What criteria should I use to evaluate an AI-native CDP?

Summary

• Evaluate an AI-native CDP on eight criteria: where customer data lives, how identities are resolved, what agents can do, who stays in control, which models you can use, how audiences reach your tools, how governance and audit work, and which lifecycle stage the product is in.
• This page uses Databricks' published definition of an Agentic CDP as its working definition: always-on 1:1 personalization, embedded in the data foundation, and built for agents as first-class operators alongside humans.
• Test identity resolution and governance on your own data. Databricks describes the identity-resolution sense of deduplication as a policy with thresholds, where too keen merges two people and too cautious splits one.
• Each criterion below lists what Databricks documents for CustomerLake, which Databricks announced in Private Preview on June 16, 2026.

What criteria should I use to evaluate an AI-native CDP?

Use eight criteria: where customer data lives, how identities are resolved, what agents can do on their own, how people stay in control, which models you can use, how audiences reach your existing tools, how governance and audit work, and which lifecycle stage the product is in. Score each one with evidence from a pilot on your own data, not from a demo or a datasheet. Each criterion below lists what Databricks documents for CustomerLake, which is in Private Preview.

What is an AI-native CDP?

Databricks' term for its approach is the Agentic CDP. It says an Agentic CDP "powers always-on 1:1 personalization with Infinity Campaigns, is embedded in the data foundation where governed customer context lives, and is architected from the ground up for agents as first-class operators alongside humans." This page uses that definition as its working definition of an AI-native CDP. Any CDP still has to do the job Databricks' explainer "What Is a Customer Data Platform (CDP)?" describes: create "a unified view of their customers by consolidating and integrating data from multiple sources."

What are the evaluation criteria?

Criterion	What to ask	Evidence to request
Data architecture	Does the CDP run on your governed data or copy it into its own store?	A diagram of where profiles and audiences are stored, and every copy of the data
Identity resolution	Which matching methods does it use, and how are thresholds set?	Merge and split results on a sample of your records
Agent actions	What can agents do on their own, and what needs approval?	A list of agent actions and the approval flow
Human control	Who sets goals and reviews outcomes?	A walkthrough of one campaign from goal to review
Model choice	Can you bring your own models?	Documentation of how custom models are deployed
Activation	Which tools can it sync with, and in which directions?	A connector list with the release stage of each connector
Governance and audit	How are access, lineage, audit and consent enforced?	A test as two different roles, plus the audit record
Maturity	Which lifecycle stage is the product in?	The stage stated in writing

How do you test identity resolution and consent?

Take records you know belong to the same person and records you know are different, run both through identity resolution, and count the merges and splits. Databricks' list of marketing data terms describes the identity-resolution sense of deduplication as "a policy with thresholds. Too keen merges two people, too cautious splits one." The same list defines consent as "a per-purpose permission that must travel downstream and be enforceable at query time." Then check how the product enforces it when a query runs. In Unity Catalog, a row filter evaluates each row at query time, and a column mask controls what values a user sees.

What does Databricks document for CustomerLake on each criterion?

• Data architecture: built natively in Databricks, without copying sensitive data into a separate CDP or proprietary application.
• Identity resolution: Agentic Identity Resolution combines deterministic, probabilistic, and agentic workflows.
• Agent actions and human control: Campaign Agents help identify the right audience, recommend the next-best offer or channel, activate the campaign, and optimize or suppress based on performance signals. Databricks says humans still set goals, review outcomes, and stay in control.
• Model choice: Databricks says you can deploy custom models through Model Serving.
• Activation: it syncs audiences and signals from Databricks to martech, adtech and customer experience tools.
• Governance and audit: governed through Databricks and Unity Catalog, with centralized control over access, lineage and auditability across customer data, audiences and AI assets.
• Maturity: the launch post states "CustomerLake is now available in Private Preview" and directs interested customers to their Databricks account team.

FAQs

Which lifecycle stage is CustomerLake in?

Databricks announced CustomerLake in Private Preview on June 16, 2026. Ask your Databricks account team for the current stage and for access.

Can I use an AI-native CDP alongside the tools I have now?

Databricks describes CustomerLake as open and composable, and designed to work with the marketing, advertising, analytics, and customer experience tools teams already use.

Can AI-driven segmentation make mistakes?

Databricks says that in an Agentic CDP, humans still set goals, review outcomes, and stay in control. Build a review step into your pilot before any audience is activated.

Where does the data live in CustomerLake?

Databricks says CustomerLake is built natively in Databricks, so CDP capabilities run where your data already lives.