AI Pilot — Export & Observability
Audience: Platform ops, SOC / compliance analysts
Time: ~15 min
Stream AI Pilot decisions and traffic events to Grafana, Splunk, Elastic, or other SIEM tools using Control Core’s event streaming — without storing raw prompts by default.
What
AI Pilot emits structured audit events (traffic, decisions, redactions, MCP tool calls, breakers, fallbacks). You configure a SIEM / streaming target once in Settings; events leave through the platform outbox with correlation IDs.
Why
Command Center charts are for operators. External SIEM and dashboards are for long-term retention, alerting, and evidence packs. Export keeps governance visible where your security team already works.
How — configure from Settings
- Open Settings → Audit logs → SIEM (or your tenant’s event-streaming page).
- Choose a destination: Splunk, Elastic, Grafana Loki / webhook, or another supported connector.
- Enter endpoint URL, auth (token or basic), and index/pipeline name as required by your tool.
- Enable AI Pilot / Advanced Audit sampling if volume is high (Settings → AI Pilot → Advanced Audit). Lineage and policy-violation style events stay at full fidelity where documented.
- Save and send a test AI request through a governed bouncer.
Troubleshooting: No events arrive? Confirm the SIEM target is enabled, credentials valid, and the bouncer is producing Activity rows first. See Advanced AI audit.
Event types (business meaning)
| Event (technical name) | Business meaning |
|---|---|
AI_TRAFFIC_LOG | Every governed AI call (who, model, allow/deny, tokens, cost) |
AI_DECISION_LINEAGE | Why the control engine allowed or denied |
AI_PII_REDACTION | Inspection redacted sensitive data in prompt/response |
AI_POLICY_VIOLATION | A control denied the request |
AI_MCP_TOOL_CALL | An MCP tool was invoked (and whether allowed) |
AI_BREAKER_TRIPPED / AI_FALLBACK_USED | Resilience: upstream unhealthy or fallback path used |
Every event carries a stable correlation ID so you can stitch request → decision → redaction in your SIEM.
Verify
- Trigger one allow and one deny (or block) through the bouncer.
- In your SIEM, search by correlation ID from AI Pilot → Audit (or Activity).
- Confirm both events landed in the expected index within your connector’s lag window.
Plain queries by business meaning
Use these as starter searches (adapt field names to your connector mapping):
| Business question | Example search idea |
|---|---|
| Denied AI calls today | Filter event = AI_POLICY_VIOLATION (or decision = deny) last 24h |
| PII redactions by team | Filter AI_PII_REDACTION, group by app or principal attribute |
| Cost burn by model | Sum cost_usd (or tokens) from AI_TRAFFIC_LOG, group by model |
| MCP tool usage | Filter AI_MCP_TOOL_CALL, group by tool name |
| Failover pressure | Count AI_FALLBACK_USED + AI_BREAKER_TRIPPED over time |
Avoid over-engineering PromQL on day one — start with these business filters, then add dashboards.
Troubleshoot
| Symptom | Where to look | Resolution |
|---|---|---|
| SIEM empty, Activity full | Settings → SIEM status / delivery logs | Fix auth, TLS, or index name; retry test event |
| Only some event types | Advanced Audit sampling / toggles | Enable the event categories you need |
| Missing prompt text | By design | Hashes only by default; raw body capture needs explicit DPO approval |