Setting up a predictive model in Salesforce usually means turning on a standard Einstein feature or building a custom one in Einstein Prediction Builder. The work is in the data, not the build button.
- Define the question and the target field
Pick the question in plain English (which leads will convert, which cases are urgent) and find the field on the object that holds the historical answer. The target field must already be populated on the historical records.
- Verify volume, freshness, and class balance
Run a report on the historical records. Confirm at least 400 examples per outcome class (or the published threshold for the standard feature). Confirm the records are from the past 12 months. Confirm the class balance is no worse than 95-5.
- Audit features for leakage
Walk the field list on the target object. Exclude any field that could only be populated after the outcome occurred. Document the excluded list.
- Train and review the model card
Launch the build. After completion, open the model card. Check AUC or accuracy. Check the feature-importance list for surprises (a leaked field tops the list almost every time).
- Wire predictions into automation and monitor
Surface the score on the page layout, build a flow that branches on the score, and create a report tracking prediction accuracy in production. Schedule a quarterly accuracy review.
Lead Scoring, Opportunity Scoring, Case Classification. Fastest path when the standard target matches the question.
Click-not-code custom predictive model on any object with any yes-no or numeric target. Same volume thresholds as standard features.
Train externally (Databricks, Vertex, SageMaker), register in Einstein Studio, surface scores in Salesforce. For predictions outside the standard scope.
How often the model retrains. Default is weekly for case features, monthly for lead and opportunity. Adjust based on how fast the underlying behavior changes.
The cutoffs that turn a score into an action. Configured in flow, workflow, or rule. Tune based on team capacity, not model perfection.
- Label leakage produces beautiful evaluation metrics and terrible production performance. Audit feature importance after every build.
- Below the volume threshold, the feature refuses to build. Confirm the data before promising a launch date.
- A model with 95 percent accuracy can still be useless if the threshold for action sends everything into the same bucket. Tune thresholds to match team capacity.
- Concept drift is silent. A model that worked at launch can be wrong six months later because customer behavior shifted. Schedule quarterly accuracy reviews.
- Predictive scores on a record do not retroactively update. A score from last week stays until the next prediction run, even if the underlying data changed.