If your business is running AI agents, you’re probably asking two questions right now: what are they accessing, and what’s it costing you?

Those questions just got easier to answer. On April 15, 2026, Databricks announced major upgrades to its AI Gateway, moving it into Unity Catalog and renaming it Unity AI Gateway. The change is not just cosmetic. It brings the same enterprise-grade governance that data teams use to manage data access to the AI infrastructure layer.

Concurrently, Salesforce rolled out similar agentic governance controls, a sign that the entire enterprise platform industry is converging on the same problem: AI agents running in production need guardrails that traditional IT controls were never built to handle.

The Problem Unity AI Gateway Solves

When a single agent queries an LLM, calls an external API, and reads from an internal database, who owns each of those interactions? Who can audit them? Who can shut them down if something goes wrong?

Until now, those questions were hard to answer without stitching together multiple monitoring tools. Unity AI Gateway changes that by extending Databricks’ governance model, already used to manage data access across organisations, to cover the entire agentic stack.

The three core capabilities are:

MCP governance controls which AI agents can access which external systems. Databricks introduced support for on-behalf-of (OBO) access, meaning an agent can call an external MCP server or API while carrying the permissions of a specific user. This closes a major audit gap. Without it, agents effectively operate with anonymous or elevated credentials that are difficult to trace back to any individual or workflow.

End-to-end observability logs every LLM call and MCP interaction inside Unity Catalog, including dollar-amount costs attributed to specific models, teams, and workflows. As AI spending comes under sharper scrutiny from finance teams, cost attribution built into governance infrastructure is a practical requirement, not a luxury.

Unified model access gives teams a single API layer across different AI models, with built-in fallbacks, rate limits, and guardrails. Organisations can swap models or apply policy changes centrally without rearchitecting individual agents every time priorities shift.

Why This Matters Now

The governance challenge in agentic AI is real and growing fast. Research from Okta found that 88% of organisations have reported suspected or confirmed AI agent security incidents, yet most have not yet moved to treat AI agents as independent, identity-bearing entities. That gap is where risk accumulates.

Databricks’ own 2026 State of AI Agents report found that companies actively practising AI governance put 12 times more AI projects into production compared to companies operating without governance frameworks. That number is worth sitting with. Governance is not slowing these organisations down. It is making their deployment pipelines more reliable, which in turn makes production rollouts more frequent and less fraught.

The reason makes intuitive sense. Teams that can monitor what agents are doing, trace unexpected costs to a specific workflow, and enforce access policies without manual intervention are more confident about deploying. Teams without those controls tend to hesitate, or worse, deploy quietly and discover problems only after something breaks.

What This Means for Business

If your organisation runs on Databricks, Unity AI Gateway should be part of your baseline infrastructure before you put AI agents into production environments. It is not an optional add-on.

Three things to review right now:

Access policies. Map which agents need access to which external systems. MCP governance enforces this at the platform level rather than relying on developers adding checks manually in each agent.

Cost visibility. If you cannot currently tell which team or workflow is driving your AI infrastructure spend, Unity Catalog’s cost attribution will surface that. Set budget thresholds and alerts before usage scales and a surprise invoice arrives at quarter close.

Model flexibility. Check whether your current agents are locked to a single model vendor. The unified API layer in Unity AI Gateway makes it practical to run fallbacks or switch providers without rewriting agent code.

For businesses not on Databricks, the broader lesson still applies: whatever AI infrastructure you are running, governance should not be retrofitted after agents are already in production. The organisations shipping AI agents at scale, reliably, are the ones that invested in observability and access control early in the process.

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.