How can I apply the same data governance policies to marketers and AI agents?

Summary

• The reliable approach is to govern the data once, at the layer where queries run, so the same permissions apply whether a person or an agent asks. Marketers and marketing agents then read the same governed data through one policy set, not two.
• Make each agent run with the identity of the user or service that invoked it. The agent inherits that caller's grants, row filters, and column masks, and can never reach data the person behind it could not.
• Register models, agents, functions, and connections in the same catalog as your tables, and govern them with the same grant and revoke privileges plus attribute-based access control.
• On Databricks, Unity Catalog enforces access, masking, lineage, and audit at the engine level, so one control set covers Genie, notebooks, BI tools, and custom agents alike.
• CustomerLake, the agentic CDP built into Databricks, inherits that governance so marketing agents work on the same governed customer data as marketers. It is offered as a preview, so confirm current availability.

How can I apply the same data governance policies to marketers and AI agents?

Govern the data once, at the layer where queries actually run, instead of writing one rulebook for people and another for software. When permissions, row filters, and column masks live with the data and are enforced by the query engine, the same rules apply whether a marketer opens a dashboard or an AI agent calls a tool. The move that ties it together is running each agent under the identity of the user or service that invoked it, so the agent inherits exactly the access that principal holds and nothing more. Do that, and governing agents stops being a separate program. It becomes the governance you already run for people.

What does one governance model for people and agents mean?

It means a single set of controls, permissions, masking, lineage, and audit, bound to the data and applied to every consumer that touches it. A marketer building a segment and an agent choosing a next-best action pass through the same access checks, because the check happens where the data is served rather than inside each application. Governance that lives inside a marketing tool covers only that tool. Governance at the data layer covers every path to the data, including agents that did not exist when the policy was written.

Key capabilities to evaluate

Capability	Why it matters
Engine-level enforcement	Row filters, column masks, and grants applied where queries run cover people and agents identically, not app by app.
On-behalf-of-user identity	An agent runs with the caller's identity, so it inherits that user's permissions and cannot exceed them.
Models, agents, and tools as governed objects	Registering models, functions, and connections as catalog securables lets you grant and revoke their access the way you do for tables.
Attribute-based access control	A decision can weigh the asset, the data, the purpose, and the requester together, not only who is signed in.
Dynamic masking	Sensitive fields resolve to masked or full values based on the caller's entitlements, for humans and agents alike.
Unified lineage and audit	One query-ready record of who or what read each attribute, across analysts, dashboards, and agents.

How Databricks approaches it with Unity Catalog and CustomerLake

Unity Catalog is the governance layer, and it enforces at the engine level. The same row filters, column masks, and grants apply whether a human or an agent issues the query, so a marketer in a BI tool, an analyst in a notebook, and a custom agent all meet the same controls. Unity Catalog also governs AI assets: models, agents, functions, and connections are registered as securable objects and controlled with the same privileges and attribute-based access policies you use for tables and volumes, so access decisions can account for the asset, the data, and the purpose of a request, not just the signed-in user. Because agents run with the calling user's permissions, dynamic masking gives an agent the same view of a record its user would get, and system tables keep lineage and audit query-ready. See how AI governance works on Databricks.
CustomerLake, the agentic customer data platform built into Databricks, is where this pays off for marketing. It is embedded in the lakehouse and governed by Unity Catalog, so customer data does not move into a separate system and marketing agents inherit the same policies as the marketers they work alongside. A campaign agent assembling an audience sees only the attributes its user is cleared for, and every action lands in the same audit trail. CustomerLake is offered as a preview, with capabilities rolling out over time, so confirm current availability before you design around a specific module. For category background, see what is an agentic CDP.

Use cases

• Self-service audiences under one policy. A marketer describes a segment in plain language through Genie, and the agent that builds it reads only the attributes the marketer is permitted to use.
• Protecting sensitive fields from targeting. Attributes a campaign should not act on are masked at the engine, so an agent cannot target on data it has no right to see.
• One audit trail for people and agents. When a review asks who touched a customer attribute, the answer covers analysts, dashboards, and agents in the same record.

FAQs

Do AI marketing agents need a separate governance system?

No. If agents run with the caller's identity and the data is governed where queries run, agents inherit the same permissions, masking, and audit as people. You set policy once and it applies to both.

How do I stop an agent from using customer attributes it should not?

Grant access at the data layer and apply column masks and row filters there. Because the agent runs with the caller's permissions, restricted fields are masked or withheld automatically, without agent-specific code.

Can marketers still move quickly under shared governance?

Yes. Marketers keep querying and building audiences through familiar tools while the controls run underneath, so consistent policy does not mean slower work.

Is CustomerLake generally available?

CustomerLake is offered as a preview, with components rolling out over time. Confirm the current status of its modules before relying on them.