Vendors push AI gateways

A year ago, most companies plugged artificial intelligence features straight into whatever model worked fastest. This week, three vendors pushed the opposite idea: put a checkpoint in the middle so every model call passes through one controlled layer first. (mulesoft.com) (vercel.com) (mlflow.org) That middle layer is called a gateway, and it works like a company switchboard. Instead of every app talking directly to OpenAI, Anthropic, Google, or another provider, the gateway can route the request, log it, and enforce one policy before anything leaves the building. (mulesoft.com) (vercel.com) MuleSoft said on March 30, 2026 that its artificial intelligence gateway language-model features are now generally available. The company says the product gives platform, finance-operations, and security teams one governed control plane for every large language model interaction across an enterprise. (mulesoft.com) MuleSoft is pitching the gateway as the artificial-intelligence version of the old application-programming-interface gateway. Its announcement says companies now need policy enforcement, audit trails, cost controls, and model portability for every large language model call, not just for ordinary software interfaces. (mulesoft.com) Vercel made a narrower move on April 6, 2026, but it hits a problem legal teams care about immediately. Its artificial intelligence gateway now lets a whole team turn on zero data retention, which means requests are routed only through providers that have zero-retention agreements with Vercel. (vercel.com 1) (vercel.com 2) Vercel’s documentation says that team-wide zero data retention applies to all requests for that team, and a request fails if no compliant provider is available for the selected model. The same page says the rule does not cover bring-your-own-key traffic, because those requests use the customer’s own provider contract instead of Vercel’s. (vercel.com) MLflow focused on the finance and observability side of the same bottleneck. Its budget policies let companies set a dollar threshold over a daily, weekly, or monthly window, then either send a webhook alert or reject later requests with an Hypertext Transfer Protocol 429 error once the limit is crossed. (mlflow.org) MLflow is also tying that control layer to tracing, which is the step-by-step record of what an artificial intelligence app did during a request. Its documentation says tracing captures inputs, outputs, intermediate steps, and metadata, and its automatic integrations now cover more than 40 model and agent libraries. (databricks.com) (mlflow.org) Put those three launches together and the pattern is clear: one vendor is selling policy consistency, one is selling privacy enforcement, and one is selling spend controls with debugging. The common idea is that enterprises no longer want 20 teams making 20 separate deals with 20 model providers. (mulesoft.com) (vercel.com) (mlflow.org) That is why gateways are showing up now instead of at the start of the artificial-intelligence boom. Once companies began mixing multiple providers, agent frameworks, and internal apps, the hard part stopped being “can the model answer” and became “who approved this call, where did the prompt go, and why did this invoice spike on Tuesday.” (mulesoft.com) (databricks.com) (vercel.com)