 | [Number] expired policy exceptions still in use | G9 | [Severity] | [Role] | [Date] | [Status] |
| HC-AIG-05 | [Finding] | [Metric] | [Severity] | [Role] | [Date] | [Status] |
Group systems with a shared root cause into one finding. Risk accepted findings carry an expiry date and the approver's name, and are re-opened at the next quarterly run.
Remediation map
| Finding type | Metrics | Where the fix is explained |
|---|---|---|
| Incomplete inventory | G1 | AI Register and AI Inventory Template; Shadow AI Discovery Guide |
| Missing or stale risk tiers | G2 | AI Use Case Risk Tiering Guide; AI Risk Assessment Guide |
| Missing impact assessments | G3 | AI Impact Assessment Template |
| Missing model or system cards | G4 | Model Card and System Card Template |
| Unassessed suppliers | G5 | Third-Party AI Risk Guide; AI Vendor Security Questionnaire Template |
| Incident handling gaps | G6 | AI Incident Management and Escalation |
| Monitoring gaps | G7 | AI Governance Guide (monitoring, incidents and decommissioning) |
| Training gaps | G8 | AI Literacy Training Plan Template |
| Exception and reporting failures | G9, G10 | AI Ethics Board and AI Review Committee Guide; Board Cyber Risk Reporting Guide |
Related guides
- AI Governance Guide: the programme this health check audits.
- AI Security Maturity Model: place the KPI results on a maturity scale.
- ISO 42001 Readiness Checklist and ISO/IEC 42001 Guide: certification preparation and the standard itself.
- NIST AI RMF Guide: the framework most of these KPIs map to.
- Agent Governance and Risk Guide: governance for AI agents, the fastest growing part of the AI register.
References
Framework anchors for the metrics: NIST AI RMF GOVERN 1.5 (ongoing monitoring and periodic review), GOVERN 1.6 (AI system inventory), GOVERN 2.2 (training), GOVERN 6.1 (third-party AI risk), MAP 1.1 (intended purpose and context), MEASURE 2.4 (production monitoring), MANAGE 4.1 (post-deployment monitoring plans) and MANAGE 4.3 (incident communication); NIST SP 800-53 PM-5 (System Inventory), PM-6 (Measures of Performance), RA-3 (Risk Assessment), SR-6 (Supplier Assessments and Reviews) and CA-7 (Continuous Monitoring); NIST CSF 2.0 GV.OV, GV.SC, ID.AM and PR.AT.
- NIST: AI Risk Management Framework (AI RMF 1.0)
- NIST: AI RMF Playbook
- NIST: SP 800-53 Rev. 5
- NIST: Cybersecurity Framework 2.0
Related NHI Mgmt Group resources: AI Governance Guide, NIST AI RMF Guide, AI Register and AI Inventory Template, Third-Party AI Risk Guide.