Development Lifecycle

Purpose

This document defines Rygen Technologies’ AI system development lifecycle, ensuring all AI systems are developed in accordance with ISO/IEC 42001:2023 requirements. It is intended to guide and inform the AI development process from end to end, ensuring that AI systems developed by Rygen are responsible and governed.

Scope

This lifecycle applies to all AI systems developed or substantially configured by Rygen:

  • AI-powered features developed for X1 Platform (IPaaS)
  • AI features developed for Corsair (TMS)
  • Internal AI tools and automations developed by Rygen
  • AI implementations requiring configuration, training, or customization

AI System Development Lifecycle

flowchart TD
    A[Problem Definition] --> B{Go / No Go?}
    B -->|No| C[Don't Implement]
    B -->|Yes| D[Data Collection and Integration]
    D --> E[Development]
    E --> F[Evaluation]
    F --> G{Acceptable?}
    G -->|No| D
    G -->|Yes| H[Deployment]
    H --> I[Monitoring]
    I --> J[Refinement]
    J --> D

Development Lifecycle Phases

Phase 1: Problem Definition

Purpose: Establish the problem, appropriate solution, feasibility, design and acceptable impacts and risks before proceeding.

Key Activities:

  • Follow AI-024 to define the problem, stakeholders, alternatives, objectives, success criteria and constraints.
  • Complete impact and risk assessments, feasibility evidence and the System Design Specification.
  • Record the authorized Go or No-Go decision, based on the assessments, and any conditions.

Deliverables:

  • AI Problem Definition
  • AI System Impact Assessment
  • AI System Risk Assessment
  • AI System Risk Report(s)
  • AI System Feasibility Study
  • Go/No-Go decision record
  • AI System Design Specification

All listed deliverables must include their required template sections. State shared context once and link it from related records. The feasibility study may reference problem and design context; it must document its evidence, results, constraints, delivery dependencies and recommendation. Binding requirements belong in the design specification, formal impact and risk findings in their assessments, and approvals in GitLab.

Definition of Done:

  • The problem, objective mapping and success basis are documented, with unresolved matters explicit.
  • Feasibility and design evidence support the proposed approach; risks and impacts are assessed and acceptable.
  • Authorized management approval and conditions are recorded through the decision record and GitLab.

Resources:

  • Human Resources: Principal AI Engineer, or a Lead Product Manager on the AI Governance Committee working with the AI Engineering Team (problem definition, impact assessment, risk assessment); Principal AI Engineer (feasibility evaluation); Product Manager or Technical Lead (requirements definition, business case); Relevant stakeholders (requirement gathering, approval)
  • Tools: Jira (project tracking); Risk assessment templates; Impact assessment templates; GitLab (documentation repository, decision records and approvals)
  • Data Resources: Historical system performance data (for feasibility assessment); Business/market requirements; Regulatory requirements (if applicable)
  • Computing Resources: Access to documentation systems (GitLab, Confluence); Development environment (for technical feasibility)

Phase 2: Data Collection and Integration

Purpose: Prepare data and infrastructure for AI system development

Key Activities (system-dependent):

  • For API Systems: Data flow documentation, integration testing
  • For Custom Models: Full data collection and preparation per AI-011

Deliverables:

  • For custom models: Data quality reports, prepared datasets
  • For API systems: the design specification’s integration and data sections

Definition of Done:

  • For custom models: All required data accessed and validated, data quality meets defined standards, processing steps documented
  • For API systems: Integration points tested and operational, data flows documented

Resources:

  • Human Resources: Data Engineers (data pipeline implementation); AI/ML Engineers (data requirements definition, quality assessment); DevOps Engineers (infrastructure setup for data storage/processing)
  • Tools: Data pipeline tools (system-specific); Data quality assessment tools; Version control systems (GitLab); Data cataloging systems
  • Data Resources: Source data systems and APIs; Training/validation/test datasets (for custom models); Data samples for integration testing (for API systems)
  • Computing Resources: Data storage infrastructure; Data processing environments; Development databases/data lakes

Phase 3: Development

Purpose: Build and configure the AI system

Key Activities (system-dependent):

  • For API-based systems: Configure API parameters, engineer prompts, build integration
  • For custom models: Train models with prepared data, validate model performance
  • Implement controls as required by risk assessment (explainability, human oversight, etc.)

Deliverables:

  • For custom models: Trained models with performance metrics
  • For API systems: Configured system with integration documentation
  • Implementation of required controls per risk assessment

Definition of Done:

  • System operational in development environment
  • Required controls implemented and tested

Resources:

  • Human Resources: AI/ML Engineers (model development for custom systems, API configuration); Software Engineers (integration development); DevOps Engineers (environment configuration)
  • Tools: Development IDEs and frameworks; ML platforms (Vertex AI, etc.); Third-party AI APIs (OpenAI, Anthropic, Google, etc.); Version control (GitLab); CI/CD tools
  • Data Resources: Prepared training datasets (custom models); API documentation and examples (API systems); Test data for validation
  • Computing Resources: Development environments; Model training infrastructure (GPU/TPU for custom models); API service accounts and quotas

Phase 4: Evaluation

Purpose: Validate system readiness for deployment

Key Activities:

  • Test against defined success criteria
  • Validate performance metrics
  • Conduct user acceptance testing (if applicable)
  • Verify compliance requirements per risk assessment
  • Document evaluation results

Deliverables:

  • Evaluation results against defined acceptance criteria, including applicable compliance verification, recorded in an evaluation report or in the system repository or MLflow and linked from the relevant Jira ticket or merge request
  • User acceptance results (when conducted)

Definition of Done:

  • Performance meets acceptable thresholds per system specifications
  • Compliance requirements verified per risk assessment
  • System approved for deployment, with applicable approvals recorded as required under Decision Authority
  • Any critical issues resolved or accepted as residual risk

Deployment approvals are recorded in the relevant GitLab merge request under Decision Authority; a separate deployment decision document is not required.

Resources:

  • Human Resources: AI/ML Engineers (performance evaluation); QA Engineers (testing); Domain experts (validation); Principal AI Engineer (compliance verification)
  • Tools: Evaluation frameworks and scripts; Testing tools; Performance monitoring tools; Compliance checklists
  • Data Resources: Test datasets; Validation datasets; Benchmark data
  • Computing Resources: Testing environments; Evaluation infrastructure

Phase 5: Deployment

Purpose: Deploy to production

Key Activities:

  • Execute deployment plan
  • Configure production monitoring
  • Provide user guidance (documentation, training, or support as needed)

Deliverables:

  • Deployment plan, recorded or referenced in the system design specification
  • System deployed in production
  • Monitoring and observability records

Definition of Done:

  • System operational in production
  • The monitoring in the system’s monitoring plan (AI-017) is in place
  • Users have access to necessary guidance for system use

Resources:

  • Human Resources: DevOps Engineers (deployment execution); AI/ML Engineers (deployment validation); Technical Writers (documentation); Support Team (preparation for operational support)
  • Tools: Deployment automation tools; Monitoring and observability platforms; Documentation systems; Communication platforms
  • Data Resources: Production data access (controlled); Migration scripts and data
  • Computing Resources: Production infrastructure; Monitoring systems; Backup systems

Phase 6: Monitoring

Purpose: Ensure ongoing performance

Key Activities:

  • Monitor performance metrics
  • Track and respond to incidents
  • Collect user feedback
  • Identify improvement needs
  • Report monitoring results at each management review (AI-017)

Deliverables:

  • Monitoring results in the management review report (AI-017)
  • Incident records (AI-023), as needed

Definition of Done:

  • Performance maintained within defined thresholds
  • Incident response procedures operational
  • Monitoring results reported at each management review (AI-017)

Resources:

  • Human Resources: DevOps Engineers (infrastructure monitoring); AI/ML Engineers (performance monitoring); Support Team (incident response); Principal AI Engineer (compliance monitoring)
  • Tools: GCP monitoring and logging; Custom monitoring dashboards; Alerting systems; Incident management tools (Jira)
  • Data Resources: Production data streams; Performance metrics; User feedback data
  • Computing Resources: Production systems; Monitoring infrastructure; Log storage systems

Phase 7: Refinement

Purpose: Improve the system. Refinement is undertaken when monitoring, incidents, changed requirements or planned development identify a need for change.

Key Activities:

  • Analyze monitoring data and feedback
  • Implement improvements
  • Update documentation
  • Reassess risks as needed

Deliverables:

  • Jira tickets and merge requests for each change (Changes to a System in Service)
  • Updated system documentation
  • Revised risk assessments (if needed)

Definition of Done:

  • Improvement opportunities identified and prioritized
  • Approved improvements implemented
  • Documentation updated to reflect current state

Resources:

  • Human Resources: AI/ML Engineers (improvement implementation); Data Engineers (data pipeline updates); Product Managers (prioritization); Principal AI Engineer (risk reassessment)
  • Tools: Development and testing tools (same as Phase 3); Analysis tools for feedback/metrics; Change management tools
  • Data Resources: Performance data and metrics; User feedback; New training data (if applicable)
  • Computing Resources: Development/staging environments; Analysis infrastructure

Changes to a System in Service

A model swap, model version update, prompt change, or code change is tracked in its Jira ticket and merge request. No governance document is written, except the design specification update that Resource Change Management requires. When the change can alter the system’s AI behavior, because it changes the model, a prompt, or code that builds the model’s input or acts on its output, the affected evaluation suites run before it is deployed, and the ticket or merge request links the result. Other code changes follow normal engineering review.

A change assessment record is written in the aims-governance repository before deployment only when the change could raise an assessed risk or impact: it sends personal or confidential client data to an AI provider not already approved for that system; it reaches new user groups or jurisdictions; it removes or weakens a control that a risk or impact assessment relies on, such as human review of output; or it otherwise materially changes an assessed risk or impact, such as a large increase in volume or exposure (AI-008). This includes a design that changes in any of these ways after its change assessment was approved. The record states whether the impact assessment (AI-009) and risk assessment (AI-008) still hold, and the documents that no longer hold are updated. Approval follows Decision Authority below.

Any other change, including a new feature or use that raises no assessed risk or impact, needs only its Jira ticket and merge request and a line in the design specification naming the feature, its model and provider, and the data it sends.

When it is unclear which applies, the Principal AI Engineer decides and notes the decision in the Jira ticket.

Resource Change Management

Changes to resources used by an AI system trigger updates to:

  • AI System Design Specification, when the system moves to a different model family or another resource the specification documents changes. A model swap or version update within the same model family needs no governance document. A model family is the provider’s named model line at the tier the design specification names, such as Gemini Flash, Gemini Flash Lite, Gemini Pro, Claude Opus or Claude Sonnet. A version change within that tier is the same family; a change of tier is a change of family.
  • Risk assessments (if resource changes affect risk profile)
  • Training requirements (if new tools or competencies needed)

Resource changes are tracked through:

  • GitLab merge requests for documentation updates
  • Jira tickets for implementation tracking
  • Version control on all resource-related documentation

Decision Authority

Risk-based approval is required: a system is approved by the authority that AI-008 (Risk Tolerance & Authority) sets for the highest residual risk level in its risk assessment.

In addition to the applicable risk-based approval:

  • AI technical approval: The Principal AI Engineer must approve new AI systems and material changes affecting AI behavior, capabilities, or risk before deployment, as defined in AI-005: Roles and Responsibilities.
  • Platform-owner approval: AI system changes that materially affect a platform must be reviewed and approved by the engineering team that owns that platform. Approval must be recorded in the relevant GitLab merge request before merge and deployment. When multiple platforms are materially affected, each platform’s owning team must approve.

Platform-owner approval covers the impact on the team’s platform and is required alongside AI technical and risk-based approvals.

Adaptive Implementation

Resource requirements adapt based on system type and complexity. The AI System Design Specification documents actual resources used for each specific system.

Revision History

VersionDateAuthorSummary of Change
1.02025-06-05Field BradleyInitial draft.
1.12025-09-02Field BradleyMigrated to markdown and gitlab
1.22025-10-08Field BradleyAdded feasibility study to Phase 1
1.32026-01-14Field BradleyAdded resource requirements to each lifecycle phase per ISO 42001 7.1, A.4.2–A.4.6
2.02026-11-04Field BradleyAligned deployment approvals with AI-005, including AI technical approval and platform-owner approval recorded in GitLab merge requests.
3.02026-11-09Field BradleyClarified Phase 1 deliverables and decisions, risk-level and model-family rules, change assessments only for changes that could raise a risk or impact, plan-based monitoring, and evaluation evidence and approvals kept in the repository, MLflow and GitLab (AI-2167, AI-2168, AI-2173, AI-2131). The Phase 5 monitoring condition is that the monitoring plan is in place.
3.12026-11-11Field BradleyPhase 1 problem definitions, impact assessments and risk assessments may be led by a Lead Product Manager on the AI Governance Committee working with the AI Engineering Team, as AI-005 allows (AI-2205)