# Nordstrom built a Kong-based AI gateway with tenant isolation and real-time cost controls

Jen Wheaton explains how Nordstrom stood up an AI gateway with Kong, moving from proof of concept to production in a few months by building on an existing internal developer platform and reusable CI/CD workflows. The talk shows why a standard API gateway was not enough for AI traffic: Nordstrom needed visibility into prompts and responses, guardrails on model output, and rate limiting based on tokens and cost rather than request count alone. The proof of concept covered LLM calls, MCP traffic, and agent-to-agent communication, while production today focuses first on LLM traffic with a self-hosted data plane so traffic stays on the Nordstrom network. A major theme is tenant isolation and security, including removing provider credentials from agent processes, validating tenant configuration through app-ID-based workflows, and preventing one tenant from affecting another. Weaten also highlights real-time cost observability and previews a move to AI Gateway 2.0 for stronger MCP governance, API-to-MCP conversion, downscoping, and deny-by-default agent communication.