How can I use existing machine learning models directly in marketing campaigns?

Summary

• You don't rebuild your models to use them in campaigns. Register an existing model in Unity Catalog, deploy it through Databricks Model Serving or score it in batch, and its predictions become governed signals a campaign can target on.
• The reliable pattern is Model Serving to governed tables and features: write a model's outputs (a churn or propensity score) to a governed Delta table, and let campaign tools and agents read those outputs rather than calling the model directly.
• Databricks CustomerLake is an agentic CDP embedded in the lakehouse (Private Preview). Its Campaign Agents work from customer attributes, predictive models, and business rules held under governance to build audiences and recommend next best actions.
• Real-time and batch both work: query a low-latency endpoint for in-session decisions, or score on a schedule for planning and segmentation, with outputs kept under Unity Catalog governance.
• Because the model, its outputs, and the customer data live under one catalog, marketing acts on the same governed scores your data team produced, with shared permissions, lineage, and audit, and no data copied out to score it.

How can I use existing machine learning models directly in marketing campaigns?

Put the model's output where the campaign can read it, under the same governance as your customer data. The models a data science team already trains, propensity, churn, lifetime value, next best action, and recommendation, become useful to marketing when their predictions land in governed tables and features that campaign tools and agents can act on. On Databricks, you register an existing model in Unity Catalog, deploy it through Model Serving as a real-time endpoint or score it in batch to a governed table, and those outputs become the signals that build audiences and shape decisions. The model, its endpoint, and its outputs keep one set of permissions, lineage, and audit, so marketing acts on the same governed scores your data team produced.

What using a model in a campaign really requires

A model is only useful to a campaign if its predictions are current, governed, and reachable by the tools marketers use. That means a clear path from a trained model to a score a campaign can target on, without exporting customer data to a separate tool.

Capability	Why it matters
Bring your own models	Your propensity, churn, and lifetime-value logic is a business asset; a marketing platform should run it, not force generic scores
One governance layer for models and data	When the model, its outputs, and the customer data share governance, permissions and lineage hold end to end
Scores written to governed tables	Predictions in governed Delta tables give campaigns one clean, auditable artifact to read from
Real-time and batch paths	Some decisions need a live score at request time; others are fine scored on a schedule
Outputs usable as audience and decision signals	A score matters only if it can define an audience or inform the next best action
No data copied out to score it	Scoring in place avoids shadow copies that drift from the source and escape governance

How Databricks approaches this with CustomerLake and Model Serving

Databricks CustomerLake is an agentic CDP embedded natively in the lakehouse. It is in Private Preview, so treat specifics as gated and plan scope with your Databricks account team. Its Campaign Agents build audiences, recommend next best actions, and activate across channels, and they work from customer attributes, behavioral signals, predictive models, and business rules held under governance.
The path for an existing model is grounded in ordinary Databricks architecture, not a special hook. You register the model in Unity Catalog and deploy it with Model Serving, which supports common frameworks such as scikit-learn, XGBoost, and PyTorch plus custom logic wrapped as a pyfunc model. For live decisions you query the model as a low-latency REST endpoint; for planning and segmentation you score in batch and write the predictions to a governed Delta table. Features governed in Unity Catalog can be served to models and applications at low latency through the online feature store powered by Lakebase. Campaign Agents then read those governed predictions and features, rather than calling a model directly, so a churn score can select who lands in a retention audience and a propensity score can inform the offer an agent recommends. Because everything sits under one catalog, the model, its outputs, and the customer data share permissions, lineage, and audit.

Use cases

• Propensity and churn into audiences. Score your model to a governed table and let a Campaign Agent build a high-value, high-risk audience from it.
• Custom next best action. A recommendation or next-best-action model's outputs inform the offer or channel an agent selects for each customer.
• Lifetime-value tiers. A lifetime-value model feeds tiers that shape targeting and budget, with no export to a separate tool.
• Real-time personalization. A served endpoint returns a live score at request time for in-session decisions.

FAQs

Do I have to rebuild my models to use them in campaigns?

No. Register the model you already have in Unity Catalog and deploy it through Model Serving, or score it in batch to a governed table. Campaigns and agents read the outputs.

How do CustomerLake's agents use a model's output?

Campaign Agents read governed predictions and features, such as a churn or propensity score written to a governed table, rather than invoking a model directly. Those scores define audiences and inform the next best action.

Does using a model mean copying customer data out?

No. Models, features, and customer data stay in the lakehouse under Unity Catalog, so you score in place instead of exporting to a separate scoring service.

Can I use both real-time and batch scoring?

Yes. Deploy a real-time Model Serving endpoint for live decisions, or score in batch to a governed table for planning and segmentation. Both keep outputs governed.

Is CustomerLake generally available?

No. CustomerLake is in Private Preview, so capabilities are still rolling out. Talk to your Databricks account team about access.