Databricks has made Unity AI Gateway generally available, positioning it as the control layer for enterprise AI traffic across models, agents and tools. The service applies governance at runtime and works with Unity Catalog to connect each request with an identity, permissions, lineage and audit history.
The scope extends beyond calls to foundation models. Databricks says the gateway can govern agents, Model Context Protocol servers, skills and tools, giving platform teams one place to observe usage and enforce policy as AI applications become more distributed.
Governance moves into the request path
Traditional governance often documents who should use a system and reviews activity after the event. Unity AI Gateway sits in the execution path, where it can apply controls when an application actually calls a model or tool. That creates a practical link between a user or workload identity and the resource it is trying to reach.
Unity Catalog provides the underlying permissions, lineage and auditing. The gateway adds contextual and runtime policies, so organisations can define guardrails around the circumstances in which a resource is used. This design is intended to make policy consistent across internally developed agents and externally supplied models rather than leaving each team to build its own wrapper.
The general-availability release includes request-level observability and cost attribution. Administrators can examine which teams, applications or agents are consuming resources, then use dashboards and Databricks Genie to explore that activity. Budgets and hard spending caps can place a limit around workloads that would otherwise continue to generate usage.
Smart Routing enters beta
Databricks is also introducing Smart Routing in beta. Instead of permanently tying an application to one model endpoint, the gateway can select a route using factors such as expected quality, price, performance, availability and the remaining budget.
That could help teams use a less expensive model for routine work while reserving a more capable model for harder requests. It could also give applications a way around an unavailable endpoint. Because Smart Routing is still a beta feature, organisations should verify supported endpoints and behaviour before relying on it for a production-critical failover path.
Routing creates a new testing question as well. If different models may handle otherwise similar requests, evaluation needs to cover the full policy rather than one fixed model. Output quality, latency and safety can change with the selected route, and logs should make that choice visible when an incident is reviewed.
A shared layer for a mixed AI estate
The gateway is aimed at companies that expect to use several model providers and a growing catalogue of agent tools. A central layer can reduce duplicated integration work and make costs easier to allocate, but it also becomes a significant piece of infrastructure. Availability, policy design and access administration deserve the same operational discipline as other gateways on a production path.
Databricks says thousands of customers have used its AI products and that the platform processed more than a quadrillion tokens over the past year. Those are company figures, but they explain the focus on controls that operate across many teams rather than within a single prototype.
General availability makes Unity AI Gateway a supported production option, while Smart Routing remains the part still under active beta evaluation. For existing Databricks customers, the notable change is that governance, tracking and spending controls can now follow AI requests across a broader set of assets without creating a separate policy system for each one.