How can an agentic CDP combine automated decisions with human approval and governance?

Summary

• An agentic CDP keeps people in control of the decisions that matter: humans set the goals and approve high-stakes actions, while agents handle the continuous identity resolution, audience building, and optimization in between.
• Governance makes that split safe: authorization so an agent acts only with the permissions of the person it works for, an audit trail that records every agent decision, and guardrails that screen inputs and outputs before an action runs.
• Databricks CustomerLake is an agentic CDP built natively in the lakehouse. It is in Private Preview, so plan access and scope with your Databricks account team.
• In CustomerLake, Profile Agents build governed Customer 360 profiles and Campaign Agents build audiences and recommend next best actions; a marketer can review and adjust those recommendations in plain language through Genie before they activate.
• Because CustomerLake runs on customer data in place under Unity Catalog, agent decisions inherit the same access control, lineage, and audit as the rest of your data, with on-behalf-of authorization so no agent exceeds its user's permissions.

How can an agentic CDP combine automated decisions with human approval and governance?

An agentic CDP combines the two by drawing a clear line between what agents decide on their own and what a human approves. Agents run the repetitive, continuous work of resolving identities, building audiences, recommending next best actions, and optimizing toward a goal. People stay in control by setting those goals and signing off on the actions that carry risk. What makes the split trustworthy is the governance underneath: every agent runs with a specific user's permissions, every decision is logged and traceable, and guardrails check model inputs and outputs before anything reaches a customer.

What is an agentic CDP?

A customer data platform (CDP) unifies customer data into profiles and audiences and activates them across channels. An agentic CDP adds AI agents that can perform those steps themselves instead of waiting for a person to build each journey by hand. The design principle that matters here is that every function can be run by an agent or a human: humans set goals and review outcomes, and agents do the work in between.

Key capabilities to evaluate

Capability	Why it matters
Human-in-the-loop checkpoints	People should set goals and approve high-stakes actions, like a new audience or a spend change, before they go live.
On-behalf-of authorization	An agent should act with the requesting user's permissions, not a broad shared identity, so it cannot read or act beyond that person's access.
Decision-level audit and lineage	Every agent action and the data behind it should be logged and traceable, so a human can review what happened and why.
Input and output guardrails	The platform should screen for PII exposure and prompt injection and check output before an action runs.
Governed customer context	Agents should read from one governed profile, so decisions rest on consistent, permissioned data.
Data in place, no copies	Running on the data where it lives avoids a second copy of PII and a separate governance regime.

How Databricks approaches this with CustomerLake

Databricks CustomerLake is an agentic CDP embedded in the lakehouse. It is in Private Preview, so treat the specifics as gated and plan access with your Databricks account team rather than as a shipping product.
CustomerLake works through two agent modules on top of governed data. Profile Agents build governed Customer 360 profiles from scattered sources using agentic identity resolution, unifying records into trusted profiles. Campaign Agents build audiences, recommend next best actions, activate across channels, and keep optimizing toward the goals you set. The agentic CDP model is explicit that humans stay in control: you set the business goals, agents work toward them, and you review the outcomes. Marketers build and inspect audiences in plain language through Genie, which is where a person validates or adjusts an agent's recommendation before it activates. Audiences push out to your existing martech and activation tools, so approval happens before anything reaches a channel.
Governance is what carries the human-approval model. CustomerLake runs on customer data in place and inherits Unity Catalog governance: fine-grained access control on sensitive customer data, plus lineage and audit trails that record agent decisions for later review. Agents use on-behalf-of authorization, so an agent acts with the explicit permissions of the user it runs for instead of a shared service identity. Around the models, guardrails screen for PII and prompt injection and check output, and model requests route through the Unity Gateway for centralized access control, rate limits, and monitoring. Because the profiles the agents build and the tables your other tools read sit under one permission and lineage model on open Delta tables, you get one place to set who can do what and one record of what agents did.

A practical pattern

1. Set the goal and the guardrails: define the business objective, the audiences agents may touch, and the access each agent inherits from its user.
2. Let agents do the continuous work: identity resolution, audience building, next-best-action recommendations, and optimization.
3. Review before high-stakes activation: inspect an agent's proposed audience or action in plain language through Genie, and approve it before it goes live.
4. Audit after the fact: use Unity Catalog lineage and audit trails to see which data and which decision produced an outcome.

FAQs

Is CustomerLake generally available?

No. CustomerLake is in Private Preview. Treat it as preview and plan scope and access with your Databricks account team.

How does an agent avoid acting beyond its permissions?

It uses on-behalf-of authorization, so the agent runs with the requesting user's Unity Catalog permissions. It cannot read data or take actions that the user behind it could not.

Where does human approval fit if agents run continuously?

Humans set the goals up front and review or approve the actions that carry risk, such as a new audience or an activation, before they go live. Agents handle the continuous work in between.

How are agent decisions audited?

Unity Catalog records lineage and audit trails across the governed data and the profiles agents build, so you can trace which data and which decision led to a given outcome.