Governance Documentation Skill
You help create, maintain, and evidence AI governance documentation that regulators, auditors, and internal oversight functions expect.
Important: You support compliance work but do not provide legal advice. Interpretations and inspection practices vary. Always verify current obligations and guidance.
1. EU AI Act documentation requirements by article
This section is organized around key documentation and recordkeeping expectations commonly mapped in governance programs. Use it to build an evidence map.
Article 11 - Technical documentation
What to maintain (typical contents):
- System description and intended purpose
- System architecture and components, including model(s), data pipelines, and interfaces
- Design and development choices and assumptions
- Risk management outputs and residual risk rationale
- Data governance approach (data sources, preprocessing, quality controls)
- Testing and validation results, including performance and robustness evidence
- Cybersecurity and resilience measures
- Human oversight design
- Change management and versioning for model, prompts, and configurations
Evidence examples:
- Technical file or model card and system card
- Architecture diagrams
- Validation reports
- Secure development lifecycle records
Article 12 - Recordkeeping and logs
What to maintain:
- Logging policy for AI systems
- Operational logs sufficient to enable traceability and investigation
- Access logs for sensitive components and datasets
- Incident related logs and investigation traces
Evidence examples:
- Logging standard and retention schedule
- Sample log extracts or log schema
- Monitoring dashboards and alerts
Article 13 - Transparency and information to users
What to maintain:
- User facing notices and instructions as required by the system type and risk tier
- Description of system capabilities and limitations
- Human oversight availability and how to seek review
- Explanation of output interpretation and safe use
Evidence examples:
- User notice templates
- Product UI screenshots of notices
- Terms of use and help center articles
Article 14 - Human oversight
What to maintain:
- Definition of human oversight measures per system
- Oversight roles, competencies, and training
- Intervention mechanisms, override, and stop conditions
- Escalation routes and appeals
Evidence examples:
- Oversight playbooks and runbooks
- Role descriptions and RACI
- Training records for operators and reviewers
Article 17 - Quality management system
What to maintain:
- Documented governance processes across the lifecycle
- Control ownership and accountability
- Internal audits and management reviews
- Corrective and preventive actions
- Supplier and subcontractor controls
Evidence examples:
- QMS policies and procedures
- Internal audit plan and reports
- Management review minutes
- CAPA register
Article 18 - Conformity assessment (high-risk)
What to maintain:
- Conformity assessment approach and results (as applicable)
- Mapping of essential requirements to controls and tests
- Evidence that post-assessment changes are controlled
Evidence examples:
- Conformity assessment report
- Essential requirements mapping table
- Change approvals and release notes
Article 19 - EU declaration of conformity (high-risk)
What to maintain:
- Declaration of conformity and supporting evidence package
- Sign-off authority and document version control
Evidence examples:
- Signed declaration of conformity
- Referenced technical documentation index
Article 26 - Deployer obligations (high-risk)
What to maintain:
- Assignment of responsibilities for deployer duties
- Operational controls ensuring use according to instructions
- Monitoring and oversight evidence in real-world use
- User information and transparency artifacts
- Incident reporting and corrective actions
Evidence examples:
- Deployer operating procedure
- Oversight and monitoring reports
- Training and competency records for users
2. What a regulatory inspection typically requests
Common inspection themes:
- Proof you know your AI systems: inventory with owners, purposes, and deployment status
- Proof you classified systems correctly: rationale, references, and decision records
- Proof you manage risks: risk assessments, controls, test results, residual risk acceptance
- Proof of traceability: logs, monitoring, change history, incident records
- Proof of human oversight: defined roles, training, intervention capability
- Proof of transparency: user notices, instructions, complaint and appeal pathways
- Proof of governance: policies, procedures, approvals, internal audits
- Proof of supplier management: vendor due diligence, contracts, and monitoring
A typical request list to be ready for:
- AI inventory register and risk classification results
- Technical documentation for selected systems
- Logging and monitoring policy plus example evidence
- DPIAs and fundamental rights assessments (if applicable)
- Incident and complaint register plus investigation notes
- Vendor assessment files, DPAs, SCCs, and security attestations
- Training evidence for staff involved in design, operation, and oversight
- Change management records for model and prompt updates
3. Document retention requirements
Retention needs depend on legal requirements, sector rules, and your own governance. Use these principles:
- Define retention by artifact type and risk tier
- Keep enough records to reconstruct decisions and system behavior during the retention period
- Align retention for AI logs with privacy and security requirements
- Keep superseded policy versions and approval records
Practical retention approach:
- Policies and procedures: keep all versions and approvals
- Inventory and risk classifications: keep current plus history
- DPIAs and assessments: keep for the life of the system plus an additional period consistent with privacy governance
- Logs and monitoring: define a retention schedule based on risk, detectability, and incident investigation needs
- Training records: keep for the duration of employment plus a defined period
- Vendor due diligence: keep for contract term plus a defined period
If unsure, align with your corporate retention schedule and consult counsel.
4. Best practices for living governance documentation
How to keep documentation audit-ready:
- Single source of truth: maintain an AI inventory register with unique system IDs
- Version control: track changes to models, prompts, datasets, and controls
- Evidence by design: create artifacts as part of the lifecycle, not after the fact
- Traceability: link each decision to an owner, date, rationale, and supporting evidence
- Risk tiering: apply heavier documentation for higher-risk systems
- Regular reviews: schedule periodic re-classification and control testing
- Minimal but sufficient: prefer concise summaries plus links to source evidence
- Redaction workflow: maintain sanitized versions for external sharing
- Cross-functional sign-offs: legal, DPO, security, product, and business owners
5. Template structures for key governance documents
Use these structures as starting points. Adapt to your organization.
AI inventory register (template)
Fields:
- System ID
- System name
- Owner (business) and owner (technical)
- Intended purpose and users
- Deployment status
- AI Act classification and rationale
- Data categories and sensitivity
- Vendors and dependencies
- Oversight measures
- Monitoring and KPIs
- Links to artifacts (DPIA, vendor assessment, policies)
- Last review date and next review date
Risk assessment log (template)
Fields:
- Risk ID
- System ID
- Risk category (bias, privacy, security, safety, etc.)
- Risk statement
- Likelihood and impact ratings
- Inherent risk score
- Controls and mitigations
- Residual risk score
- Owner
- Review cadence
- Evidence links
Incident log (template)
Fields:
- Incident ID
- Date detected and date closed
- System ID
- Description
- Severity
- Impacted users and data (high-level)
- Root cause
- Immediate containment actions
- Corrective and preventive actions
- Notifications made (internal and external)
- Lessons learned
6. ISO 42001 alignment
If pursuing ISO 42001 (AI management system), map your documentation to:
- Context of the organization and scope of the AIMS
- Leadership, roles, and responsibilities
- Planning and risk treatment
- Support (competence, awareness, communication, documented information)
- Operation (lifecycle controls and supplier management)
- Performance evaluation (monitoring, internal audit, management review)
- Improvement (nonconformity, corrective action, continual improvement)
Practical alignment tips:
- Use a common document control process for both ISO 42001 and AI Act evidence
- Maintain an evidence map that links ISO clauses and AI Act obligations to artifacts
- Ensure internal audit and management review outputs are documented and actionable