Trust isn’t just about AI making good decisions. It’s about knowing who accessed the system, what they did, how much it cost, and whether policies were followed. Governance makes AI operations auditable.

The Four Pillars of AI Governance

Access Control: RBAC for AI

Role-Based Access Control (RBAC) maps users to roles, and roles to permissions. This creates a clear hierarchy:

# Role definitions
roles = {
    "admin": {
        "permissions": ["query", "admin", "audit", "configure"],
        "quotas": {"daily_queries": 10000, "daily_cost": 100.0}
    },
    "analyst": {
        "permissions": ["query", "audit"],
        "quotas": {"daily_queries": 1000, "daily_cost": 25.0}
    },
    "viewer": {
        "permissions": ["query"],
        "quotas": {"daily_queries": 100, "daily_cost": 5.0}
    }
}

How RBAC is Enforced

Services call the governance endpoint before executing any operation. The check is synchronous and fail-secure (deny on error):

# POST /api/v1/access/check
{
  "tenant_id": "acme-corp",
  "user_id": "alice@acme.com",
  "resource": "knowledge-base",
  "action": "query"
}

# Response
{
  "allowed": true,
  "matched_roles": ["analyst"],
  "matched_permissions": ["kb-query"]
}

# Or if denied:
{
  "allowed": false,
  "reason": "No matching permissions"
}

The enforcement pattern: User → Roles lookup → Permissions collected → Match against (resource, action). Every check is logged to the audit trail.

Policy Engine: Rules That Enforce Behavior

Policies define what’s allowed and what happens when rules are violated. Each policy has conditions and actions:

# Policy rule structure
policy_rule = {
    "name": "cost-limit-per-request",
    "description": "Block requests that exceed $0.50",
    "conditions": {
        "estimated_cost": {"operator": "gt", "value": 0.50}
    },
    "action": "BLOCK",
    "message": "Request exceeds per-request cost limit"
}

Policy Actions

• BLOCK: Stop the request immediately with an error
• WARN: Allow but log a warning for review
• REQUIRE_APPROVAL: Queue for human approval before execution
• RATE_LIMIT: Slow down rather than block

Common Policy Patterns

# Sensitive data policy
{
    "name": "no-pii-to-external-models",
    "conditions": {
        "has_pii": {"operator": "eq", "value": True},
        "model_type": {"operator": "eq", "value": "external"}
    },
    "action": "BLOCK"
}

# Time-based policy
{
    "name": "off-hours-approval",
    "conditions": {
        "hour_of_day": {"operator": "not_in", "value": [9,10,...,17]}
    },
    "action": "REQUIRE_APPROVAL"
}

Cost Tracking: Know Your Spend

AI costs can spiral quickly. GPT-4 at $0.03/1K tokens adds up. Cost tracking provides:

• Real-time visibility into spending
• Per-user and per-role budgets
• Alerts when approaching limits
• Historical trends for forecasting

# Cost tracking aggregates
cost_summary = {
    "period": "2024-12-07",
    "total_requests": 1247,
    "total_tokens": 892350,
    "total_cost_usd": 42.50,
    "by_model": {
        "gpt-4o": {"requests": 423, "cost": 28.40},
        "claude-3-haiku": {"requests": 824, "cost": 14.10}
    }
}

Real-Time Tracking with Redis

The POC uses Redis for fast, real-time cost aggregation across multiple time periods:

# Redis keys for cost tracking (from governance-service)
usage:daily:{tenant}:{user}:tokens:2024-12-07     # Daily aggregate
usage:weekly:{tenant}:{user}:tokens:2024-12-02    # Weekly aggregate
usage:monthly:{tenant}:{user}:tokens:2024-12      # Monthly aggregate

# Each request increments atomically
redis.incrbyfloat(daily_key, token_count)

# Check current usage in milliseconds
current_usage = redis.get(daily_key)  # Fast quota checks

Multi-Period Quotas

Quotas can be enforced at daily, weekly, or monthly levels – each with its own limit and alert threshold:

# Quota configuration
quota = {
    "metric_type": "TOKENS",
    "quota_type": "daily",       # daily | weekly | monthly
    "limit": 100000,
    "alert_threshold": 0.8,      # Alert at 80%
    "current_usage": 75000,
    "usage_percentage": 0.75
}

# When threshold exceeded, alert generated automatically

Audit Logging: Every Action Recorded

Audit logs capture everything. Every query, every response, every policy decision. This creates an immutable record for:

• Compliance audits
• Incident investigation
• Usage pattern analysis
• Security forensics

Audit Log Structure

# Each audit entry captures comprehensive context
audit_entry = {
    "timestamp": "2024-12-07T14:32:18.234Z",
    "event_type": "QUERY_EXECUTED",
    "user_id": "alice@company.com",
    "role": "analyst",
    "request_id": "req_abc123",
    "model": "gpt-4o",
    "tokens": {"prompt": 847, "completion": 400},
    "cost_usd": 0.08,
    "policies_evaluated": ["cost-limit", "pii-check"],
    "policies_passed": True,
    "latency_ms": 1847
}

Compliance Reports

Governance data feeds into compliance reports. Common requirements:

• Access reports: Who accessed what data and when
• Cost reports: Spending by team, user, or model
• Policy reports: What was blocked and why
• Usage reports: Patterns and anomalies

# Compliance report structure
compliance_report = {
    "report_type": "MONTHLY_AUDIT",
    "period": "2024-12",
    "summary": {
        "total_requests": 42847,
        "unique_users": 127,
        "total_cost": 1847.23,
        "policies_blocked": 234,
        "pii_incidents": 12
    },
    "top_users_by_cost": [...],
    "blocked_by_policy": {...},
    "recommendations": [
        "Consider increasing analyst quotas - 15% hit limits"
    ]
}

What to Measure: Governance Metrics

Governance produces measurable signals. Track these to know your AI system is under control:

# Key governance metrics (from compliance_reporter.py)
governance_metrics = {
    # Access Control
    "active_users": 127,
    "total_roles": 5,
    "access_denied_count": 34,      # Attempts blocked by RBAC

    # Policy Enforcement
    "policy_violations": 12,        # Requests blocked by policy
    "pii_incidents": 3,             # PII detected and blocked

    # Cost Control
    "total_cost_usd": 1847.23,
    "quota_breaches": 7,            # Users who hit limits
    "cost_by_model": {"gpt-4o": 1200, "claude": 500},

    # Audit Coverage
    "total_events_audited": 42847,
    "audit_gaps": 0                 # Operations without audit record
}

Why Local Governance Matters

Cloud platforms provide governance features, but local control gives you:

• Custom policies: Define rules that match your specific requirements
• Complete audit access: Export full logs, not just summaries
• Integration flexibility: Connect to your existing SIEM, IAM, or compliance tools
• Data sovereignty: Audit logs never leave your control

Coming Up Next

Governance tracks who did what. But for agentic AI workflows, we need deeper observability. In the final post, we’ll explore Agent Observability – tracking task completion, workflow efficiency, and cost analysis for AI agents.