What is AI runtime governance?

Summary

• Governance that runs with your AI. AI runtime governance enforces policy on AI assets — models, agents, tools, and MCP servers, plus their requests and responses — while they run, not only when they are built.
• One model for data and AI. On Databricks it extends Unity Catalog governance to every AI asset and adds the Unity AI Gateway as a central control plane for AI traffic and behavior.
• Real-time guardrails. Service policies evaluate every request before the model runs and every response after, with built-in protection for PII, unsafe content, and jailbreak or prompt-injection attempts.
• Cost and access control. Fine-grained rate limits, budgets with hard spend caps, and attribute-based access control govern who can call which model or tool and how much they can spend.
• Full observability. Every model call, tool invocation, and agent action is logged with end-to-end lineage and audit trails, stored as governed tables in the lakehouse.

What is AI runtime governance?

AI runtime governance is the practice of applying and enforcing policy on AI systems while they operate — controlling which models and tools can be called, what data can flow in and out, how much can be spent, and capturing a complete record of every interaction. It complements design-time controls by acting on live requests and responses. On the Databricks Data Intelligence Platform, AI runtime governance builds on two components: Unity Catalog governs AI assets as securable objects, and the Unity AI Gateway governs the traffic and behavior of every AI call from a central control plane.

Why Databricks governs AI at runtime with the Unity AI Gateway

• Asset governance in Unity Catalog. Every model, MCP server, function, and connection is managed as a securable object and governed with the same privileges and attribute-based access control (ABAC) policies used for tables and volumes. See How AI governance works on Databricks.
• Traffic governance. The Unity AI Gateway routes every model-service and MCP-service request through a central control plane, applying fine-grained rate limits, fallbacks, budgets, and usage tracking at the model level.
• Behavior governance with service policies. Service policies (guardrails) are evaluated in real time on every call — before the model runs and after it responds. Built-in guardrails provide deterministic PII detection with redaction or blocking, unsafe-content filtering, and jailbreak or prompt-injection protection; custom LLM-as-judge and SQL rules add context-specific decisions scoped by caller and tool. See service policies, guardrails, observability, and cost controls.
• Cost controls. Budgets set per-user, per-team, per-department, and per-application spend thresholds, and hard spend caps automatically halt requests when a budget is exceeded rather than alerting after the fact. Spend is attributed by user, team, tool, and use case.
• Identity and access. Access uses ABAC with tag policies and privileges such as EXECUTE (use a model or tool), CREATE SERVICE or CONNECTION (register an asset or connect a provider), and MANAGE (administer configuration and access). Agent identity is propagated through On-Behalf-Of authorization or service principals, so an agent acts with the caller's exact permissions.
• Observability, lineage, and audit. Payload logging captures full request and response data; every model call, tool invocation, and agent interaction is recorded in unified tracing tables stored as Delta tables in the lakehouse, with end-to-end lineage, audit trails, and built-in dashboards for usage, latency, and cost.

This lets you govern AI from a single place — Databricks-hosted foundation models and custom models, external models accessed through the gateway, registered MCP servers, and external coding agents routed through Databricks model services — without limiting developer choice of models, frameworks, or vendors. The Unity AI Gateway became generally available on August 4, 2026.

Getting started

• Read How AI governance works on Databricks for the runtime governance model.
• Review Unity AI Gateway is generally available for the current capability set.
• See what's new in service policies, guardrails, observability, and cost controls to plan your guardrails and budgets.
• Learn how to govern AI agents at scale with Unity Catalog.

FAQs

What does "runtime" mean in AI governance?

Runtime governance acts on live AI activity — the actual requests and responses — enforcing policy on every call rather than only reviewing configurations at build time.

How does Databricks stop sensitive data from leaking through an AI model?

Built-in guardrails detect PII on every request and response and can redact or block it, and service policies can deny or require approval for risky requests based on who is calling and what the content contains.

Can Databricks govern external models and agents?

Yes. The Unity AI Gateway governs Databricks-hosted models, external models accessed through the gateway, registered MCP servers, and external coding agents routed through Databricks model services.

Is AI activity auditable?

Yes. Payload logging and unified tracing record every model call, tool invocation, and agent action as governed tables in the lakehouse, with end-to-end lineage and audit trails.