Domain 2: AI Strategy and Governance

Study Guide: CRAGE Domain 2 – AI Strategy and Governance

Section 1: Detailed Domain Review

Module 1: AI Strategic Planning and Organizational Alignment

AI strategy establishes why an organization uses artificial intelligence, defines its operational scope, and specifies its risk appetite before any technical tools, vendors, or models are selected. Governance operationalizes this strategy by creating policy, phased implementation roadmaps, and measurable monitoring structures.

Strategic Vision and Objectives Alignment

• Foundational Strategy: An AI strategy functions like a zoning plan, establishing acceptable use cases, value goals, and risk boundaries prior to technical deployment. Governance derives directly from strategic business objectives.
• ISO/IEC 38507:2022: Provides guidance for governing bodies to enable and govern AI use effectively, efficiently, and acceptably across the organization.
• ISO/IEC 42001 AIMS Alignment: Context, interested parties, and scope are established in Clause 4; top management sets the AI policy in Clause 5.2; AI objectives are defined in Clause 6.2. Annex C details organizational objectives and potential AI risk sources.
• NIST AI Risk Management Framework (AI RMF 1.0):
  • GOVERN 1.3: Tailors risk management processes to the organization’s explicit risk tolerance.
  • MAP 1.1: Documents intended purpose and deployment context for every AI system.
  • MAP 3: Benchmarks capabilities, expected benefits, and costs to rank use cases by value versus risk.
  • MANAGE 1.1: Assesses whether a system meets its intended purpose, establishing that choosing not to deploy is a valid strategic outcome.
• EC-Council ADG (Adopt, Defend, Govern) Framework: The GOVERN pillar establishes AI risk appetite, decision rights, escalation paths, board reporting formats, investment criteria, and regulatory mappings.

The ADG 24-Month Governance Roadmap

The ADG framework structures AI governance implementation over a 24-month lifecycle across four distinct phases:

Policies, Procedures, and Exceptions

• Policy vs. Procedure: Policies define organizational requirements for AI and are approved by top management; procedures detail the step-by-step actions required to meet those policies.
• NIST AI RMF Policy Requirements: GOVERN 1.1 requires mapping legal and regulatory obligations; GOVERN 1.2 integrates trustworthy AI characteristics into policies; GOVERN 1.4 defines transparent controls; GOVERN 1.5 plans periodic policy reviews; GOVERN 1.6 mandates system inventory mechanisms; GOVERN 1.7 establishes safe decommissioning procedures.
• ADG Minimum Controls (MC-1 to MC-12): Translates policies into 12 concrete controls backed by inspectable evidence artifacts.
• Governance Exception Register: Policy deviations must be formally recorded with a documented business justification, risk acceptance sign-off, and explicit expiry date. Expired, open exceptions serve as a key tracking metric.
• Agent Authority Statements: Formal policy artifacts that define specific access rights, decision boundaries, execution limits, and escalation triggers for autonomous AI agents.

Organizational Culture and Literacy

• Culture as a Safeguard: Governance relies on workforce habit and understanding. NIST AI RMF GOVERN 4 mandates a safety-first, critical-thinking culture, requiring clear risk communication (GOVERN 4.2), incident information sharing (GOVERN 4.3), targeted risk training (GOVERN 2.2), and diverse input into risk decisions (GOVERN 3.1).
• EU AI Act Article 4 (AI Literacy): Amended by Regulation (EU) 2026/1744 (in effect since 2 February 2025). Requires AI providers and deployers to take measures to ensure a sufficient level of AI literacy among staff, while not guaranteeing a specific outcome level. The 2026 amendment strengthens the role of the European Commission and Member States in promoting literacy.
• ADG Capability Baseline: Treats workforce literacy as a binding operational constraint, establishing a standardized AI vocabulary across all organizational roles.

Performance Measurement Architecture

ADG organizes governance performance into a three-tiered measurement pyramid, supplemented by ISO/IEC 42001 check-and-improve loops:

/\
         /  \     Tier 3: Board-Level Indicators
        /    \    (Composite Risk Posture, Incident Trends, Compliance Rate)
       /------\
      /        \   Tier 2: Outcome Metrics
     /          \  (MTTD, MTTC, Expired Exceptions, Drift Rates)
    /------------\
   /              \  Tier 1: Control Effectiveness
  /                \ (% Classified Systems, Uncontrolled Changes, RACI Coverage)
 /------------------\

• Tier 1 (Control Effectiveness): Operative control counts (e.g., share of systems with completed risk classifications or RACI matrices, number of uncontrolled production changes per quarter).
• Tier 2 (Outcome Metrics): Operational performance outcomes (e.g., Mean Time to Detect [MTTD], Mean Time to Contain [MTTC] AI incidents, expired open exceptions, systems operating within approved configurations).
• Tier 3 (Board Indicators): Executive-level aggregate metrics (e.g., composite AI risk posture score, regulatory compliance coverage, exception backlog trends, vendor AI exposure).
• ISO/IEC 42001 Evaluation: Clause 9 governs monitoring and measurement (9.1), internal audit (9.2), and management review (9.3). Clause 10 mandates continual improvement through Plan-Do-Check-Act (PDCA) cycles.

Module 2: AI Governance Frameworks and Operational Models

AI governance relies on structural separation of duties, standardized frameworks (such as NIST AI RMF and ISO/IEC 42001), and formal control mechanisms throughout the AI system lifecycle.

The ADG Operating Model and Separation of Duties

The ADG framework divides AI responsibilities across three distinct operational pillars to prevent conflicts of interest:

+-----------------------------+
                      |       GOVERN PILLAR         |
                      | Oversees, Assures, Decides  |
                      +--------------+--------------+
                                     |
           +-------------------------+-------------------------+
           | Escalates Blockers                                | Escalates Unresolved Risks
           v                                                   v
+--------------------------+                       +--------------------------+
|       ADOPT PILLAR       |                       |      DEFEND PILLAR       |
| Executes and Delivers    |                       | Secures and Validates    |
+--------------------------+                       +--------------------------+

• ADOPT: Focuses on implementation and business delivery. Key roles include AI Product Owners and Prompt Designers.
• DEFEND: Focuses on security, robustness, and safety validation. Key roles include AI Red Team Leads, Guardrail Engineers, and Bias Auditors.
• GOVERN: Focuses on oversight, policy, alignment, and final approval authority. Key roles include Model Risk Officers and Privacy Counsel.
• Minimum Control MC-3 (Separation of Duties): Strictly divides deployment ownership, security validation, and final approval. Evidenced by a system-specific RACI matrix. Disagreements between Adopt and Defend escalate directly to Govern.

AI Governance Council and Decision Rights

• Council Function: A cross-functional body chartered during the Foundation phase (Months 0–3) comprising representatives from Adopt, Defend, and Govern. It serves as the ultimate escalation authority for dispute resolution and go/no-go deployment decisions.
• Decision Rights: Owns the Governance Exception Register, defines risk appetite, sets release thresholds, and monitors MC-10 governance reviews.
• Human Oversight Models:
  • Human-in-the-Loop (HITL): Human operator must review and approve every AI output prior to execution.
  • Human-on-the-Loop (HOTL): Human operator monitors operation in real time and can intervene or trigger circuit breakers.
  • Human-out-of-the-Loop (HOOTL): System operates fully autonomously, subject to post-hoc forensic audits and periodic reviews.
  • Note: Selecting higher autonomy tiers (HOTL/HOOTL) mandates the implementation of ADG Agentic Orchestration and Agentic Hardening overlays.

NIST AI RMF Core Functions and Profiles

• NIST AI RMF 1.0 (NIST AI 100-1): Voluntary framework structured around four core functions:
  • GOVERN: Cross-cutting function defining policies, accountability, culture, workforce diversity, and third-party risk management (6 categories).
  • MAP: Contextualizes system purpose, categorizes risks, maps socio-technical impacts, and evaluates component capabilities (5 categories).
  • MEASURE: Conducts Test, Evaluation, Verification, and Validation (TEVV), tracks metrics, and validates measurement efficacy (4 categories).
  • MANAGE: Allocates risk response strategies (mitigate, transfer, avoid, accept), manages incidents, and governs decommissioning (4 categories).
• Trustworthy AI Characteristics: AI systems must be valid and reliable (the foundational requirement), safe, secure and resilient, accountable and transparent (relates to all characteristics), explainable and interpretable, privacy-enhanced, and fair (harmful bias managed).
• Profiles & Companion Frameworks:
  • Profiles: Customize the core framework for specific use cases, sectors, or temporal states (comparing a Current Profile to a Target Profile for gap analysis).
  • NIST AI 600-1 (Generative AI Profile): Released 26 July 2024; details 12 novel or heightened risks associated with generative AI, including confabulation (confidently generated false outputs/hallucinations), IP infringement, and value-chain integration failures.

ISO/IEC AI Standards Architecture

Lifecycle Governance Stages

ADG governs six repeating lifecycle stages:

[Sourcing/Pre-training] ---> [Post-training/Alignment] ---> [Build/Integrate]
          ^                                                        |
          |                                                        v
   [Retire/Learn]      <---     [Run/Monitor]     <---  [Deploy/Authorize]

1. Pre-training / Sourcing: Adopt selects fit-for-purpose datasets; Defend inspects provenance abuse and data contamination; Govern validates data licensing, ethics, and jurisdictional constraints.
2. Post-training / Alignment: Defend tests for safety bypasses, security regressions, and fairness drift; Govern reviews responsible AI criteria and alignment documentation.
3. Build / Integrate: Govern classifies system risk (MC-2); Defend validates interface controls, secret management, and input sanitization.
4. Deploy / Authorize: MC-4 testing across four harm classes: Technical (security/integrity), Societal (bias/rights), Operational (reliability/accuracy), and Systemic (emergent AI-to-AI interaction risks). Final authorization is granted via the AI Release Gate Checklist.
5. Run / Monitor: Production runtime monitoring (MC-8) for configuration drift, prompt modifications, unauthorized usage, and emerging bias.
6. Retire / Learn: Decommissioning by Adopt, evidence preservation by Defend, and policy updates/lessons-learned loops by Govern.

Change Control and Modifications

• ADG Change Control (MC-5 & MC-11): All prompt updates, context alterations, retrieval source shifts, and model swaps must be recorded in a Model and Prompt Change Log and undergo fairness re-evaluation.
• EU AI Act Substantial Modifications (Article 43(4)): Any unplanned change affecting a high-risk system's compliance or intended purpose mandates a new conformity assessment. Pre-determined changes incorporated into the initial assessment during design (Recital 128) do not constitute substantial modifications.
• Role Reclassification (Article 25): Any deployer or distributor making a substantial modification to a high-risk system, or changing its intended purpose to a high-risk category, legally assumes the duties of a Provider.

Module 3: AI Compliance, Legal Frameworks, and Regulatory Requirements

Navigating global AI legal mandates requires structured regulatory mapping, risk classification, and alignment between voluntary governance frameworks and binding statutory requirements.

Global AI Regulatory Overview

+-----------------------------------+
                      |      GLOBAL AI REGULATORY MAP     |
                      +-----------------+-----------------+
                                        |
       +--------------------------------+--------------------------------+
       |                                |                                |
       v                                v                                v
[EUROPEAN UNION]                 [UNITED STATES]                   [ASIA-PACIFIC]
- EU AI Act (Binding)            - EO 14110 revoked Jan 2025      - S. Korea: AI Basic Act
- Risk-based classification      - EO 14179 / July 2025 Action    - Japan: AI Promotion Act
- Product Liability Directive      Plan (Voluntary/National)        - China: GAI Interim Rules
                                 - State laws (Colorado, NYC)       - CETS 225: Int'l Treaty

• Standard vs. Regulation: Frameworks like ISO/IEC 42001 and NIST AI RMF are voluntary guidelines; regulations like the EU AI Act (Regulation (EU) 2024/1689) are binding laws carrying severe administrative fines.
• United States: Lacks a comprehensive federal AI statute. Executive Order 14110 was revoked on 20 January 2025. Policy is guided by EO 14179 and the July 2025 AI Action Plan. Individual state acts persist, such as Colorado's re-enacted Automated Decision-Making Technology law (effective 1 January 2027) and NYC Local Law 144 (hiring bias audits). Federal preemption of state laws was urged by the White House in March 2026.
• Asia-Pacific: China regulates targeted services via the Generative AI Interim Measures (August 2023) and Content Labeling Measures (September 2025). South Korea’s AI Basic Act took effect 22 January 2026. Japan’s AI Promotion Act (May 2025) provides an innovation-first regime led by an AI Strategic Headquarters.
• International Treaties: Council of Europe Framework Convention on AI (CETS 225, opened September 2024) is the first legally binding international treaty on AI, requiring five ratifications to enter into force.

EU AI Act Risk Tiers and Statutory Obligations

Regulation (EU) 2024/1689 establishes four horizontal risk categories based on intended purpose:

[ UNACCEPTABLE RISK ] -> Banned Practices (Art. 5) -> Fines up to €35M or 7%
[    HIGH RISK     ] -> Annex I (Products) & Annex III (Use Cases) -> Full Compliance (Arts. 9-15)
[ TRANSPARENCY RISK] -> Article 50 Disclosure & Watermarking -> Fines up to €15M or 3%
[ MINIMAL / NO RISK] -> Unregulated / Code of Conduct Voluntary

1. Unacceptable Risk (Prohibited - Article 5): Banned practices include harmful cognitive manipulation, social scoring, untargeted facial image scraping, workplace emotion recognition, and AI-enabled nudification or child abuse material generation.
2. High Risk (Articles 6, 9–15, Annex I & III):Annex I: AI components embedded in safety products requiring third-party assessment under EU harmonized laws (e.g., machinery, toys, medical devices).Annex III: Eight specific critical domains: biometrics, critical infrastructure, education, employment, essential private/public services (credit, insurance), law enforcement, migration, and justice/democratic processes.Profiling Rule (Article 6(3)): An Annex III system that profiles natural persons is always classified as high-risk.

• Annex I: AI components embedded in safety products requiring third-party assessment under EU harmonized laws (e.g., machinery, toys, medical devices).
• Annex III: Eight specific critical domains: biometrics, critical infrastructure, education, employment, essential private/public services (credit, insurance), law enforcement, migration, and justice/democratic processes.
• Profiling Rule (Article 6(3)): An Annex III system that profiles natural persons is always classified as high-risk.

3. Transparency Risk (Limited - Article 50): Mandates clear notification when humans interact with AI, synthetic content machine-readable labeling/watermarking, deepfake disclosures, and emotion recognition notices.
4. Minimal / No Risk: Video games, spam filters, general utility tools. No mandatory regulatory requirements.

Implementation Timeline (Post-AI Omnibus Revisions)

Regulation (EU) 2026/1744 (the AI Omnibus, adopted 8 July 2026, in force 27 July 2026) revised several compliance deadlines originally specified in the 2024 Act:

01 Aug 2024: Entry into Force
02 Feb 2025: Prohibitions & AI Literacy Enforcement Starts
02 Aug 2025: GPAI General Rules & EU Governance Apply
02 Aug 2026: Article 50 Transparency Enforcement Starts
02 Dec 2026: New Nudification/CSAM Ban & Product Liability Rules Apply
02 Dec 2027: High-Risk Rule Enforcement for Annex III Systems
02 Aug 2028: High-Risk Rule Enforcement for Annex I Product Embedded AI
02 Aug 2030: High-Risk Deadline for Public Authority Legacy Systems (Art. 111)

General-Purpose AI (GPAI) Model Provisions

• Basic GPAI Provider Duties (Article 53): Maintain technical documentation, supply integration data to downstream deployers, establish a copyright policy complying with Directive (EU) 2019/790 Article 4(3) text-and-data-mining opt-outs, and publish a comprehensive training content summary using the EU AI Office template.
• Systemic Risk Threshold (Article 51): GPAI models trained using cumulative computational power exceeding 1025floating-point operations (FLOP) are presumed to present systemic risks.
• Systemic GPAI Duties (Article 55): In addition to Article 53 duties, systemic model providers must perform continuous adversarial red-teaming, evaluate and mitigate systemic risks, report serious security incidents to the AI Office, and ensure robust cybersecurity protection.
• Open-Source Exceptions: Open-source models with publicly available parameters skip technical documentation requirements under Article 53(1)(a) and (b), but must comply with copyright policy rules and training summaries. Open-source models with systemic risk receive no exemptions.
• Enforcement: Oversight is conducted by the EU AI Office. Fines of up to €15 million or 3% of global annual turnover apply starting 2 August 2026 under Article 101.

High-Risk Requirements: Providers vs. Deployers

AI Liability Frameworks and Data Protection Integration

• EU Product Liability Directive (EU) 2024/2853: Takes effect 9 December 2026 (replacing Directive 85/374/EEC). Explicitly classifies software and AI systems as "products." Establishes strict no-fault liability for defective AI software causing harm.
• Judicial Presumption of Defectiveness (Article 10): Courts may legally presume an AI product defective if the defendant fails to disclose relevant technical documentation or operational logs, highlighting the legal necessity of log retention.
• Withdrawal of AI Liability Directive: The separate, fault-based proposed AI Liability Directive was formally withdrawn by the European Commission in October 2025.
• GDPR Integration:
  • Article 22: Prohibits decisions based solely on automated processing (including profiling) that produce legal or similarly significant effects on individuals, unless specific exceptions apply.
  • Article 35: Requires Data Protection Impact Assessments (DPIAs) for high-risk processing. The AI Omnibus allows deployers to cross-reference existing DPIAs when fulfilling AI Act FRIA requirements.
  • Article 4a (AI Act Amendment): Permits high-risk AI providers to process special categories of personal data solely to detect and correct algorithmic bias, provided strict encryption, access limitations, and post-use data deletion safeguards are maintained.

Sector-Specific AI Compliance Rules

• Financial Services:
  • Credit scoring and life/health insurance risk pricing are Annex III high-risk use cases. Financial entities fulfill deployer operational logging and monitoring duties via existing financial services governance regulations (Article 26(5), (6)).
  • US Guidance Update: On 17 April 2026, US banking regulators issued SR 26-2, replacing SR 11-7 model risk management guidance. SR 26-2 applies to traditional ML but explicitly excludes generative and agentic AI systems.
• Healthcare: Embedded AI in medical devices requiring third-party assessments falls under Annex I high-risk requirements (enforceable 2 August 2028). The US FDA’s December 2024 guidance on Predetermined Change Control Plans (PCCPs) allows pre-approved iterative ML model modifications without requiring new marketing submissions for each change.
• Employment: AI tools used for recruitment, hiring, evaluation, or termination are Annex III high-risk systems. NYC Local Law 144 requires employers using Automated Employment Decision Tools (AEDTs) to conduct annual independent bias audits, publish summary metrics, and notify candidates 10 business days prior to use.

Section 2: Practice Quiz

Questions

1. Under the NIST AI Risk Management Framework (AI RMF 1.0), which core function wraps continuously around all other functions, and what are the specific roles of the MAP and MANAGE functions?
2. In the EC-Council ADG (Adopt, Defend, Govern) framework, what specific operational roles and responsibilities are assigned to the DEFEND pillar, and how does it interact with the GOVERN pillar?
3. What core activities, deliverables, and Minimum Controls are executed during the Foundation Phase (Months 0 to 3) of the ADG 24-month governance roadmap?
4. How does the EU AI Act (Regulation (EU) 2024/1689) differentiate between an AI Provider and an AI Deployer, and under what conditions does a deployer legally assume the duties of a provider?
5. Under the EU AI Act, what computational threshold triggers the presumption that a General-Purpose AI (GPAI) model possesses high-impact capabilities presenting systemic risk, and what extra obligations apply to its provider?
6. Following the adoption of Regulation (EU) 2026/1744 (the AI Omnibus), what are the revised enforcement start dates for Annex III high-risk AI systems and Annex I regulated product-embedded AI systems?
7. What are the specific reporting timelines mandated by Article 73 of the EU AI Act for high-risk AI providers reporting serious incidents to market surveillance authorities?
8. How does the EU Product Liability Directive (Directive (EU) 2024/2853) treat software and AI systems, and how does Article 10 affect the legal burden of proof regarding technical logs?
9. In US banking regulation, what update occurred on 17 April 2026 regarding model risk management guidance, and what specific types of AI does this revised guidance explicitly exclude?
10. What is the distinction between ISO/IEC 42001 and ISO/IEC 42005 within the ISO/IEC AI standards family regarding certification and operational scope?

Section 3: Answer Key

1. Answer: The GOVERN function is cross-cutting and infused throughout the entire lifecycle, setting policies, accountability, risk tolerance, and governance structures. The MAP function establishes context, categorizes the AI system, and identifies benefits, costs, and socio-technical risks. The MANAGE function prioritizes and executes risk response strategies (mitigation, avoidance, transfer, or acceptance) and manages post-deployment incident responses and system decommissioning.
2. Answer: The DEFEND pillar is responsible for securing, testing, and validating AI systems, utilizing roles such as AI Red Team Leads, Guardrail Engineers, and Bias Auditors. It actively conducts adversarial testing, safety checks, and fairness evaluations to try to break systems before release. When unresolved security or safety risks are identified by DEFEND that conflict with ADOPT's delivery goals, DEFEND escalates the impasse directly to GOVERN for binding dispute resolution.
3. Answer: The Foundation Phase focuses on establishing visibility, baseline control, and oversight structures within the first three months of governance implementation. Core deliverables include compiling a comprehensive AI system inventory, naming accountable owners, classifying criticality tiers, establishing workforce literacy baselines, and chartering the cross-functional AI Governance Council. This phase specifically implements ADG Minimum Controls MC-1 (Inventory & Categorization), MC-2 (Risk Classification), and MC-3 (Separation of Duties).
4. Answer: A Provider develops an AI system or GPAI model (or has one developed) and places it on the market under its own name or trademark, whereas a Deployer operates an AI system under its authority for professional activities. A Deployer legally becomes a Provider under Article 25 if it places its own trademark on a high-risk system, makes a substantial modification to an existing high-risk system, or modifies a lower-risk system's intended purpose such that it falls into a high-risk category.
5. Answer: A GPAI model is presumed to present systemic risk when its training used a total computational volume exceeding 1025 floating-point operations (FLOP). Providers of systemic models must conduct continuous adversarial model evaluations, assess and mitigate systemic risks, report serious incidents to the EU AI Office, and enforce heightened cybersecurity protections. These requirements apply in addition to baseline Article 53 GPAI obligations such as maintaining technical documentation and publishing training content summaries.
6. Answer: Regulation (EU) 2026/1744 (the AI Omnibus) deferred the application dates for high-risk AI systems past the original August 2026 milestone. High-risk AI systems categorized under Annex III (such as biometrics, employment, credit, and law enforcement) must comply by 2 December 2027. High-risk systems embedded as safety components in Annex I regulated products (such as medical devices, machinery, and toys) must comply by 2 August 2028.
7. Answer: High-risk AI providers must report any serious incident to market surveillance authorities within a baseline maximum of 15 days after becoming aware of the event. However, this deadline shrinks to 10 days if the serious incident results in a person's death or severe harm to health. If the incident involves a widespread infringement or a critical infrastructure disruption, the notification must be submitted within 2 days.
8. Answer: Directive (EU) 2024/2853 explicitly classifies software and AI systems as "products," making manufacturers strictly liable without fault for harm caused by defective AI systems starting 9 December 2026. Under Article 10, if a defendant company fails to disclose relevant technical documentation or operational logs during legal proceedings, a court may legally presume the AI product was defective, shifting the burden of proof to the defendant.
9. Answer: US federal bank regulators replaced the legacy SR 11-7 model risk management guidance with SR 26-2 on 17 April 2026. SR 26-2 updates model risk expectations for traditional statistical models and non-generative machine learning applications. However, the guidance explicitly excludes generative AI models and autonomous agentic AI architectures from its scope.
10. Answer: ISO/IEC 42001:2023 is a certifiable requirements standard that specifies how an organization establishes, implements, and continually improves an AI Management System (AIMS). In contrast, ISO/IEC 42005:2025 is a non-certifiable guidance standard that defines methodologies for conducting AI system impact assessments regarding impacts on individuals, groups, and society.

Section 4: Essay Examination Questions

1. Strategic Risk Alignment and Policy Translation: Analyze the process of converting an enterprise's stated risk appetite into operational AI policies, Minimum Controls, and measurable governance metrics. Compare how the NIST AI RMF, ISO/IEC 42001, and the ADG framework approach this lifecycle.
2. Operational Separation of Duties and Autonomy Governance: Evaluate the structural friction between business adoption (ADOPT) and security validation (DEFEND) in high-autonomy AI deployments. Discuss how an AI Governance Council uses decision rights, oversight models (HITL, HOTL, HOOTL), and Agent Authority Statements to manage risk without halting innovation.
3. Comparative Analysis of Global Regulatory Architectures: Contrast the binding, horizontal, risk-tiered approach of the EU AI Act with the decentralized, sector-specific, and state-level regulatory landscape of the United States. Assess the operational challenges faced by multinational entities attempting to construct a unified global compliance engine.
4. Lifecycle Governance and Change Control Dynamics: Examine the legal, governance, and technical implications of modifying a deployed machine learning model. Detail the criteria that distinguish routine operational maintenance from a "substantial modification" under the EU AI Act, and describe the corresponding governance workflows required under ADG Minimum Control MC-5.
5. Data Protection, Rights Enforcement, and Strict Liability: Synthesize the legal interlocks between GDPR Article 22, the EU AI Act's Fundamental Rights Impact Assessment (FRIA), and the updated Product Liability Directive (EU) 2024/2853. Explain how retention of technical evidence and automated logging serves as both a statutory requirement and a legal defense against judicial presumptions of defectiveness.

Section 5: Comprehensive Glossary

• ADG Framework (Adopt, Defend, Govern): An EC-Council AI security governance framework structuring enterprise AI management into three operational pillars over a 24-month, four-phase implementation roadmap.
• ADG Phase 1 (Foundation): The initial 0–3 month roadmap phase focused on building the AI inventory, establishing ownership, setting literacy baselines, and chartering the AI Governance Council.
• ADG Phase 2 (Control Deployment): The 3–9 month roadmap phase focused on deploying release gates, change control, context policies, runtime monitoring, adversarial red-teaming, and vendor due diligence.
• ADG Phase 3 (Agentic Readiness): The 9–15 month roadmap phase focused on implementing Agent Authority Statements, tool trust tiering, Model Context Protocol governance, circuit breakers, and forensic replay.
• ADG Phase 4 (Maturity): The 15–24 month roadmap phase focused on continuous evaluation, persistent adversarial monitoring, drift governance, responsible AI metrics, and board-level reporting.
• ADOPT Pillar: The ADG framework pillar responsible for building, integrating, and executing AI solutions to deliver business value.
• DEFEND Pillar: The ADG framework pillar responsible for testing, validating, red-teaming, and securing AI systems against vulnerability, abuse, and bias.
• GOVERN Pillar: The ADG framework pillar responsible for setting risk appetite, defining policies, granting deployment authorizations, managing exceptions, and conducting board reporting.
• Agent Authority Statement: An ADG policy artifact defining the precise boundaries of access, execution authority, decision limits, and mandatory human escalation paths for an autonomous AI agent.
• AI Governance Council: A cross-functional executive body comprising Adopt, Defend, and Govern stakeholders that acts as the final decision-making and tie-breaking authority for AI deployment and risk exceptions.
• AI Literacy: The skills, knowledge, and understanding required to enable staff and actors to make informed deployments and decisions regarding AI systems, mandated by EU AI Act Article 4.
• AI Management System (AIMS): An organization's integrated management structure defining policies, processes, and controls for AI governance, standardized under ISO/IEC 42001.
• AI Release Gate Checklist: A formal deployment artifact verifying that an AI system has passed mandatory quality, safety, security, fairness, and governance reviews before entering production.
• AI Risk Management Framework (NIST AI RMF 1.0): A voluntary guidance document published by NIST (NIST AI 100-1) structured around GOVERN, MAP, MEASURE, and MANAGE functions to manage AI socio-technical risks.
• Automated Employment Decision Tool (AEDT): AI systems used to automate or assist hiring, promotion, termination, or employment evaluation decisions, subject to bias audits under NYC Local Law 144.
• CETS 225: The Council of Europe Framework Convention on Artificial Intelligence, the first legally binding international treaty on AI, human rights, democracy, and the rule of law.
• Circuit Breaker: An automated safety control mechanism that immediately halts an AI agent's execution or revokes tool access upon detecting anomalous or out-of-bounds behavior.
• Confabulation: Confidently presented false, unsupported, or inaccurate output generated by an AI model (colloquially referred to as hallucination); identified as a key risk in NIST AI 600-1.
• Conformity Assessment: The process conducted by a provider under EU AI Act Article 43 to verify that a high-risk AI system satisfies all mandatory statutory requirements prior to market entry.
• Defectiveness Presumption: A legal rule under EU Product Liability Directive (EU) 2024/2853 Article 10 allowing courts to presume an AI system defective if the manufacturer fails to produce requested technical evidence or logs.
• Deployer: Any natural or legal person, public authority, agency, or other body using an AI system under its authority, except where used for personal, non-professional activity (EU AI Act Article 3(4)).
• EN 18286:2026: The first harmonised Quality Management System standard published in the Official Journal of the EU to support EU AI Act Article 17 compliance.
• FLOP (Floating-Point Operations): A measure of compute performance; used under EU AI Act Article 51 (1025 FLOP) to identify General-Purpose AI models with systemic risk.
• Fundamental Rights Impact Assessment (FRIA): A mandatory assessment required under EU AI Act Article 27 by public bodies and specific deployers prior to deploying high-risk AI systems to evaluate societal and individual impacts.
• General-Purpose AI (GPAI) Model: An AI model, including foundation models, trained on broad data at scale, capable of competently performing a wide range of distinct tasks, and integrable into multiple downstream applications.
• Governance Exception Register: A formal log owned by GOVERN recording all temporary policy deviations, including risk sign-offs, justifications, and hard expiration dates.
• Human-in-the-Loop (HITL): An oversight mechanism requiring direct human intervention and approval before an AI system can finalize or execute an output.
• Human-on-the-Loop (HOTL): An oversight mechanism where a human continuously monitors an active AI system and retains the technical capability to intervene or abort execution in real time.
• Human-out-of-the-Loop (HOOTL): Fully autonomous AI operation where human review occurs retrospectively through audits and forensic logs.
• ISO/IEC 23894:2023: An international standard providing specific guidance on managing AI-related organizational risks.
• ISO/IEC 38507:2022: An international standard offering governance guidance specifically tailored for members of the governing body (board of directors) regarding organizational AI usage.
• ISO/IEC 42001:2023: The international, certifiable management system standard specifying requirements for setting up, operating, and improving an AI Management System (AIMS).
• ISO/IEC 42005:2025: An international guidance standard providing methodologies for assessing the impacts of AI systems on individuals, groups, and society.
• ISO/IEC 42006:2025: An international standard setting qualification and operational requirements for accreditation bodies certifying ISO/IEC 42001 AIMS implementations.
• Mean Time to Contain (MTTC): A Tier 2 outcome metric measuring the average operational time elapsed from the initial detection of an AI incident to its containment or safe isolation.
• Mean Time to Detect (MTTD): A Tier 2 outcome metric measuring the average operational time required to identify an active AI security, bias, or operational failure.
• Model Context Protocol (MCP): A standardized framework governing how AI agents connect, exchange data context, and access external technical tools securely.
• NIST AI 600-1: The Generative AI Profile companion resource to the NIST AI RMF detailing 12 novel/heightened risks associated with generative AI deployments.
• Post-Market Monitoring System: A documented operational regime required under EU AI Act Article 72 forcing high-risk providers to collect, analyze, and report production operational data throughout a system's lifecycle.
• Predetermined Change Control Plan (PCCP): FDA-approved documentation allowing medical device providers to execute planned, iterative machine-learning modifications post-market without new marketing submissions.
• Provider: Any natural or legal person developing an AI system or GPAI model (or having one developed) and placing it on the market under its own name or trademark (EU AI Act Article 3(3)).
• RACI Matrix: A governance artifact detailing which operational roles are Responsible, Accountable, Consulted, and Informed for specific system lifecycle controls.
• Regulation (EU) 2024/1689: The EU AI Act; a comprehensive horizontal regulatory framework establishing harmonized rules, prohibitions, and compliance tiers for artificial intelligence in the European Union.
• Regulation (EU) 2026/1744 (AI Omnibus): An EU amending regulation adopted in July 2026 that revised and extended compliance implementation timelines for high-risk AI systems.
• Separation of Duties (MC-3): An ADG minimum control mandating that system development/operation (Adopt), safety validation (Defend), and final release authorization (Govern) must be assigned to distinct entities.
• SR 26-2: Federal Reserve Bank supervision guidance issued on 17 April 2026 replacing SR 11-7 for ML model risk management, explicitly excluding generative and agentic AI.
• Statement of Applicability (SoA): A formal document required by ISO/IEC 42001 Clause 6.1.3 identifying which Annex A controls have been selected or excluded, along with technical justifications.
• Substantial Modification: An unplanned change to an AI system post-deployment that alters its statutory compliance or fundamental intended purpose, triggering a new conformity assessment under EU AI Act Article 43(4).
• TEVV (Test, Evaluation, Verification, and Validation): The foundational technical evaluation workflow prescribed within the NIST AI RMF MEASURE function to establish system trustworthiness.
• Uncontrolled Production Change: A Tier 1 metric measuring the number of unauthorized modifications made to production prompts, parameters, retrieval datasets, or code bases without change control approval.