AI Pilot Configuration

Audience: Platform engineers, AI governance owners Time: ~20 min Prerequisites:

• A registered bouncer in the target environment (Bouncer Deployment)
• At least one upstream AI provider credential (OpenAI key, Azure deployment, AWS role, etc.)

This guide walks through all tabs of the AI Pilot dashboard (both the core tabs and the enterprise tabs added in 2026-Q2) and the underlying AI Pilot manifests each tab produces.

TL;DR

1. Open AI Pilot and select a bouncer on the Overview tab.
2. In Routing & models, add one backend per provider, then a route that sends the model to it.
3. In Guardrails & DLP, enable classifiers and DLP profiles that match your risk posture.
4. (Enterprise) Configure Cost Optimization, Observability, Cache & Rate Limits, mTLS, Resilience, MCP Proxy, Upstream Auth, and Advanced Audit as needed.
5. Save. The Control Plane compiles your inputs into AI Pilot manifests (9 CRD kinds) and pushes them to the selected bouncer.
6. Validate in Analytics and Traffic log; verify AI_TRAFFIC_LOG events in the audit trail.
7. (Enterprise) Verify Token Ledger, Audit stream, and Decision Lineage on the /pilot dashboard.

1. Overview tab

The Overview tab shows all bouncers in the current environment. Four summary cards give you:

• Bouncer count (connected / intercepting)
• Current environment (Sandbox or Production)
• AI Pilot status — Managed when Control Plane has at least one bouncer under management
• SIEM outbox (AI) — delivered / failed / pending counts for AI events, plus last delivery timestamp

Quick links open the audit trail pre-filtered to AI events, guardrails, or traffic, and to the SIEM settings page.

2. Analytics tab

Aggregated AI telemetry. All numbers come from audit events captured by the bouncer's AI Pilot filters — no separate metrics pipeline is required.

• Time series: Requests, Tokens (in/out), Estimated cost, TTFT
• Breakdowns: by Model, Provider, Guardrail category
• Selectable window: 1h, 24h, 7d, 30d

Troubleshooting: If charts are empty, confirm the bouncer is intercepting traffic and that AI_TRAFFIC_LOG events are being emitted (see the Traffic log tab).

3. Traffic log tab

A per-request log view. Each row is an AI_TRAFFIC_LOG audit event with rich AI context:

• Outcome and guardrail / DLP action
• Model, provider, tool name
• Tokens in/out, estimated cost
• Duration, TTFT, ITL
• Cache status (hit / miss / bypass)

Filter by bouncer, model, provider, outcome, and guardrail action; export filtered results to CSV.

4. Routing & models tab

This tab is the visual front-end for AI Pilot manifests. Your inputs are compiled into a multi-document YAML bundle and delivered to the bouncer.

Service backends

Each backend describes one upstream provider.

Credentials are stored in the Control Plane's credential vault; only a reference is ever sent to the bouncer.

Route rules

Each rule is conditional on request attributes. All conditions in a rule use AND.

Condition types:

• model — caller-supplied model name (supports exact match or prefix)
• header — any request header
• tool_name — MCP tool invocation name
• metadata — decision metadata key

For each matched rule you provide:

• Primary backend (required)
• Fallback chain (ordered list used on 5xx, rate-limit, or circuit break)
• Rate limit descriptor reference
• Model mapping to rewrite the caller's model name to the provider's actual model ID

Rate limits

Rate limits are token-aware (not only request-count). Each descriptor names:

• name — unique identifier
• tokens_per_minute
• concurrent_requests
• budget_usd_per_day (optional hard cap)
• scope — global, per_user, or per_api_key

Preview

A live YAML panel shows the exact multi-document bundle that will be pushed. Use it to review before saving.

Troubleshooting: If save returns a 400, the preview panel highlights the offending document. Most common cause: a referenced credential ref does not exist in the vault.

5. Guardrails & DLP tab

These controls run in the bouncer's external processor filter alongside AI Pilot. They are Control Core's value-add. Everything the ext_proc does is captured as audit events and streamed to SIEM.

Guardrails

• Classifier provider — Llama Guard (self-hosted), Azure AI Content Safety, OpenAI moderation, or custom
• Endpoint / model / timeout
• Apply to prompt / Apply to response toggles
• Category policies — for each category (e.g. self-harm, violence, jailbreak), choose the action (block, redact, annotate, off) and threshold

DLP profiles

Each profile bundles detectors plus an action. Supported built-in detectors include:

credit_card, ssn, iban, us_bank_routing, email, phone, passport, ip_address, jwt, aws_access_key, aws_secret, gcp_service_account, pgp_private, ssh_private, api_key_generic.

You can also supply custom regex patterns and an allowlist of terms that should never trigger a match.

Prompt cache

• Enabled toggle
• TTL (seconds)
• Semantic match (optional; requires an embedding backend)
• Include system prompt — if off, two requests that differ only in their system prompt will share a cache entry

Troubleshooting: If a request that should be blocked is not, check the category threshold — scores below the threshold are treated as safe.

6. Enterprise tabs (2026-Q2)

The following tabs were added for multi-provider, multi-target AI governance. They appear on the Settings → AI Pilot page alongside the original tabs.

Cache & Rate Limits tab

Configure how the bouncer stores rate-limit counters and semantic cache entries.

See Cache & Rate Limits for deployment instructions.

mTLS tab

Enforce mutual TLS between the bouncer and internal AI targets.

See mTLS Enforcement.

Resilience tab

Circuit breakers, retries, hedging, and fallback chains.

See Resilience & Circuit Breaker.

MCP Proxy tab

Configure the bouncer as an intelligent proxy between MCP clients and MCP servers.

See MCP Proxy Mode.

Upstream Auth tab

Manage credentials for connected AI/LLM providers at the bouncer, eliminating API keys in app code.

See Upstream Auth.

Advanced Audit tab

Deep audit trail configuration for compliance and forensics.

See Advanced AI Audit.

Cost Optimization tab (redesigned)

The Cost tab now supports multi-target cost rules instead of a single global limit.

Each cost rule specifies:

See Cost Optimization (multi-provider).

Observability tab (redesigned)

The Observability tab now supports dimensional filtering. Apply any combination of:

• Time window (1h, 6h, 24h, 7d)
• Target type (llm, mcp, rag, agent, api)
• Provider, Model
• MCP server, AI agent, RAG service
• Application, User
• Decision outcome (allow, deny, rate_limited)

See Observability Dimensions.

Audit & SIEM linkage

Every decision AI Pilot and ext_proc make lands in the audit log with one of these event types:

• AI_TRAFFIC_LOG — a successful call
• AI_POLICY_VIOLATION — a guardrail or DLP block
• AI_PII_REDACTION — content was redacted before reaching the model or the caller
• AI_TOKEN_USAGE — per-transaction token consumption and cost (2026-Q2)
• AI_MCP_TOOL_CALL — MCP tool invocation audit (2026-Q2)
• AI_DECISION_LINEAGE — full decision chain including policy path and obligations (2026-Q2)
• PILOT_* — configuration changes

Deep links in the Overview tab jump to the audit page with the right filter applied:

• /audit?quickFilter=ai-pilot — everything AI-related
• /audit?quickFilter=ai-guardrails — guardrail and DLP events only
• /audit?quickFilter=ai-traffic — traffic log only

To export these to Splunk, Elastic, QRadar, Sentinel, or Grafana Loki, configure a SIEM target at Settings -> Audit logs. The dashboard's SIEM outbox (AI) card shows pending / delivered / failed counts and surfaces the most recent delivery error.

Bouncer Plugin Admin Endpoints (2026-Q2)

The bouncer's plugin admin server (default port 9998) exposes these endpoints for the Control Plane to query:

These endpoints are consumed by the Control Plane's /pilot/bouncer/{pep_id}/* API routes and displayed on the /pilot dashboard.

Validation checklist

• Bouncer selected in the intended environment
• At least one backend saved successfully; preview YAML passes lint
• Guardrails category policies reviewed for false positives
• DLP profile tested with a synthetic prompt
• AI_TRAFFIC_LOG events visible in the traffic log
• SIEM outbox card shows non-zero delivered after first traffic
• (Enterprise) Cache topology configured and probe returns healthy
• (Enterprise) Cost rules cover all active providers and models
• (Enterprise) Observability filters return expected dimensional data
• (Enterprise) Decision lineage events visible in Audit tab

Next steps

• AI Pilot Troubleshooting
• AI Pilot Enterprise Onboarding
• Observe & Audit — audit schema and export
• Controls Manager Guide — author controls that consume AI attributes