Implementation Guide

Governing AI Agents in Databricks Genie and AI/BI

Databricks Genie and AI/BI agents inherit governance from Unity Catalog, which controls table, view, row, and column access and logs query execution. It does not enforce runtime restrictions on which tools an agent calls, in what sequence, or under what intent. Enterprises scaling Genie beyond pilots need a separate runtime layer that governs agent identity, least-privilege tool access, and full audit reconstruction of agent behavior, not just final query results.

Where Databricks Governance Ends

Four layers shape what a Genie or AI/BI agent can see and do today. Each one governs data or configuration, none of them governs runtime agent behavior.

Unity Catalog

Governs table, view, row, and column access for queries executed by Genie and AI/BI agents.

Genie Space Configuration

Curators define table scope, sample queries, and instructions that shape agent query generation.

Audit Logs

Capture discrete query and object-access events, not full agent reasoning or tool-call chains.

Mosaic AI Agent Framework

Extends agents with custom tools governed by the same Unity Catalog permission model.

How to Approach the Governance Gap

Before defining controls, teams need a clear picture of what each Genie space or agent deployment can actually reach and how that maps to identity.

  1. Mapping the Governance Gap

    Inventory every Genie space and agent deployment, then trace which tables, functions, and identities each one can reach in practice, not just what its configuration claims.

Practical Controls to Implement

Runtime governance is the layer that sits above Unity Catalog's data permissions and evaluates what an agent is actually attempting to do.

Practical Controls to Implement

Establish agent identity, least-privilege tool access, approval workflows for sensitive tool calls, and unified audit logging that ties prompts, reasoning, and executed queries together.

What Genie and AI/BI Agents Actually Do

Databricks Genie is a conversational interface that translates natural-language questions into SQL, executed against tables and views governed by Unity Catalog. AI/BI, including Genie and dashboards, runs on the Databricks Data Intelligence Platform, which unifies governance across BI, ML, and SQL workloads through Unity Catalog. A Genie space is configured by a curator who defines which tables the agent can reference, along with sample queries and instructions that guide query generation. This means the agent's effective scope is shaped by two layers at once: the underlying Unity Catalog grants on the data, and the configuration choices made in the Genie space itself. Query execution runs under either the invoking user's identity or a service principal, and Unity Catalog permissions are applied based on whichever identity is executing the query. Which model a given workspace uses materially affects whether access control is enforced per user or pooled under a shared, broader identity.

Where Native Databricks Controls Stop

Unity Catalog provides real governance value: centralized access control, lineage tracking, data discovery, and audit logging for tables, views, and other objects. It also supports row-level and column-level security, dynamic views, and attribute-based access policies to restrict what data is visible to a given identity. These controls operate at the object level. They determine whether a query is allowed to touch a table or column, and they log that a query executed and what it accessed. What they do not do is evaluate the sequence of actions an agent takes, restrict which tools an agent may invoke before it invokes them, or provide a unified record of the agent's reasoning path alongside the resulting SQL. Databricks documentation does not describe a built-in mechanism for restricting or auditing individual tool calls beyond the SQL query and object-access logging that Unity Catalog already provides. For agents using the Mosaic AI Agent Framework, which supports registering custom tools including Unity Catalog functions, governance still relies on the same object-permission model applied to the underlying functions and data assets, not on a distinct policy layer for tool invocation itself.

Risks From Broad Query and Tool-Call Permissions

The practical risk is not that Unity Catalog fails to enforce permissions. It is that permissions are typically granted at a coarse grain relative to what an agent needs for any single interaction. A Genie space configured against a broad set of tables, or an agent workload running under a shared service principal with wide catalog access, gives the natural-language interface far more reach than any individual question requires. Because Genie's SQL generation is influenced by table descriptions, sample queries, and curator instructions rather than solely by underlying grants, ambiguous or adversarial prompts can generate queries that touch data outside the intended use case, even when technically permitted by Unity Catalog. As Mosaic AI agents add tool-calling capability, this exposure extends beyond static SQL generation to include whatever functions or actions have been registered and made callable. Without a way to restrict or approve individual tool calls at runtime, the effective blast radius of a single agent deployment can be considerably larger than the governance team assumes.

Data-Layer Permissions Are Not Agent-Layer Permissions

A useful distinction for governance teams is between what Unity Catalog controls and what a runtime governance layer would need to control. Unity Catalog answers the question of whether a given identity is allowed to read a given table or column. It does not answer the question of whether an agent should be permitted to invoke a particular tool, chain a particular sequence of actions, or act on a particular inferred intent within a session. These are agent-layer permissions, and they are not equivalent to data-layer permissions even when they overlap in effect. This is the specific gap that runtime policy enforcement addresses: evaluating and, where necessary, blocking an agent's action before it executes, based on the agent's identity, the tool or query it is attempting, and the context of the session, rather than relying solely on the static grants already applied to the underlying data. Trussed AI's runtime governance capabilities are built for this layer, providing agent identity, least-privilege enforcement, tool approval workflows, and audit logging that sits alongside Databricks' native Unity Catalog controls rather than replacing them.

Frequently Asked Questions

Does Unity Catalog govern what tools a Genie agent can call?

Unity Catalog governs access to data objects and, for Mosaic AI Agent Framework tools, the underlying functions those tools rely on. It does not provide a separate policy layer for approving or restricting individual tool calls at runtime.

Is Genie's query execution identity always the end user?

No. Genie and AI/BI workloads can execute under either the invoking user's identity or a service principal, depending on workspace configuration. This choice determines how Unity Catalog permissions are applied and materially affects least-privilege posture.

Can existing Databricks audit logs reconstruct an agent's full reasoning?

Audit logs capture discrete events such as query execution and object access. They do not natively unify the originating prompt, agent reasoning, and tool-call sequence into a single record, which typically requires external correlation.

What is the difference between data-layer and agent-layer governance?

Data-layer governance, provided by Unity Catalog, controls whether an identity can access a table or column. Agent-layer governance controls whether an agent should be permitted to take a specific action or invoke a specific tool, which is a distinct and currently unaddressed control point in native Databricks tooling.

Close the Gap Between Data Governance and Agent Governance

Unity Catalog governs the data. Runtime governance governs what your Genie and AI/BI agents are actually permitted to do with it.

Request a Demo