Cloudflare launches Identity-Aware AI Gateway to give IT and security teams visibility into AI usage by tying every AI request to a verified user identity.

Flexport, a customer, notes that shared credentials blocked tracing; with Cloudflare Access integrated into the AI Gateway, per-request identity is established without a separate authentication system.

The system supports rate limiting, request caching, and filtering for sensitive information to manage costs and data risk without hindering workflows.

The centralized control point enables cost control, data sanitization, and visibility across tools and model providers.

User Insights provides admins with a rogue-behavior feed for investigations, coaching, or policy adjustments, helping distinguish insider risk from legitimate high activity.

All AI requests pass through a single control point for centralized visibility, caching of repeats to cut costs, rate limiting, and automatic filtering of sensitive data from logs.

Industry voices from Flexport and others stress authenticating identities at the gateway and applying existing access policies rather than relying on shared API keys for AI usage.

Figures illustrate visual methods for interpreting session cost anomalies and classify alerts into four categories, explaining why some high costs can be ignored if they align with a user’s normal behavior.

User Insights helps detect anomalies early, identify cost-saving opportunities by matching tasks to cost-effective models, and cache repeated requests to reduce waste while enforcing rate limits.

A two-tier baseline approach uses an account-level p99 ceiling and per-user p95 baselines with a rolling window to adapt to usage shifts, targeting meaningful anomalies rather than routine variance.

Future plans include task-based smart routing to optimize cost and quality across models, plus prompt classification to detect misuse or misalignment with intended tasks.

Per-user spend limits can be enforced to prevent surprise invoices, with routing to cheaper models when limits are reached, and future plans for group-based controls.