AI Pilot Admin User Guide
Audience: Enterprise admins, AI governance owners
Time: ~15 min read
Use this guide to navigate the AI Pilot Command Center views and operate day-to-day governance with controls language — not application code changes.
What
The Command Center is the Control Plane surface for AI Pilot: health of governed AI traffic, live activity, models and routing, protection (inspection + controls), cost, and audit. Each bouncer shows a badge for Gateway Mode or App-Embedded so you know the deployment topology at a glance.
Why
Operators need one place to answer: Is AI traffic flowing? Which models are in use? Are safety and data-protection controls firing? Are we within budget? Can we prove decisions for compliance?
How — Command Center views
Open AI Pilot from the sidebar (or Settings → AI Pilot for configuration). Select a bouncer; note the Gateway Mode / App-Embedded badge.
Overview
- Fleet health for the selected bouncer (heartbeat, sync, intercept scope)
- High-level counters: requests, allows/denies, inspection actions, spend trend
- Quick links into Activity, Protection, and Cost
Verify: Counters move when you send test traffic through that bouncer.
Activity (Traffic)
- Live or recent request stream: caller context, model/path, decision, latency
- Filters by outcome (allow / deny / block / rate-limit) and resource type (LLM, MCP, agent)
Verify: A known test call appears within seconds with the expected decision.
Troubleshooting: Empty Activity with a healthy heartbeat usually means clients bypass the bouncer. Confirm base URLs. See Troubleshooting.
Models & routing
- Which models and providers are enabled for this bouncer
- Fallback order and route bindings (multi-provider)
- Allow/deny overlays driven by controls (who may call which model)
Verify: Disable a model in settings (sandbox), retry — expect 403 from the control, not a provider error.
Protection
- AI Inspection Engine outcomes: PII redaction, prompt safety / guardrails
- Content safety thresholds and trusted-domain exceptions
- Links to author or tune related controls
Verify: A prompt with a known PII pattern shows redaction flags in Activity or audit; a forbidden phrase yields 400 when set to block.
Cost & usage
- Token and spend trends per user / model / app (as configured)
- Daily/monthly caps and rate limits
- Fallback and cache hit indicators when those features are on
Verify: After traffic, charts update for the selected window. Over-limit callers receive 429.
Troubleshooting: Inconsistent caps across replicas → shared Redis required. See Cost section in Troubleshooting.
Audit
- Decision lineage and AI event types for compliance
- Correlation IDs for SIEM stitching
- Optional live audit stream on the Audit tab
Verify: Allow and deny paths both produce audit rows. Export path: Export & observability.
Controls vs gateway config
| You want to… | Prefer |
|---|---|
| Who may call which model / MCP tool | A control (Rego / Visual Builder) |
| Prompt PII / jailbreak inspection | Protection + AI Inspection Engine settings |
| Token budgets / fallbacks | Cost & Models & routing (gateway config) |
| Prove decisions externally | Audit + SIEM export |
Troubleshoot
| Symptom | Where to look | Resolution |
|---|---|---|
| Wrong mode badge | Deployment of that PEP | Confirm Gateway vs App-Embedded install; see Deployment modes |
| Overview empty | Bouncer selection / registration | Pick the correct PEP; confirm registration |
| Protection never fires | Intercept scope; LLM paths | Bind AI resources; ensure paths are LLM/GenAI routes |