Ten years ago you introduced an API gateway because APIs multiplied and you needed to know who calls them, how to secure them and how to monitor them. With AI systems the story repeats, but with new rules. Consumption is measured in tokens, the same prompt doesn’t return the same answer, attacks arrive through the content itself, and more and more often the caller isn’t a user but an agent. How do you set token limits and chargeback per team? How do you use semantic caching when the answer is never the same? How do you stop prompt injection before it reaches the model? What happens when your PTU deployment runs out of capacity? And how do you give an agent access to an MCP server without sharing keys? We’ll show how Azure API Management became an AI gateway for models, MCP servers and agents, on a platform with multiple teams, models and regions.