Chapter 14

Annex A Controls: AI System Life Cycle (A.6)

Detailed guidance on implementing Annex A controls for AI system lifecycle management (A.6), the largest control domain with 12 controls.

25 min read

Chapter Overview

This chapter covers the AI System Life Cycle domain (A.6), the largest control domain with 12 controls. These controls ensure AI systems are managed responsibly throughout their entire lifecycle from design to decommissioning.

A.6 AI System Life Cycle

This domain is divided into two sub-sections: A.6.1 (General) and A.6.2 (AI System Life Cycle Stages).

A.6.1 General Controls

A.6.1.2 Managing AI System Life Cycle

AttributeDetails
ControlProcesses shall be defined to manage the AI system life cycle.
PurposeEstablish systematic lifecycle management
Related Clause8.1 (Operational planning and control)

Implementation Guidance

  • Define lifecycle stages for your organization
  • Establish processes for each stage
  • Define stage gates and approval criteria
  • Document lifecycle management procedures
  • Integrate with existing development methodologies
  • Ensure traceability across stages

AI System Lifecycle Stages

StageKey Activities
ConceptionIdea generation, feasibility assessment
DesignRequirements, architecture, approach selection
Data CollectionData acquisition, preparation, labeling
DevelopmentModel development, training, optimization
VerificationTesting, validation, bias assessment
DeploymentProduction release, integration
OperationDay-to-day operation, support
MonitoringPerformance monitoring, drift detection
RetirementDecommissioning, archival, transition
Audit Questions - A.6.1.2

• What lifecycle stages do you define?
• Show me your lifecycle management process
• What are the stage gate criteria?
• How do you track AI systems through the lifecycle?

A.6.1.3 Responsible AI

AttributeDetails
ControlPrinciples of responsible AI relevant to the organization shall be defined and implemented throughout the AI system life cycle.
PurposeEmbed ethical AI throughout development and use
Related Clause5.2 (AI Policy)

Implementation Guidance

  • Define responsible AI principles
  • Integrate principles into development processes
  • Train personnel on responsible AI
  • Implement checkpoints for principle adherence
  • Review systems against principles
  • Update principles based on emerging standards

Common Responsible AI Principles

PrincipleDescription
FairnessAI systems treat all people equitably
TransparencyAI operations are understandable and open
AccountabilityClear responsibility for AI outcomes
PrivacyPersonal data is protected
SafetyAI systems operate safely and reliably
Human OversightHumans maintain appropriate control
BeneficenceAI provides benefit to users and society
Audit Questions - A.6.1.3

• What responsible AI principles have you defined?
• How are principles implemented in development?
• How do you verify adherence to principles?
• Show me how principles are embedded in your processes

A.6.1.4 AI System Life Cycle Documentation

AttributeDetails
ControlAI systems shall be documented according to defined requirements throughout their life cycle.
PurposeMaintain comprehensive AI system records
Related Clause7.5 (Documented information)

Documentation Requirements

Lifecycle StageDocumentation
DesignRequirements, design decisions, architecture
DataData sources, preparation, quality assessments
DevelopmentModel specifications, training parameters, experiments
TestingTest plans, results, validation reports
DeploymentDeployment procedures, configurations
OperationUser guides, operational procedures
MonitoringMonitoring specifications, thresholds
Audit Questions - A.6.1.4

• What documentation requirements exist for AI systems?
• Show me documentation for [specific AI system]
• How do you ensure documentation is maintained?
• What templates do you use?

A.6.2 AI System Life Cycle Stages Controls

A.6.2.2 Defining Objectives

AttributeDetails
ControlObjectives for the AI system and its intended use shall be defined and documented.
PurposeEstablish clear purpose and success criteria

Implementation Guidance

  • Define business objectives for each AI system
  • Document intended use cases
  • Specify success criteria and metrics
  • Identify constraints and boundaries
  • Document what the AI system should NOT do
  • Obtain stakeholder agreement on objectives
Audit Questions - A.6.2.2

• What are the objectives of [specific AI system]?
• How are objectives documented?
• Who approves AI system objectives?
• How do you define intended use?

A.6.2.3 Assessing Feasibility

AttributeDetails
ControlFeasibility of achieving objectives shall be assessed and documented prior to development or acquisition.
PurposeEnsure AI projects are viable before investment

Feasibility Assessment Areas

AreaAssessment Questions
TechnicalIs this technically achievable with current methods?
DataIs sufficient quality data available?
ResourceDo we have skills, budget, infrastructure?
EthicalCan this be done responsibly?
LegalAre there regulatory barriers?
BusinessDoes the business case justify investment?
Audit Questions - A.6.2.3

• How do you assess feasibility before development?
• Show me a feasibility assessment
• What criteria determine go/no-go decisions?
• Have any projects been rejected based on feasibility?

A.6.2.4 Technical Documentation

AttributeDetails
ControlTechnical documentation shall be produced and maintained throughout the AI system life cycle.
PurposeEnable understanding and maintenance of AI systems

Technical Documentation Content

  • System architecture and design
  • Model specifications and parameters
  • Training methodology and data
  • Performance metrics and benchmarks
  • API specifications and interfaces
  • Dependencies and requirements
  • Known limitations and constraints
Audit Questions - A.6.2.4

• What technical documentation do you maintain?
• Show me technical documentation for [AI system]
• How is documentation kept current?
• Who is responsible for technical documentation?

A.6.2.5 Maintaining Records

AttributeDetails
ControlRecords related to AI systems shall be maintained throughout the AI system life cycle.
PurposeEnsure traceability and auditability

Records to Maintain

  • Decision records and approvals
  • Change records and version history
  • Testing and validation records
  • Incident and issue records
  • Performance monitoring records
  • Training data provenance
  • Model versions and experiments
Audit Questions - A.6.2.5

• What records do you maintain for AI systems?
• How long are records retained?
• Show me records for [specific decision/change]
• How do you ensure record integrity?

A.6.2.6 Engaging Interested Parties

AttributeDetails
ControlRelevant interested parties shall be engaged throughout the AI system life cycle.
PurposeIncorporate stakeholder perspectives

Stakeholder Engagement Activities

StageEngagement Activities
DesignRequirements gathering, user research
DevelopmentFeedback on prototypes, beta testing
DeploymentUser training, change management
OperationSupport channels, feedback collection
MonitoringUser satisfaction surveys, complaints
Audit Questions - A.6.2.6

• How do you engage stakeholders in AI development?
• Which stakeholders are involved at each stage?
• Show me evidence of stakeholder engagement
• How do you incorporate stakeholder feedback?

A.6.2.7 Approaches for Achieving Objectives

AttributeDetails
ControlApproaches for achieving objectives shall be defined.
PurposeSelect appropriate methods for AI development
Audit Questions - A.6.2.7

• How do you select AI approaches/methods?
• What alternatives were considered?
• Why was this approach chosen?
• How do you document approach decisions?

A.6.2.8 Defining System Requirements

AttributeDetails
ControlRequirements for AI systems shall be defined and documented.
PurposeEstablish clear specifications for AI systems

Requirement Types

TypeExamples
FunctionalWhat the system must do
PerformanceAccuracy, latency, throughput
SecurityAccess control, data protection
ComplianceRegulatory requirements
UsabilityUser interface, accessibility
EthicalFairness, transparency requirements
Audit Questions - A.6.2.8

• How are AI system requirements defined?
• Show me requirements for [AI system]
• How do you handle requirement changes?
• Who approves requirements?

A.6.2.9 Verification and Validation

AttributeDetails
ControlVerification and validation of AI systems shall be performed, including system performance against defined objectives.
PurposeEnsure AI systems meet requirements and objectives

Verification vs Validation

AspectVerificationValidation
QuestionAre we building it right?Are we building the right thing?
FocusTechnical correctnessBusiness value and fitness
MethodsTesting, code review, analysisUser acceptance, real-world testing

AI-Specific Testing

  • Model accuracy and performance testing
  • Bias and fairness testing
  • Robustness and adversarial testing
  • Edge case and boundary testing
  • Integration testing
  • User acceptance testing
Audit Questions - A.6.2.9

• How do you verify and validate AI systems?
• Show me test results for [AI system]
• How do you test for bias?
• What acceptance criteria must be met?

A.6.2.10 AI System Operation and Monitoring

AttributeDetails
ControlOperations and performance of AI systems shall be monitored.
PurposeEnsure ongoing AI system effectiveness
Related Clause9.1 (Monitoring, measurement, analysis and evaluation)

Monitoring Areas

AreaMetrics
PerformanceAccuracy, precision, recall, latency
DriftData drift, concept drift, model degradation
UsageVolume, user patterns, adoption
IncidentsErrors, failures, complaints
FairnessOutcomes across groups
ResourcesCompute, memory, costs
Audit Questions - A.6.2.10

• How do you monitor AI systems in production?
• What metrics do you track?
• How do you detect model drift?
• What triggers investigation or intervention?
• Show me monitoring dashboards

Control Implementation Summary

ControlKey Evidence
A.6.1.2 Lifecycle ManagementLifecycle process documentation, stage gates
A.6.1.3 Responsible AIPrinciples document, integration evidence
A.6.1.4 Lifecycle DocumentationDocumentation standards, examples
A.6.2.2 ObjectivesObjective statements, success criteria
A.6.2.3 FeasibilityFeasibility assessments, decisions
A.6.2.4 Technical DocumentationTechnical specs, architecture docs
A.6.2.5 RecordsRecord retention, audit trails
A.6.2.6 Stakeholder EngagementEngagement records, feedback
A.6.2.7 ApproachesApproach selection rationale
A.6.2.8 RequirementsRequirements documents
A.6.2.9 Verification & ValidationTest plans, results, sign-offs
A.6.2.10 MonitoringMonitoring specs, dashboards, alerts
Key Takeaways - A.6

1. A.6 is the largest domain with 12 controls
2. Lifecycle management requires defined processes and stage gates
3. Responsible AI principles must be defined AND implemented
4. Documentation is required throughout the lifecycle
5. Verification and validation must include AI-specific testing (bias, robustness)
6. Monitoring must continue throughout operation

AI Assistant
00:00