AI Governance

Non-negotiable governance rules for all AI capabilities in AIIA.

Core Principles

Principle	Enforcement
AI obeys RBAC	Permission check at AI Gateway before any retrieval or generation
No cross-role leakage	RAG filters by user's role and org permissions
Traceable outputs	Every AI response includes citations + confidence scores
Human approval required	AI = suggestion only; user must "Apply/Approve"
Full audit logging	AIRequest → AIResult → AuditLogEvent chain
Permission-aware retrieval	user_roles passed to RAG pipeline for filtering

AI Gateway Flow

Audit Trail for AI

Every AI interaction creates three records:

Record	Content
AIRequest	User identity, prompt, context data, timestamp
AIResult	AI response, citations, confidence, model used
AuditLogEvent	Action=AI_ACTION, linking request to result

IPPF Prompt Governance

AI prompts are governed assets:

• Versioned — each prompt has immutable version history
• Approved — changes require admin signoff
• Auditable — version changes are logged
• Methodology-aligned — IPPF overlays ensure professional standards

Permission-Aware Retrieval

The RAG pipeline respects user permissions:

1. User's roles and org_id extracted from JWT
2. Vector search filtered by accessible engagements
3. Results limited to user-authorized data
4. Citations reference only visible internal objects