Building a Prediction Builder model takes 15 to 30 minutes for the configuration plus a few hours for the platform to train and score the model. The output is a deployable prediction admins can reference from Flow, Apex, or Lightning page.
- Confirm licensing
Einstein Prediction Builder is included with Service Cloud or Sales Cloud Enterprise. Confirm the org has the relevant license.
- Open Einstein Prediction Builder
Setup, Quick Find Einstein Prediction Builder, click the link. The page lists existing predictions. Click New.
- Pick the object and prediction field
Choose the object (Case, Opportunity, custom object). Choose the field to predict (a checkbox, picklist value, or numeric field).
- Define the example set
Configure which records count as positive examples (Closed Cases with Resolution = Refund) and which should be excluded (test cases).
- Set predictors
Pick which fields the model can use as predictors. Exclude leak variables and any field that should not influence the prediction.
- Build and review the model
Click Build. Wait for the platform to train and score the model. Review the quality score: Insufficient, Acceptable, Good, Excellent.
- Deploy the model
Click Deploy. The prediction becomes a Prediction Definition referenceable from Flow, Apex, and Lightning App Builder.
- Surface the prediction on the record page
Lightning App Builder, edit the record page, drag the Einstein Predictions component, configure the Prediction Definition reference.
The checkbox, picklist, or numeric field the model predicts. Binary or regression.
The historical records used as positive examples for training the model.
The input fields the model uses to predict the target. Excluded fields are not considered.
Platform-assigned model quality rating: Insufficient Data, Acceptable, Good, Excellent.
Simple explanation of which variables drive each prediction, shown inline on records.
Automatic retraining schedule, default 7 days. Adjustable per prediction.
- Prediction Builder is single-object only. Multi-object joins require Einstein Discovery instead.
- Insufficient Data means the org does not have enough historical examples. The platform needs hundreds to thousands of records of each outcome class for a useful model.
- Leak variables (fields only known after the outcome) produce great-looking models that fail in production. Audit the predictor list before deployment.
- Bias concerns apply. Fields that proxy for protected characteristics should be excluded unless explicitly justified.
- The 7-day refresh cadence is automatic. Models retrain whether the admin wants them to or not. Adjust the cadence for predictions where stability matters more than freshness.