2.1 The "Three Lines of Defense" Model

The Three Lines of Defense (3LoD) model, originally developed for financial services risk management, provides a tested structure for assigning AI governance responsibilities across your organization, giving clear accountability while preventing gaps or overlaps in risk coverage.

Overview: The Three Lines

2.1.1 First Line: Development & Data Teams (Operational Risk)

The first line consists of those who build, deploy, and operate AI systems. They are the primary risk owners because they make day-to-day decisions that directly impact AI behavior and outcomes.

First Line Roles

• Data Scientists & ML Engineers: Build models, select algorithms, tune hyperparameters
• Data Engineers: Manage data pipelines, ensure data quality, maintain lineage
• Software Developers: Integrate AI into applications, implement guardrails
• Product Managers: Define use cases, set requirements, prioritize features
• DevOps/MLOps Engineers: Deploy, monitor, and maintain AI systems in production
• Business Analysts: Translate business needs, validate outcomes

First Line Responsibilities

Activity	Description	Deliverables
Risk Identification	Identify potential biases, safety issues, and ethical concerns in design	Risk register entries, preliminary risk assessment
Control Implementation	Build safeguards into systems (guardrails, validation, monitoring)	Technical controls, test results, deployment checklists
Documentation	Create and maintain model cards, system documentation, audit trails	Model cards, technical docs, change logs
Testing & Validation	Execute fairness tests, adversarial testing, performance validation	Test reports, fairness metrics, validation results
Monitoring	Track model performance, detect drift, identify incidents	Monitoring dashboards, alert configurations, incident reports
Issue Escalation	Report concerns and incidents to second line	Escalation tickets, incident notifications

First Line Controls for AI

2.1.2 Second Line: RAI Council & Risk Officers (Oversight & Policy)

The second line provides expertise, support, and challenge to the first line. They don't build AI systems but set the standards, review compliance, and monitor overall risk posture.

Second Line Roles

• Responsible AI Council/Committee: Cross-functional body setting AI ethics policy and reviewing high-risk systems
• Chief AI Officer (CAIO): Executive accountable for AI strategy and governance
• AI Risk Officers: Specialists assessing and monitoring AI-specific risks
• Data Privacy Officers: Ensuring AI compliance with privacy regulations
• Legal/Compliance: Interpreting regulations, reviewing contracts, advising on liability
• Ethics Advisors: Providing ethical guidance on complex or novel use cases

Second Line Responsibilities

Activity	Description	Deliverables
Policy Development	Create AI governance policies, standards, and procedures	AI Policy Manual, Acceptable Use Policy, Risk Taxonomy
Framework Management	Maintain and evolve the Responsible AI Framework	Framework documentation, templates, guidance
Risk Assessment Review	Review and challenge first-line risk assessments	Assessment reviews, recommendations, approvals
Monitoring & Reporting	Aggregate risk metrics, report to leadership	Risk dashboards, Board reports, trend analysis
Training & Awareness	Develop and deliver AI ethics training	Training materials, completion tracking
Regulatory Monitoring	Track evolving regulations, assess compliance gaps	Regulatory updates, compliance roadmaps

RAI Council Structure

Second Line Decision Framework

Risk Tier	Review Level	Approval Authority
Minimal	Self-assessment by first line	Model Owner
Limited	Second-line review of documentation	AI Risk Officer
High	Full RAI Council review	RAI Council vote
Prohibited	Automatic rejection	N/A (cannot proceed)

2.1.3 Third Line: Internal Audit (Independent Assurance)

The third line provides independent, objective assurance on the effectiveness of governance and risk management. They report to the Board/Audit Committee, not to management, ensuring true independence.

Third Line Roles

• Internal Audit: Traditional audit function with AI-specific capabilities
• AI Audit Specialists: Technical auditors who can assess ML systems
• External Auditors: Third-party firms for independent assessments (especially for high-risk systems)
• Regulatory Examiners: External regulators conducting compliance reviews

Third Line Responsibilities

Activity	Description	Deliverables
Framework Effectiveness	Assess whether RAI framework is properly designed and implemented	Framework audit report, gap analysis
Control Testing	Independently test first and second line controls	Control testing results, deficiency reports
Model Validation	Independent validation of high-risk AI models	Model validation reports, technical findings
Compliance Verification	Verify compliance with policies and regulations	Compliance audit reports
Thematic Reviews	Deep-dive audits on specific AI risk areas	Thematic audit reports (e.g., bias, security)
Board Reporting	Report findings directly to Board/Audit Committee	Board presentations, management letters

AI Audit Competencies Needed

Audit Focus Areas for AI

• AI inventory completeness and accuracy
• Risk classification appropriateness
• Documentation quality (model cards, impact assessments)
• Control design and operating effectiveness
• Training data governance and provenance
• Bias testing methodology and results
• Model monitoring and drift detection
• Incident response and remediation
• Vendor oversight and due diligence
• Regulatory compliance (EU AI Act, sector-specific)

Making the Model Work

Common Implementation Challenges

Challenge	Symptom	Solution
Silo Mentality	Lines don't communicate; duplicate or conflicting activities	Regular cross-line coordination meetings; shared platforms
First Line Resistance	Developers view governance as obstacle; bypass controls	Involve first line in policy development; automate where possible
Second Line Overload	Too many reviews; bottlenecks; rubber-stamping	Risk-based prioritization; delegate lower-risk reviews
Third Line Gaps	Auditors lack AI expertise; superficial reviews	Invest in training; use external specialists; develop AI audit tools
Information Asymmetry	Second/third lines only see what first line shows them	Independent data access; automated monitoring; surprise testing

Keys to Success

Implementation Steps