How can I use predictive models directly in audience segmentation?

Summary

• Predictive scores (propensity to convert, churn risk, predicted lifetime value, offer responsiveness) become audience signals the moment a model's output lands in a governed table next to your customer data. A segment turns into "customers whose churn probability is high and whose predicted value is in the top tier," not a static hand-written rule.
• The path is standard: train the model, register it in Unity Catalog, then score customers in batch into a governed predictions table, or serve scores at low latency through the online feature store powered by Lakebase for in-the-moment decisions.
• Because the model, its scores, and the audiences built from them sit in one catalog, they share the same permissions, lineage, and audit as the source customer data.
• Score in place. Copying data to a separate segmentation tool creates shadow copies that drift from the source and fall outside governance.
• In Databricks CustomerLake, Campaign Agents build audiences from natural-language requests using customer attributes, behavioral signals, and predictive models together. CustomerLake is currently in Private Preview.

How can I use predictive models directly in audience segmentation?

You use a predictive model in segmentation by turning its output into a governed attribute that segment logic can filter on. Train the model, register it in Unity Catalog, score your customers, and write the results (a churn probability, a lifetime-value estimate, an offer-response score) back to a governed table alongside the rest of the customer data. A segment then becomes a query over those scores, for example "customers whose churn probability is high and whose predicted value sits in the top tier," and it refreshes as the model rescores rather than staying a rule someone maintains by hand.

What does predictive segmentation mean?

Traditional segmentation groups customers by descriptive attributes: geography, past purchases, demographics. Predictive segmentation groups them by a model's forward-looking estimate: how likely someone is to convert, to churn, or to respond to an offer, and how much they're worth over time. The model scores every customer, and those scores become attributes you segment on, the same way you'd segment on a region or a purchase count. You target on what a customer is likely to do next, not only on what they've already done.

What to look for when models drive segmentation

Capability	Why it matters
Model output stored as a governed attribute	A score is only useful for segmentation if it lands next to customer data and segment logic can filter on it.
Scoring in place, no data copied out	Exporting to a separate segmentation tool creates copies that drift from the source and escape governance.
One catalog for model, scores, and audiences	Permissions, lineage, and audit hold end to end, from the training data through the score to the audience.
Fresh scores on the right cadence	Batch rescoring fits most segments; real-time decisions need scores and features served in the moment.
Bring your own models	Your propensity, churn, and value logic is a business asset, so the platform should run your models, not force generic scores.
Natural-language audience building	Marketers should be able to express a predictive segment without writing SQL or filing a data-team ticket.

How Databricks approaches predictive segmentation

On Databricks the path is standard MLOps feeding marketing. Train a propensity, churn, lifetime-value, or conversion model and track it with MLflow, then register it in Unity Catalog so the model and the data it learned from share one governance layer. Score customers in batch and write the results to a governed predictions table, the simplest and most common pattern; that table becomes the source of truth other tools read, so dashboards, apps, and agents consume the same scores instead of calling the model directly. When a decision has to happen in the moment, deploy the model as a Model Serving endpoint and serve features and scores at low latency through the online feature store powered by Lakebase.
Databricks CustomerLake, an agentic CDP built into the platform, is where those scores become audiences. Campaign Agents build audiences and recommend next-best actions with access to customer attributes, behavioral signals, and predictive models in one governed place, so a marketer can describe a segment in plain language, such as "high-value subscribers likely to churn who haven't engaged recently," and the agent turns it into a data-backed audience. CustomerLake is currently in Private Preview, so treat the specifics as gated and plan scope with your account team. Because everything runs under Unity Catalog, the model, its scores, and the audiences inherit the same access control, lineage, and PII handling as the source data.

Use cases

• Churn-risk retention audiences. Score churn probability, then segment the high-risk, high-value customers for a retention offer.
• Propensity-to-convert targeting. Build an audience of the customers most likely to respond so spend goes where it converts.
• Lifetime-value tiers. Turn a value model into tiers that shape targeting and budget across campaigns.
• Offer-responsiveness segments. Score how likely each customer is to respond to a specific offer and target accordingly.

FAQs

Do I have to move data into a segmentation tool to use model scores?

No. Score in place and write the results to a governed table, so the model, scores, and audiences stay in one catalog instead of in copies that drift from the source.

Should I use batch or real-time scoring for segmentation?

Most segments are fine with batch rescoring into a predictions table. Use a Model Serving endpoint and the Lakebase-powered online feature store when a decision has to happen in the moment.

Can I use my own models, or only built-in ones?

Use your own. Register your propensity, churn, or value models in Unity Catalog and their outputs become segment signals.

How does a predictive score become a segment?

The score is stored as an attribute on the customer record, so a segment is a filter over it, for example a churn probability above a threshold combined with a value tier.

Is CustomerLake generally available?

No. CustomerLake is currently in Private Preview. Plan access and scope with your Databricks account team.