BK-GA³™ AI Assessment Framework
The BK-GA³™ framework helps organizations identify, assess, and manage AI-related risks. It can be used to evaluate internal AI risk, support the development or maturation of AI risk management programs, and assess third-party suppliers.
How to Use BK-GA³™
For community users
The BK-GA³™ Framework is publicly available below, and can also be accessed as an Excel spreadsheet. This can be used for manual mapping or to share with vendors via email for third-party assessments.
For Black Kite customers
BK-GA³™ is available within the Cyber Assessments module, enabling automated AI risk assessments directly in the platform.
BK-GA³™ Benefits
Standardized Risk Assessments
Apply a single AI framework across all vendors regardless of their industry, size, or region.
Efficient Third-Party Oversight
Reduce duplicate efforts across your vendor network, making it easier to manage AI risk at scale.
Gap Identification & Prioritization
Identify specific areas where AI controls are lacking so you can more easily prioritize vendor outreach based on the severity of their gaps.
BK-GA³™ Framework Categories
AI Governance & Strategy
AI Legal & Regulatory Compliance
AI Data Governance & Management
AI Model Risk & Control
AI System Security & Access Control
AI Accountability & Incident Response
AI Transparency & Disclosure
Human Oversight of AI Systems
Bias, Fairness & Non-Discrimination in AI
AI Privacy & Data Confidentiality
AI Explainability & Interpretability
AI Testing, Validation & Evaluation
AI System Documentation & Traceability
AI Training & Awareness Programs
Third-Party & Vendor AI Risk Management
AI Continuity & Operational Resilience
AI Monitoring, Auditing & Logging
AI Infrastructure & System Architecture
AI Ethics, Sustainability & Responsible Use
AI Third-Party Governance & Risk Management
Category
AI Governance & Strategy
ID | Description | Standards Mapping |
|---|---|---|
GOV-001 | The organization establishes and maintains a board-approved AI governance framework that defines AI-specific roles, responsibilities, and decision authority. | ISO/IEC 42001 (AI governance), NIST AI RMF (Govern function) |
GOV-002 | An AI steering committee operates under a documented charter and meets regularly to oversee AI-related risks, ethics, and compliance escalations. | OECD AI Principles (Technical leadership), ISO/IEC 42001 (Roles & Responsibilities) |
GOV-003 | A central inventory of AI systems is maintained, classified by risk level, business criticality, and applicable AI regulations or ethical standards. | ISO/IEC 42001 (AI inventory), NIST AI RMF (Govern - Asset Management), EU AI Act (Article 29) |
GOV-004 | Periodic AI governance maturity assessments are conducted using recognized frameworks (e.g., ISO/IEC 42001, NIST RMF), with documented plans for improvement. | CMMI (Maturity models), ISO/IEC 38507 (Governance of IT for AI) |
GOV-005 | All AI-related investment proposals undergo approval based on documented business justification, AI risk assessments, and projected outcomes. | COBIT 2019 (Value delivery & investment planning), ISO/IEC 42001 (Strategic alignment) |
GOV-006 | The organization maintains an AI strategy aligned with enterprise goals, supported by a roadmap, success metrics, and dedicated AI-specific resource plans. | ISO/IEC 38500 (Strategic planning), ISO/IEC 42001 (AI strategy alignment) |
GOV-007 | All AI governance documentation is version-controlled, reviewed annually, and accessible to stakeholders involved in AI development, deployment, or oversight. | ISO 9001 (Documented information control), ISO/IEC 42001 (Policy documentation) |
GOV-008 | A cross-functional governance structure is in place to oversee AI systems, ensuring involvement from legal, risk, compliance, data science, and business units. | TOGAF (Enterprise architecture governance), ISO/IEC 42001 (Organizational structure) |
GOV-009 | Organization has a designated AI governance officer or equivalent role responsible for AI policy enforcement and coordination. | ISO/IEC 42001 (Governance roles), NIST AI RMF (Organizational responsibilities) |
GOV-010 | Organizational AI governance policy is mapped to external frameworks (e.g. ISO/IEC 42001, NIST AI RMF, EU AI Act) | ISO/IEC 42001 (Policy integration), COBIT 2019 (Strategic alignment) |
GOV-011 | All strategic AI decisions are documented with clear accountability, rationale, and stakeholder sign-off | ISO/IEC 42001 (Decision accountability), NIST AI RMF (Govern – Documentation) |
GOV-012 | AI governance KPIs (e.g., number of incidents, completion of risk assessments) are reviewed at executive level | ISO/IEC 42001 (Governance KPIs), NIST AI RMF (Measurement and monitoring) |
Category
AI Legal & Regulatory Compliance
ID | Description | Standards Mapping |
|---|---|---|
LEG-001 | A comprehensive legal review process is conducted for all AI system deployments, covering liability, IP ownership, data rights, and regulatory exposure. | ISO/IEC 42001 (Clause 6.1.1 - Legal compliance), ABA Guidelines (AI Risk & Responsibility) |
LEG-002 | A regulatory compliance map identifies all AI-related laws, directives, and standards (e.g., EU AI Act, GDPR, CCPA) applicable across operating jurisdictions. | NIST AI RMF (Govern - Legal obligations), EU AI Act (Annex III risk categories) |
LEG-003 | All contracts involving AI include specific clauses covering training data usage, model ownership, liability allocation, and third-party audit rights. | ISO/IEC 42001 (Supplier agreements), EU AI Act (Article 28 - Obligations for deployers) |
LEG-004 | An AI compliance monitoring program is in place to track adherence to applicable AI-related laws and policies, with results reported to senior leadership. | ISO 37301 (Compliance management systems), ISO/IEC 42001 (Clause 9.1 - Evaluation) |
LEG-005 | Legal hold and preservation procedures are implemented for AI-related datasets, model artifacts, and audit logs in the event of legal or regulatory inquiry. | E-Discovery Standards (Sedona Principles), ISO 27001 (Annex A.12.3 - Backup) |
LEG-006 | Cross-border transfer of AI-related personal data complies with international data protection frameworks (e.g., SCCs, UK IDTA, BCRs). | GDPR (Chapter V - International transfers), SCCs, ISO/IEC 42001 (Clause 8.3) |
LEG-007 | A regulatory change management process ensures that updates to AI laws or policies are identified, assessed for impact, and implemented in a timely manner. | ISO 37301 (Clause 9.3 - Compliance updates), EU AI Act (Regulatory change adaptation) |
LEG-008 | AI systems are reviewed to determine whether they meet the definition of "high-risk AI" under applicable legal frameworks (e.g., EU AI Act). | EU AI Act (Title III - High-risk classification), NIST AI RMF (Map Function) |
LEG-009 | For each AI deployment, Data Processing Agreements (DPAs) are reviewed to ensure lawful basis, data protection responsibilities, and AI-specific clauses. | GDPR (Articles 28-29), ISO/IEC 42001 (Clause 8.2 - Third-party agreements) |
LEG-010 | Legal teams validate whether AI-generated outputs may constitute copyrighted or sensitive material subject to reuse or redistribution restrictions. | EU AI Act (Article 10 - Data governance), OECD AI Principles (Accountability), ISO/IEC 42001 (Clause 6.1.1) |
Category
AI Data Governance & Management
ID | Description | Standards Mapping |
|---|---|---|
DATA-001 | The AI data governance framework defines ownership, data stewardship, quality standards, and lifecycle management for all AI training and inference data. | ISO/IEC 42001 (Clause 6.1.1 - Data governance), DAMA-DMBOK (Data ownership & stewardship) |
DATA-002 | AI dataset lineage and provenance are documented across all stages — from original data source through preprocessing to model deployment. | ISO/IEC 5259-1 (Data provenance), NIST AI RMF (Data management practices) |
DATA-003 | All datasets used in AI systems undergo quality controls, including validation rules, completeness checks, and accuracy verification prior to use. | ISO/IEC 25012 (Data quality characteristics), ISO/IEC 42001 (Clause 8.1 - Data quality) |
DATA-004 | Consent management procedures are applied to AI datasets, ensuring lawful basis for data processing with full audit trails linking consent to usage. | GDPR (Articles 6-7 - Lawful basis & consent), ISO/IEC 29184 (Online consent) |
DATA-005 | Data retention and disposal policies specific to AI datasets align with legal, contractual, and ethical requirements, and are regularly reviewed. | ISO 15489 (Records management), ISO/IEC 42001 (Clause 8.3 - Data lifecycle) |
DATA-006 | A Master Data Management (MDM) process ensures consistency and reliability of shared datasets across all AI systems and business applications. | MDM Best Practices, DAMA-DMBOK (Master data domains), ISO/IEC 11179 |
DATA-007 | A data catalog documents all AI training and evaluation datasets with metadata including sensitivity classification, use purpose, and access restrictions. | ISO/IEC 11179 (Metadata registry), ISO/IEC 42001 (Clause 6.1.2 - Data cataloguing) |
DATA-008 | Anonymization and pseudonymization techniques used for AI training or inference are validated, documented, and periodically re-evaluated for re-identification risk. | ISO/IEC 20889 (Privacy engineering), ISO/IEC 27559 (Anonymization & pseudonymization) |
DATA-009 | Data used in high-risk AI systems undergo bias and representativeness analysis before training to ensure demographic balance and mitigate discriminatory outcomes. | EU AI Act (Annex III - Bias risk), ISO/IEC 24029-1 (AI Bias evaluation) |
DATA-010 | Synthetic or augmented data used for AI training is clearly labeled, validated for accuracy, and assessed for risk of hallucination or misrepresentation. | NIST GenAI Guidelines (Synthetic data integrity), ISO/IEC 27090 (AI training data) |
DATA-011 | Access to AI datasets is role-based and logged, with approval workflows in place for sensitive or personal data usage. | ISO/IEC 42001 (Access control to data), ISO/IEC 27001 (Annex A.9 - Access Management) |
Category
AI Model Risk & Control
ID | Description | Standards Mapping |
|---|---|---|
RISK-001 | The organization maintains an AI model risk management framework that categorizes models by materiality and applies tiered governance and control requirements accordingly. | Federal Reserve SR 11-7 (Model Risk Management), ISO/IEC 42001 (Clause 8.4) |
RISK-002 | All AI models undergo independent validation by a function separate from development before production deployment, with validation reports formally documented. | OCC 2011-12 (Model validation), ISO/IEC 42001 (Clause 9.2 - Evaluation) |
RISK-003 | The AI model inventory includes detailed documentation of each model’s assumptions, limitations, intended use, and data dependencies. | NIST AI RMF (Govern - Documentation), ISO/IEC 42001 (Clause 6.1.2 - Model inventory) |
RISK-004 | A formal AI model monitoring program detects performance degradation, drift, or anomalous behavior, and includes escalation and remediation protocols. | ISO/IEC 24028 (Monitoring performance of AI), NIST AI RMF (Measure Function) |
RISK-005 | The organization defines and documents a model risk appetite statement specifying acceptable risk levels and tolerance thresholds for AI system performance. | ISO 31000 (Risk appetite definition), COSO ERM Framework, ISO/IEC 42001 (Clause 6.1.1) |
RISK-006 | Periodic scenario analysis and stress testing is conducted to evaluate AI model behavior under adverse, rare, or boundary-case conditions. | Basel III Guidelines (Stress testing), ECB Guide on AI Models, ISO/IEC 24029-1 |
RISK-007 | A model change control process is enforced to ensure all modifications to AI models are tested, reviewed, approved, and version-controlled prior to release. | ISO/IEC 42001 (Change management), SR 11-7 (Model change control), DevOps for AI (ModelOps) |
RISK-008 | Each AI model is assigned a formal model risk rating (e.g., high, medium, low) based on complexity, autonomy, impact, and regulatory exposure. | NIST AI RMF (Risk categorization), ISO/IEC 42001 (Clause 6.1.3), FCA AI Risk Categorization |
RISK-009 | The organization maintains an incident log for AI model failures, misclassifications, or near-miss events, reviewed by risk committees. | ISO/IEC 42001 (Clause 9.1 - Incident tracking), EBA Guidelines on ICT and Security Risk |
RISK-010 | Use of pretrained or third-party AI models is subject to risk assessment, including evaluation of origin, license, and robustness testing. | EU AI Act (Article 28-29), NIST AI RMF (Third-party model risk), ISO/IEC 42001 (Third-party model usage) |
Category
AI System Security & Access Control
ID | Description | Standards Mapping |
|---|---|---|
SEC-001 | The organization implements a defense-in-depth architecture for AI systems, including layered controls across model storage, APIs, data pipelines, and endpoints. | NIST SP 800-207 (Zero Trust Architecture), ISO/IEC 27001 (Annex A.13) |
SEC-002 | Access to AI models, datasets, and configuration tools is restricted via role-based controls that enforce least privilege and separation of duties. | ISO/IEC 27001 (Annex A.9 - Access Control), NIST SP 800-53 (AC-2) |
SEC-003 | Encryption standards (e.g., AES-256, TLS 1.3) are applied to AI model artifacts, training datasets, and inference APIs both in transit and at rest. | NIST SP 800-175B (Cryptographic Standards), ISO/IEC 27001 (Annex A.10) |
SEC-004 | AI infrastructure includes security monitoring and logging, with threat detection, incident response, and forensic traceability for AI-related components. | SOC 2 (Security & Monitoring), ISO/IEC 27002 (SIEM integration), NIST SP 800-137 |
SEC-005 | A vulnerability management program addresses AI-specific threats such as adversarial inputs, model inversion, and poisoning attacks. | MITRE ATLAS (Adversarial threats to AI), ISO/IEC 27001 (A.12.6), NIST AI RMF (Manage) |
SEC-006 | The secure development lifecycle (SDLC) for AI integrates security requirements at each stage — including data ingestion, model training, and deployment. | OWASP SAMM & Secure SDLC, ISO/IEC 27034 (Application Security) |
SEC-007 | External security assessments or penetration tests are conducted periodically on AI systems to evaluate control effectiveness and identify exposure gaps. | ISO/IEC 27001 (A.18.2 - Audit & Testing), NIST SP 800-115 (Penetration Testing Guide) |
SEC-008 | Model access logs are maintained for all production AI systems, capturing user, purpose, and timestamp to support incident investigation and compliance audits. | ISO/IEC 42001 (Logging AI system access), NIST SP 800-92 (Log Management) |
SEC-009 | AI inference endpoints exposed via APIs are protected with rate limiting, authentication, and input validation against adversarial manipulation. | OWASP Top 10 for LLMs (Injection & exposure risks), NIST SP 800-204A (API security) |
SEC-010 | A model tampering detection mechanism is implemented to verify the integrity of deployed AI models via hashing or secure attestation. | ISO/IEC 27036-4 (Tamper protection), NIST AI RMF (Govern/Map - Model Integrity) |
Category
AI Accountability & Incident Response
ID | Description | Standards Mapping |
|---|---|---|
ACC-001 | A formal AI accountability matrix defines roles and responsibilities across development, validation, deployment, and monitoring of AI systems. | ISO/IEC 42001 (Clause 5.3 - Roles and responsibilities), RACI Models, OECD AI Principle 2.4 |
ACC-002 | Incident response procedures are established for AI-related events (e.g., misclassification, bias, failure), including triage, root cause analysis, and remediation. | ISO/IEC 27035 (Information security incident management), ITIL v4 (Incident Response) |
ACC-003 | The organization’s insurance coverage includes AI-specific liabilities such as algorithmic errors, harm from misuse, and model failure risks. | Cyber Insurance Guidelines (NAIC, NIST), ISO 31000 (Risk transfer mechanisms) |
ACC-004 | Service Level Agreements (SLAs) for AI services include performance expectations, reliability thresholds, and remedies for model underperformance or errors. | ISO/IEC 20000-1 (Service delivery agreements), SLA Best Practices, COBIT 2019 (BAI09) |
ACC-005 | Immutable audit trails record significant actions taken during AI system design, deployment, inference, and update phases for traceability and accountability. | ISO/IEC 42001 (Clause 9.1 - Logging), SOC 2 (Audit trails), ISO 27001 (Annex A.12.4) |
ACC-006 | AI-related incidents are classified by severity and impact on stakeholders, and major events are reported to executive leadership and regulators if applicable. | NIST AI RMF (Manage - Incident response), ISO/IEC 27035 (Incident classification) |
ACC-007 | A responsible AI policy is in place and acknowledged by all teams involved in the AI lifecycle, outlining acceptable use, legal obligations, and ethical commitments. | OECD AI Principles (Accountability), ISO/IEC 42001 (Clause 5.2 - Governance commitment) |
ACC-008 | A designated AI incident coordinator is assigned to manage high-impact events involving AI systems, ensuring cross-functional resolution and reporting. | ISO/IEC 27035 (Incident roles), NIST AI RMF (Govern - Assigned roles), COBIT 2019 (EDM05) |
Category
AI Transparency & Disclosure
ID | Description | Standards Mapping |
|---|---|---|
TRAN-001 | Each deployed AI model is accompanied by documentation that describes its intended purpose, capabilities, known limitations, and appropriate use scenarios. | ISO/IEC 42001 (Clause 8.1.3), NIST AI RMF (Map Function) |
TRAN-002 | A formal stakeholder communication plan ensures that AI usage is disclosed clearly to users, employees, and partners, including potential impacts or risks. | OECD AI Principles (Transparency and Awareness), ISO/IEC 42001 (Clause 5.2) |
TRAN-003 | The organization publishes public reports or transparency statements disclosing key AI systems, governance practices, and performance evaluation metrics. | ESG Standards (SASB, GRI), ISO/IEC 42001 (Public transparency), EU AI Act (Article 60) |
TRAN-004 | Decision-making logic for high-impact AI systems is documented in a way that is understandable to affected stakeholders, including rationale and key inputs. | NIST Explainability Guidelines, EU AI Act (Annex III), ISO/IEC 42001 (Clause 8.1) |
TRAN-005 | Feedback and grievance mechanisms allow individuals or groups to raise concerns about AI system behavior and receive timely, documented responses. | OECD AI Principles (Stakeholder Feedback), ISO/IEC 42001 (Clause 9.1 - Continuous Improvement) |
TRAN-006 | External-facing AI systems (e.g., chatbots, recommendation engines) clearly disclose when users are interacting with an AI rather than a human. | EU AI Act (Article 52 - Transparency obligations), ISO/IEC 42001 (Clause 6.1.2) |
TRAN-007 | High-risk AI deployments include plain-language summaries of model functionality, risks, and safeguards, made available to impacted individuals. | ISO/IEC 42001 (Plain-language requirements), EU AI Act (Annex IV), Garante AI Guidelines |
TRAN-008 | A review process is established to verify the accuracy and clarity of AI disclosures before public release or stakeholder communication. | ISO/IEC 42001 (Clause 9.2 - Public communications), NIST AI RMF (Govern - Transparency controls) |
Category
Human Oversight of AI Systems
ID | Description | Standards Mapping |
|---|---|---|
HUM-001 | Human-in-the-loop (HITL) or human-on-the-loop mechanisms are implemented for all AI systems involved in high-impact or safety-critical decision-making. | EU AI Act (Article 14 - Human Oversight), ISO/IEC 42001 (Clause 6.1.1), NIST AI RMF (Manage - Oversight) |
HUM-002 | Override and fail-safe mechanisms are in place to allow authorized personnel to intervene, suspend, or reverse AI-driven processes when required. | ISO/IEC 42001 (Clause 8.2 - Safety & control mechanisms), IEEE 7009 (Fail-safe design) |
HUM-003 | Personnel responsible for AI oversight receive training on AI system behavior, limitations, known failure modes, and escalation protocols. | ISO/IEC 42001 (Clause 5.3 - Roles & competencies), OECD AI Principles (Human capacity building) |
HUM-004 | Escalation and review criteria are defined for when human intervention is mandatory, based on risk level, context, or AI confidence thresholds. | ISO 31000 (Escalation & risk response), ISO/IEC 27035 (Incident response) |
HUM-005 | The effectiveness of human oversight activities (e.g., intervention accuracy, latency, override frequency) is continuously monitored and reported. | ISO/IEC 42001 (Clause 9.1 - Monitoring oversight effectiveness), NIST AI RMF (Measure) |
HUM-006 | Documentation of human oversight responsibilities is maintained, including assigned reviewers, review logs, and decision outcomes. | ISO/IEC 42001 (Clause 6.1.2 - Oversight documentation), NIST AI RMF (Govern - Roles) |
HUM-007 | AI systems that enable or automate decisions affecting rights, safety, or access (e.g., hiring, finance, healthcare) undergo mandatory human review checkpoints. | EU AI Act (Title III - High-risk systems), ISO/IEC 42001 (Clause 8.1 - Operational Controls) |
HUM-008 | Human reviewers are empowered to challenge or veto AI outputs without penalty, and escalation paths exist for disagreement with AI-based recommendations. | OECD AI Principles (Accountability & Redress), ISO/IEC 42001 (Human override mechanisms) |
Category
Bias, Fairness & Non-Discrimination in AI
ID | Description | Standards Mapping |
|---|---|---|
BIAS-001 | A formal bias assessment methodology is applied to all AI models, measuring outputs across protected characteristics (e.g., gender, race, age) during testing. | EU AI Act (Article 14 - Human Oversight), ISO/IEC 42001 (Clause 6.1.1), NIST AI RMF (Manage - Oversight) |
BIAS-002 | Fairness criteria are defined in advance (e.g., statistical parity, equal opportunity) and integrated into the AI model development and evaluation lifecycle. | ISO/IEC 42001 (Clause 8.2 - Safety & control mechanisms), IEEE 7009 (Fail-safe design) |
BIAS-003 | Regular algorithmic fairness audits are conducted to detect and document disparate impacts across population groups in real-world usage. | ISO/IEC 42001 (Clause 5.3 - Roles & competencies), OECD AI Principles (Human capacity building) |
BIAS-004 | Identified bias issues in AI models are addressed through remediation strategies (e.g., reweighting, counterfactual testing), with results documented and validated. | ISO 31000 (Escalation & risk response), ISO/IEC 27035 (Incident response) |
BIAS-005 | The AI development and review process involves cross-functional teams with diverse demographic, disciplinary, and ethical backgrounds. | ISO/IEC 42001 (Clause 9.1 - Monitoring oversight effectiveness), NIST AI RMF (Measure) |
BIAS-006 | All datasets used for AI training and evaluation are reviewed for representativeness and coverage across relevant population segments. | ISO/IEC 42001 (Clause 6.1.2 - Oversight documentation), NIST AI RMF (Govern - Roles) |
BIAS-007 | Audit logs of bias testing and fairness interventions are retained for each high-impact model and reviewed during governance cycles. | EU AI Act (Title III - High-risk systems), ISO/IEC 42001 (Clause 8.1 - Operational Controls) |
BIAS-008 | The organization defines acceptable fairness thresholdsand documents rationale for selected metrics, trade-offs, and decisions. | OECD AI Principles (Accountability & Redress), ISO/IEC 42001 (Human override mechanisms) |
Category
AI Privacy & Data Confidentiality
ID | Description | Standards Mapping |
|---|---|---|
PRIV-001 | Privacy-by-design principles are applied throughout the AI system lifecycle, from dataset selection to model deployment, ensuring data minimization and purpose limitation. | GDPR Article 25 (Privacy by Design), ISO/IEC 29134, ISO/IEC 42001 (Clause 6.1.1) |
PRIV-002 | Data Protection Impact Assessments (DPIAs) are conducted for all high-risk AI systems to evaluate privacy risks and define mitigation controls before processing begins. | GDPR Article 35 (DPIA), ISO/IEC 29134 (Privacy Impact Assessment), NIST Privacy Framework (Assess) |
PRIV-003 | AI systems implement technical privacy safeguards, including access controls, encryption, and audit logging for training and inference data pipelines. | ISO/IEC 27701 (Annex A.10 - Privacy controls), ISO/IEC 42001 (Clause 8.1.2) |
PRIV-004 | Privacy notices and consent statements transparently explain how personal data is used in AI systems, including automated decision-making logic where applicable. | GDPR Articles 12-13 (Transparent notice), ISO/IEC 29184 (Online Privacy Notices) |
PRIV-005 | Operational processes are in place to support individual data rights, including access, correction, objection, and deletion for data used in AI systems. | GDPR Chapter 3 (Data Subject Rights), NIST Privacy Framework (Control-P) |
PRIV-006 | All datasets used in AI systems are classified by sensitivity level, with documented handling rules based on data subject rights, reidentification risk, and legal basis. | ISO/IEC 27701 (PII Classification), ISO/IEC 5259-1, ISO/IEC 42001 (Clause 6.1.2) |
PRIV-007 | Model training and evaluation processes are reviewed to ensure personal data is not exposed through memorization, inversion, or reconstruction vulnerabilities. | ISO/IEC 42001 (Clause 8.1.1), EU AI Act (Article 10 - Data Governance), NIST AI RMF (Manage - Data) |
PRIV-008 | Records of processing activities (RoPAs) for AI systems involving personal data are maintained in compliance with Article 30 of GDPR, and available upon request. | GDPR Article 30 (RoPA), ISO/IEC 27701 (Record-Keeping), EDPB Guidelines (Accountability Principle) |
Category
AI Explainability & Interpretability
ID | Description | Standards Mapping |
|---|---|---|
EXPL-001 | Explainability requirements are defined per AI system based on risk level, affected rights, and the needs of different internal and external stakeholders. | ISO/IEC 42001 (Clause 6.1.1 - Explainability), EU AI Act (Article 13), NIST AI RMF (Map) |
EXPL-002 | Technical explainability methods (e.g., SHAP, LIME, feature importance) are applied to provide insight into model behavior and decision logic. | SHAP, LIME, Anchors (XAI Techniques), ISO/IEC 42001 (Clause 8.1.2) |
EXPL-003 | Explanations are adapted to audience needs, ensuring regulators, business users, and impacted individuals receive information appropriate to their literacy level. | OECD AI Principles (Transparency), Google PAIR (User-Centered Explainability) |
EXPL-004 | Validation of explainability outputs is conducted to ensure that explanations correctly reflect underlying model logic and are not misleading or overly simplified. | NIST AI RMF (Measure - Explanation quality), Fidelity Metrics, ISO/IEC 24028 |
EXPL-005 | AI documentation includes both global model summaries and individual decision explanations to enable traceability and external accountability. | ISO/IEC 42001 (Documentation), DARPA XAI Program (Global & local explanations) |
EXPL-006 | Explainability limitations are documented for each model, including types of decisions that cannot be fully explained or methods that may obscure logic. | EU AI Act (Annex III - High-risk requirements), ISO/IEC 42001 (Clause 8.1) |
EXPL-007 | For high-risk use cases, explanations are user-tested for clarity and usability, with results documented and fed into iterative improvement. | EU AI Act (Article 13 - Human Understanding), NIST AI RMF (Govern - Communication) |
EXPL-008 | Where models are updated or retrained, versioned explanation outputs are maintained to track changes in interpretability over time. | ISO/IEC 42001 (Model versioning & audit), NIST AI RMF (Manage - Model lifecycle) |
Category
AI Testing, Validation & Evaluation
ID | Description | Standards Mapping |
|---|---|---|
TEST-001 | A documented AI-specific testing strategy is implemented, covering functional correctness, model accuracy, adversarial robustness, and cybersecurity resilience. | ISO/IEC 29119 (Test strategy), ISO/IEC 42001 (Clause 8.2 - Testing AI systems) |
TEST-002 | Test datasets used during model validation are reviewed for representativeness, bias, data subject rights compliance, and alignment with the intended deployment context. | ISO/IEC 5259-1 (Test dataset documentation), NIST SP 1270 (Synthetic test data practices) |
TEST-003 | Automated testing and validation pipelines are integrated into the AI development lifecycle to ensure ongoing checks across data preprocessing, training, and inference. | CI/CD Guidelines (MLOps), ISO/IEC 42001 (Automated testing pipelines), NIST AI RMF (Manage) |
TEST-004 | User acceptance testing (UAT) is conducted before deployment, verifying that AI system outputs meet documented business requirements and are understood by end users. | ISO/IEC 25051 (Acceptance Testing), EU AI Act (User validation for high-risk systems) |
TEST-005 | AI models are subjected to performance benchmarking against pre-defined technical baselines, thresholds, and target KPIs prior to production release. | ISO/IEC 25010 (Performance attributes), NIST AI RMF (Measure - Evaluation) |
TEST-006 | Regression testing is conducted after each model update or retraining to ensure previously validated functionality and fairness is not adversely affected. | ISO/IEC 29119-2 (Regression Testing), ISO/IEC 42001 (Change impact evaluation) |
TEST-007 | For high-risk AI systems, stress testing or adversarial scenario analysis is performed to evaluate system behavior under edge cases or hostile inputs. | EU AI Act (Article 15 - Robustness testing), ISO/IEC 24029-1 (Adversarial testing methods) |
TEST-008 | Validation artifacts (e.g., test cases, test logs, evaluation results) are version-controlled and stored to support audit, incident investigation, or certification. | ISO/IEC 42001 (Clause 9.2 - Validation records), NIST AI RMF (Govern - Documentation) |
Category
AI System Documentation & Traceability
ID | Description | Standards Mapping |
|---|---|---|
DOC-001 | Standardized documentation templates define required content, formatting, and update cadence for all AI system artifacts across the lifecycle. | ISO/IEC 26514 (Documentation templates), ISO/IEC 42001 (Clause 6.1.1 - Documentation structure) |
DOC-002 | Technical documentation includes details on AI model architecture, data flows, input-output schema, and integration with external systems or APIs. | ISO/IEC 42001 (Clause 6.1.2 - Technical details), EU AI Act (Article 11 - Documentation content) |
DOC-003 | Operational documentation for AI systems includes deployment playbooks, performance monitoring procedures, issue escalation, and rollback steps. | ISO 10013 (Operational documentation), ISO/IEC 27001 (Operational procedures) |
DOC-004 | Version control is implemented for all AI documentation, with change histories, reviewer comments, and approval workflows retained for audit traceability. | ISO 9001 (Document control), GitOps Principles, ISO/IEC 42001 (Version tracking) |
DOC-005 | A centralized, searchable knowledge repository provides role-based access to all current AI-related documentation, logs, validations, and model metadata. | ISO/IEC 42001 (Clause 6.1.3 - Knowledge management), ITIL Knowledge Management |
DOC-006 | Documentation for high-risk AI systems includes risk assessments, training configuration parameters, and expected limitations as required under applicable regulations. | EU AI Act (Article 11 - High-risk system documentation), ISO/IEC 42001 (Clause 8.2) |
DOC-007 | Audit trail of documentation changes is retained for each AI model version, linking decisions to corresponding governance or validation actions. | ISO/IEC 27001 (Annex A.12.4 - Audit logs), ISO/IEC 42001 (Traceability requirements) |
DOC-008 | AI documentation is periodically reviewed and updated based on changes to model logic, regulatory environment, or organizational policy. | ISO/IEC 42001 (Clause 9.2 - Review frequency), NIST AI RMF (Govern - Continuous Improvement) |
Category
AI Training & Awareness Programs
ID | Description | Standards Mapping |
|---|---|---|
TRAIN-001 | AI literacy training is provided to all employees, covering basic AI concepts, ethical considerations, and organizational AI use cases. | OECD AI Principles (AI education), ISO/IEC 42001 (Clause 7.2 - General AI literacy) |
TRAIN-002 | Role-based training programs address specific responsibilities across AI governance roles (e.g., developers, reviewers, business owners, and compliance teams). | ISO/IEC 42001 (Clause 7.2 - Role-specific training), NIST AI RMF (Govern - Training) |
TRAIN-003 | A continuous learning program ensures AI-relevant staff stay up to date with emerging technologies, regulatory changes, and risk mitigation practices. | Continuous Learning Standards (ISO 10015), ISO/IEC 42001 (Skill refresh), HR L&D Programs |
TRAIN-004 | Training effectiveness is assessed through knowledge checks, scenario-based exercises, or documented application of concepts in assigned roles. | Kirkpatrick Model (Training evaluation), ISO 30422 (Workforce metrics), NIST AI RMF (Measure) |
TRAIN-005 | External certifications and expert-led programs are used to supplement internal AI training, especially for high-impact or sensitive AI domains. | Professional Certifications (ISO/IEC 17024), External courses (Coursera, IEEE, etc.) |
TRAIN-006 | Training completion records are tracked and reviewed as part of access control or role enablement for AI system interaction or development. | ISO/IEC 42001 (Training records), ISO 27001 (Clause A.7.2.2 - Awareness tracking) |
TRAIN-007 | High-risk AI roles (e.g., model validators, fairness auditors) are required to undergo certified or externally validated trainingon bias, ethics, and legal duties. | EU AI Act (Article 14 - High-risk oversight roles), ISO/IEC 42001 (Clause 5.3 - Responsibility) |
TRAIN-008 | The organization maintains a training calendar and review cycle to update AI training materials at least annually or upon regulatory or policy changes. | ISO 10015 (Training scheduling), ISO/IEC 42001 (Clause 9.1 - Updating competence plans) |
Category
Third-Party & Vendor AI Risk Management
ID | Description | Standards Mapping |
|---|---|---|
VEND-001 | Third-party and vendor selection includes formal assessment criteria for AI-related competencies, model transparency, training data governance, and regulatory alignment. | ISO/IEC 27036-1 & 27036-4 (Supplier security), NIST AI RMF (Govern - Third-party) |
VEND-002 | Contractual agreements with AI vendors explicitly address model ownership, intellectual property rights, liability allocation, service-level expectations, and audit access. | EU AI Act (Article 28 - Provider obligations), ISO/IEC 42001 (Clause 8.2 - Supplier agreements) |
VEND-003 | An AI vendor monitoring program tracks key performance indicators, risk metrics, and incident reports throughout the lifecycle of outsourced AI systems. | COBIT 2019 (MEA03 - Vendor monitoring), ISO/IEC 27036 (Ongoing assessment), NIST AI RMF (Manage) |
VEND-004 | Vendor risk assessments include dependency mapping and concentration risk analysis for externally sourced AI capabilities and data infrastructure. | ISO/IEC 27036 (Risk dependencies), ENISA AI Threat Landscape (Supply Chain), ISO 31000 |
VEND-005 | Documented vendor exit strategies are in place for AI-related services, including transition plans, knowledge transfer, and access to models or datasets upon contract end. | ISO 22301 (Business continuity), ISO/IEC 42001 (Clause 8.4 - Transition), ITIL Exit Strategies |
VEND-006 | Due diligence processes require validation of vendor adherence to recognized AI risk frameworks (e.g., ISO 42001, NIST AI RMF) before onboarding. | ISO/IEC 42001 (Clause 6.1.1 - Due diligence), SOC 2 (Vendor evaluation controls) |
VEND-007 | Vendors providing AI systems must supply technical documentation, bias testing results, and explainability methods for high-risk use cases. | EU AI Act (Article 11 - Technical documentation), ISO/IEC 42001 (Documentation for external AI) |
VEND-008 | Contingency plans and alternative vendor strategies are maintained for critical AI services to ensure service continuity during third-party disruption. | NIST SP 800-161 (Supply Chain Risk), ISO 22317 (Continuity strategy), ISO/IEC 27001 (A.15) |
Category
AI Continuity & Operational Resilience
ID | Description | Standards Mapping |
|---|---|---|
CONT-001 | Business continuity and disaster recovery plans explicitly address AI system outages, model unavailability, corrupted inference pipelines, and third-party model failures. | ISO 22301 (Business Continuity), ISO/IEC 42001 (Clause 8.4), DORA (Operational Resilience) |
CONT-002 | Continuity plans for AI systems are tested periodically, simulating model rollback, corrupted predictions, and failover to backup workflows to identify operational gaps. | ISO/IEC 27031 (Continuity testing), ISO 22398 (Exercise programs), NIST AI RMF (Recover) |
CONT-003 | Fallback procedures, such as manual overrides, alternative algorithms, or non-AI process substitutions, are defined for critical AI-supported business processes. | ISO/IEC 42001 (Clause 8.2 - Manual override), ITIL v4 (Contingency Procedures) |
CONT-004 | During AI-related service disruptions, stakeholder communication protocols ensure timely alerts, impact summaries, and escalation to governance bodies. | ISO 22320 (Crisis Communication), NIST SP 800-61 (Incident Handling) |
CONT-005 | Recovery Time Objectives (RTOs) and Recovery Point Objectives (RPOs) are defined for AI systems based on business impact assessments and interdependencies. | ISO/IEC 27031 (RTO/RPO Planning), ISO 22317 (Business Impact Analysis) |
CONT-006 | AI continuity plans include model redeployment procedures, ensuring reproducibility from version-controlled artifacts and pre-approved configurations. | ISO/IEC 42001 (AI model lifecycle), NIST AI RMF (Recover Function) |
CONT-007 | Critical AI model dependencies, such as data sources, APIs, and external model providers, are mapped and monitored for resiliency risk. | ENISA (AI Supply Chain Resilience), ISO/IEC 27001 (A.15 - Supplier Dependency) |
CONT-008 | Post-incident reviews involving AI system failures are conducted, with root cause analyses and updates to continuity documentation and governance records. | ISO 22301 (Post-Incident Review), ISO/IEC 42001 (Clause 9.2), NIST AI RMF (Learn from Incidents) |
Category
AI Monitoring, Auditing & Logging
ID | Description | Standards Mapping |
|---|---|---|
MON-001 | Continuous monitoring is implemented for AI model performance degradation, drift, inference anomalies, input data integrity, and system latency or resource failures. | ISO/IEC 42001 (Clause 9.1 - Continuous monitoring), NIST AI RMF (Measure - Monitoring) |
MON-002 | The AI audit program includes technical review of models, operational process compliance, privacy checks, and regulatory alignment assessments, conducted periodically. | ISO/IEC 19011 (Audit principles), ISO/IEC 42001 (Clause 9.2 - Audit program) |
MON-003 | Monitoring alerts for AI-related deviations (e.g., bias shift, threshold breaches, or unrecognized inputs) trigger documented investigation and corrective procedures. | ISO/IEC 27001 (Annex A.16 - Incident alerting), NIST AI RMF (Manage - Response triggers) |
MON-004 | Performance and compliance dashboards provide real-time visibility into AI operations, tailored for different stakeholder groups including risk, IT, and executives. | GRC Dashboards, ISO/IEC 42001 (Clause 9.1 - Performance evaluation), ITIL v4 (Event management) |
MON-005 | Audit outcomes related to AI governance, fairness, and data handling feed into the continuous improvement cycle with documented follow-ups and change implementation. | ISO 9001 (Corrective actions), ISO/IEC 42001 (Governance improvement) |
MON-006 | AI systems maintain immutable logging of model inputs, outputs, versions, and decisions to support explainability, audit, and forensic investigations. | ISO/IEC 27001 (Annex A.12.4 - Logging), ISO/IEC 42001 (Traceability), NIST AI RMF (Govern) |
MON-007 | Automated compliance checks are integrated into AI system pipelines to detect violations of policy, threshold breaches, or use of unauthorized datasets. | NIST AI RMF (Automated policy enforcement), ISO/IEC 42001 (Clause 8.2) |
MON-008 | A model monitoring governance review is conducted quarterly to assess monitoring scope, effectiveness, and escalation response metrics. | ISO/IEC 42001 (Governance reviews), COBIT 2019 (MEA - Monitoring and evaluation), NIST AI RMF (Govern) |
Category
AI Infrastructure & System Architecture
ID | Description | Standards Mapping |
|---|---|---|
INFRA-001 | AI-specific architecture standards define principles for scalability, reliability, fault tolerance, and maintainability of AI models, data pipelines, and services. | ISO/IEC 42001 (Clause 6.1.2 - AI architecture), TOGAF (Enterprise Architecture) |
INFRA-002 | Infrastructure is designed to support required model training and inference compute capacity, with elastic scalability and cost governance mechanisms. | ISO/IEC 42001 (Clause 8.3 - Capacity planning), NIST SP 800-160 (System Design & Resilience) |
INFRA-003 | The AI technology stack is selected based on compatibility with organizational platforms, regulatory compliance, and long-term support plans. | ISO/IEC 42001 (Clause 6.1.1 - Technology suitability), NIST AI RMF (Map - System Characteristics) |
INFRA-004 | Deployment architecture supports controlled rollout, canary testing, rollback of AI models and services, and maintains version isolation for validation. | DevOps/MLOps Practices (Blue-Green, Canary Deployments), ISO/IEC 27001 (A.12.1.2) |
INFRA-005 | Integration patterns support interoperability of AI systems with existing enterprise applications, APIs, monitoring tools, and security infrastructure. | ISO/IEC 42001 (Clause 6.1.2 - Interoperability), API Standards (REST, OpenAPI), TOGAF |
INFRA-006 | AI-specific resource provisioning policies govern access to GPUs, large memory nodes, and distributed training environments to prevent bottlenecks or overuse. | ISO/IEC 42001 (AI resource governance), Cloud Architecture Best Practices (AWS, Azure) |
INFRA-007 | Infrastructure includes isolation controls and sandboxing for testing AI models before production release. | ISO/IEC 27001 (Annex A.13.1 - Network isolation), NIST SP 800-53 (SC-7, SC-39) |
INFRA-008 | Architecture diagrams and system inventories for AI platforms are maintained and updated as part of technical documentation and governance reviews. | ISO/IEC 42001 (Clause 9.2 - System diagrams), ITIL CMDB, TOGAF Architecture Repository |
Category
AI Ethics, Sustainability & Responsible Use
ID | Description | Standards Mapping |
|---|---|---|
ETH-001 | Organizational policies require that AI design, development, and deployment decisions are explicitly aligned with documented ethical principles. | OECD AI Principles (Human-centered values), ISO/IEC 42001 (Clause 5.2 - Ethics), IEEE EAD |
ETH-002 | A structured ethics review process is conducted for all high-impact AI use cases, evaluating societal, environmental, and human rights implications. | UNESCO AI Ethics (Review governance), ISO/IEC 42001 (Clause 6.1.1), AI HLEG Ethics Guidelines |
ETH-003 | Stakeholder engagement includes consultation with affected users, domain experts, and civil society organizations during AI system design and evaluation stages. | OECD AI Principles (Stakeholder participation), EU AI Act (Recital 76), ISO 26000 (Social responsibility) |
ETH-004 | Responsible AI metrics are defined and tracked, covering fairness, human agency, sustainability, and impact on vulnerable populations. | ISO 30414 (Ethics KPIs), NIST AI RMF (Measure - Societal impact), OECD Responsible Innovation |
ETH-005 | Ethical risk management includes a process for updating principles and controls as new ethical dilemmas or societal norms emerge. | IEEE EAD (Evolving norms), ISO/IEC 42001 (Clause 9.1 - Improvement & ethics), NIST AI RMF (Manage) |
ETH-006 | The organization appoints an AI Ethics Officer or committeewith authority to review and veto deployments that do not align with ethical risk thresholds. | ISO/IEC 42001 (Clause 5.3 - Role definition), OECD AI Governance, Corporate Ethics Oversight |
ETH-007 | AI projects undergo environmental impact assessments, especially for large-scale models with significant compute resource demands. | SASB Standards (Environmental), ISO 14001 (Environmental impact), EU Green Deal for AI |
ETH-008 | Ethics training modules are integrated into the onboarding and development process for all roles involved in AI lifecycle activities. | ISO 10015 (Training), NIST AI RMF (Govern - Culture & Ethics), OECD (AI workforce skills) |
Category
AI Third-Party Governance & Risk Management
ID | Description | Standards Mapping |
|---|---|---|
TP-001 | A formal third-party AI risk policy defines criteria for selecting, onboarding, and monitoring vendors providing AI models, datasets, or services. | ISO/IEC 42001 (Clause 8.2 - AI supplier policy), NIST AI RMF (Govern - Third-party controls) |
TP-002 | Vendors are classified by AI risk exposure (e.g., high-risk models, critical APIs, data processors), and appropriate control tiers are applied accordingly. | NIST AI RMF (Map - Categorize third-party risk), ISO/IEC 27036-4 (Risk tiering) |
TP-003 | Contracts with AI vendors mandate transparency artifacts including model cards, data sources, explainability methods, and audit rights. | EU AI Act (Article 28 - Provider transparency), OECD AI Principles (Accountability) |
TP-004 | Due diligence includes review of vendors’ AI governance, data handling, bias mitigation practices, and regulatory compliance posture. | ISO/IEC 42001 (Third-party governance), ISO/IEC 27036 (Supplier practices) |
TP-005 | AI-related third-party systems must undergo security assessment including adversarial robustness and data leakage risks before integration. | ISO/IEC 27001 (Annex A.15 - Supplier review), NIST SP 800-53 (RA-3, CA-2) |
TP-006 | All critical AI vendor models are subject to input/output testing and validation to ensure consistency with enterprise fairness, accuracy, and explainability goals. | EU AI Act (Annex IV - Technical documentation), ISO/IEC 42001 (Clause 9.2 - Auditing) |
TP-007 | Model update procedures from third parties must be reviewed prior to deployment, and logs of version changes must be maintained. | ISO/IEC 42001 (Clause 8.3 - Model updates), ISO 9001 (Change control procedures) |
TP-008 | Vendors are required to notify of known incidents (e.g., hallucinations, data breach, model failure) within a predefined response timeframe. | NIST IR 8269 (Incident reporting obligations), ISO/IEC 27001 (A.16 - Incident response) |
TP-009 | Exit strategies are maintained to replace AI vendors in the event of risk violations, service disruption, or contract expiration. | ISO 22301 (Clause 8.4.3 - Transition plans), ISO/IEC 42001 (Business continuity) |
TP-010 | Periodic risk re-assessment is conducted for all AI suppliers, especially after incidents, regulation changes, or model version upgrades. | ISO/IEC 42001 (Clause 9.1 - Risk re-evaluation), NIST AI RMF (Manage - Risk reviews) |