Appendices
Reference materials including glossary, definitions, architectures, maturity model, FAQ, internal audit programme, and continuous improvement processes.
Glossary & Definitions
This glossary provides standardised definitions for AI governance terminology used throughout the framework. All terms align with ISO/IEC 22989, ISO/IEC 23053, and NIST AI RMF where applicable.
Core Terminology
| Term | Definition |
|---|---|
| Artificial Intelligence (AI) | An engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives or parameters. |
| Machine Learning (ML) | A subset of AI in which a system learns from data to improve its performance on a specific task without being explicitly programmed. |
| Foundation Model | An AI model trained on broad data at scale, designed for generality of output and adaptability to a wide range of downstream tasks. |
| Generative AI | AI systems that create new content including text, images, audio, video, or code in response to prompts or inputs. |
| Large Language Model (LLM) | A type of foundation model trained on vast text corpora, capable of understanding and generating human language. |
| Prompt Injection | An attack technique where malicious instructions are embedded in user inputs to manipulate AI system behaviour or extract sensitive information. |
| Model Poisoning | An attack on the training data or training process that causes a model to behave incorrectly or maliciously. |
| Hallucination | A phenomenon where an AI system generates plausible but false or ungrounded information. |
| Bias | Systematic error in model outputs that results in unfair outcomes for specific groups or individuals. |
| Explainability | The degree to which a human can understand the reasoning behind an AI system's output or decision. |
| Human-in-the-Loop (HITL) | A governance approach where human oversight and approval are required for AI system decisions or outputs. |
| Data Lineage | The complete lifecycle of data including its origins, transformations, and movement through the organisation. |
| AI Risk Appetite | The amount and type of AI risk that an organisation is willing to pursue or retain. |
| Residual Risk | The remaining risk after risk treatment and control implementation. |
| RACI Matrix | A responsibility assignment matrix defining who is Responsible, Accountable, Consulted, and Informed for each task or decision. |
Governance Maturity Model
The AI Governance Maturity Model provides a five-level scale for assessing and benchmarking an organisation's AI governance capabilities. Each level builds upon the previous, with clear criteria for advancement.
Maturity Levels
| Level | Name | Characteristics | Typical Indicators |
|---|---|---|---|
| 1 | Initial | Ad hoc, reactive governance; no formal policies; shadow AI prevalent | No AI inventory; no risk assessments; incidents managed informally |
| 2 | Developing | Basic governance established; initial policies drafted; some awareness | AI inventory started; basic risk register; informal review processes |
| 3 | Defined | Formal governance framework implemented; policies approved; roles assigned | All 12 volumes operational; regular reporting; training programme active |
| 4 | Managed | Quantitative governance; metrics-driven; continuous monitoring; external validation | KPI dashboards; internal audit programme; ISO 42001 aligned |
| 5 | Optimising | Industry-leading governance; innovation balanced with control; external certification; thought leadership | ISO 42001 certified; public transparency reports; industry standards contribution |
Assessment Criteria by Domain
| Domain | Level 1 | Level 3 | Level 5 |
|---|---|---|---|
| Strategy | No AI strategy | Board-approved AI strategy with 3-year roadmap | AI strategy reviewed annually, contributes to industry standards |
| Risk | No risk management | Formal risk framework with quarterly reviews | Predictive risk analytics, stress testing, board-level dashboards |
| Data | No data governance | Data classification, quality standards, lineage tools | Automated data quality monitoring, real-time lineage |
| Security | Basic IT security | AI-specific security controls, threat modelling | Zero-trust AI architecture, automated threat response |
| Operations | No operational oversight | Monitoring, SLAs, incident response procedures | Self-healing systems, predictive maintenance, automated governance |
| People | No AI training | Role-based training, competency framework | AI literacy across all levels, external faculty, research partnerships |
Assessment Guidance
Maturity assessments should be conducted annually by Internal Audit or an external assessor. Organisations should target Level 3 within 12 months of framework adoption and Level 4-5 within 24 months. Assessment results are reported to the Board Risk Committee.
Frequently Asked Questions
Common questions from executives, board members, legal counsel, and operational staff regarding the AI Governance Framework implementation and operation.
Executive & Board Questions
- Q: How does this framework differ from our existing IT governance? A: AI governance extends beyond traditional IT governance to address model-specific risks including bias, hallucination, explainability, and ethical considerations that standard IT frameworks do not cover.
- Q: What is the Board's liability exposure if we do not implement AI governance? A: Directors may face personal liability under corporate governance laws for failure to exercise reasonable care in overseeing AI risks, particularly where AI decisions cause harm or regulatory breach.
- Q: How much should we budget for full implementation? A: Budget ranges from $500K for SMEs to $5M+ for large multinationals, spread over 18-24 months. The cost of non-compliance typically exceeds implementation investment.
- Q: Can we implement this framework if we use third-party AI services exclusively? A: Yes. The framework applies to all AI use regardless of deployment model. Third-party use requires additional vendor governance controls per Volume 5.
Operational Questions
- Q: Who is responsible when an AI system makes a harmful decision? A: Accountability follows the RACI matrix. The system owner (business unit leader) retains accountability. The AI Ethics Board reviews systemic issues.
- Q: How do we handle AI systems developed before this framework existed? A: All existing AI systems must be inventoried and assessed within 90 days of framework adoption. High-risk legacy systems require immediate remediation plans.
- Q: What training is required for staff who do not work directly with AI? A: All staff require AI literacy awareness training. Staff interacting with AI systems require role-specific training per the competency framework.
- Q: How often should policies be reviewed? A: Core policies require annual review. Operational standards require 6-monthly review. All policies require immediate review after significant incidents or regulatory changes.
Technical Questions
- Q: Do we need to build our own AI models to comply? A: No. The framework supports build, buy, and hybrid strategies. Third-party models require vendor assessment and contractual governance clauses.
- Q: How do we monitor for model drift in production? A: Implement statistical drift detection comparing production input distributions and output distributions against training baselines. Alert thresholds require calibration per use case.
- Q: What is the minimum documentation required for a production AI system? A: At minimum: Model Card, Risk Assessment, Data Governance Plan, Security Review, and Operations Runbook. See Volume 7 for complete requirements.
- Q: How do we handle prompt injection attacks? A: Implement input validation, prompt sanitisation, output filtering, and adversarial testing. See Volume 5 (Security) and Volume 6 (Incident Response) for detailed controls.