AI assistants can now query your Kubernetes clusters, diagnose failures, and apply remediations — but should they? We built an MCP bridge that lets AI providers interact with live multi-cluster environments through a local agent, then watched what happened when operators used it. This talk covers: (1) connecting AI models to live cluster state — why a local agent with the user's own kubeconfig matters more than a service account, (2) guided missions that walk operators through installing CNCF projects across clusters, and why pre-flight checks and rollback are non-negotiable when AI suggests the steps, (3) the trust boundary — operators want AI help but not kubectl apply without review, and how we landed on "suggest, confirm, execute," (4) failure modes in production: AI recommending actions on the wrong cluster, hallucinating resource names, and the guardrails we added after each. Attendees leave with patterns for safely integrating AI assistants into Kubernetes operations.