Predictive Model
A predictive model is a machine learning model trained on historical data to forecast a future outcome, classify a record into a category, or score a record on a likelihood scale.
Definition
A predictive model is a machine learning model trained on historical data to forecast a future outcome, classify a record into a category, or score a record on a likelihood scale. In Salesforce, predictive models power Einstein Lead Scoring, Einstein Opportunity Scoring, Einstein Case Classification, Einstein Forecasting, Next Best Action recommendations, and any custom prediction built in Einstein Prediction Builder. The model reads features from the record (industry, source, amount, age, related-record counts) and returns a number or label that downstream code consumes.
Predictive models differ from generative models in what they output. A generative model produces free text. A predictive model produces a structured value: a probability between 0 and 1, a category label, or a number. That structured output makes predictive models easy to wire into existing automation. A flow can branch on a score. A workflow can route on a category. The hard work is upstream: collecting clean historical data, defining the right target field, and validating that the model actually learned the pattern instead of a confounding signal.
How predictive models work and where they sit in Salesforce
Classification, regression, ranking
Predictive models fall into three task types. Classification predicts a category: this lead will convert or not, this case is Type A or Type B, this opportunity will close won or lost. Regression predicts a number: how many days until this opportunity closes, what dollar amount this deal will land at. Ranking orders records by a score: which leads should the rep call first. Salesforce features map cleanly to one of the three. Einstein Lead Scoring is ranking, Einstein Case Classification is classification, Einstein Opportunity Scoring is classification with an underlying ranking score.
Features, target, and what the model actually learns
Every predictive model has features (the input columns) and a target (the column being predicted). The model learns a function from features to target by minimizing prediction error on the historical examples. In Salesforce predictive features, the feature list is chosen by the platform from the standard and custom fields on the object. The admin's job is to know which fields to exclude (leaky ones, irrelevant ones, sensitive ones) and which fields to add (formula fields that summarize related-record activity often help). The target field is what the team picks at setup.
Training, evaluation, deployment as a pipeline
Salesforce predictive features run a training pipeline on a schedule. The pipeline reads historical records, splits them into training and validation sets, trains the model on the training set, measures accuracy on the validation set, publishes the model card, and replaces the previous model in production. The admin sees the model card with metrics like AUC or accuracy and approves or rejects the new model. If accuracy degrades meaningfully from the previous version, the old model stays in production until the data issue is resolved.
Einstein Prediction Builder for custom targets
The standard predictive features cover the common targets (lead conversion, opportunity close, case type). For any other target, Einstein Prediction Builder is the click-not-code path. The admin picks an object, picks a yes-no field or a number field as the target, and the platform builds a custom predictive model with the same training pipeline. Prediction Builder predictions land on the record as fields and can drive any downstream automation. The model still needs the volume thresholds: at least 400 examples of each target outcome.
Scores, thresholds, and how predictions drive action
A prediction is rarely the final action. The score lands on the record, but the action lives in flow, workflow, or rule. The typical pattern: score above 75 routes to a fast-track queue, score between 25 and 75 stays in the normal queue, score below 25 routes to a nurture sequence. Picking the thresholds requires a business decision about how to spend the team's capacity. A model that predicts perfectly is worth nothing if the thresholds send everything to the same queue.
When a predictive model goes wrong
Predictive models fail in three predictable ways. Label leakage produces models that score great in evaluation and terribly in production because the leaky feature is not populated at prediction time. Class imbalance produces models that predict the majority class for everything, technically accurate, operationally useless. Concept drift produces models that were right at launch and wrong six months later because the underlying behavior changed. Each failure has its own audit. A model card showing 95 percent accuracy without a feature-importance review is a red flag, not green.
Predictive versus generative, and when to use which
Generative AI is exciting, but for many problems predictive is the right tool. If the question is which lead should the rep call first, that is a ranking problem and a predictive model is faster, cheaper, and more measurable than asking an LLM. If the question is what should the rep say on the call, that is generation. Use predictive for scoring, classifying, and ranking. Use generative for drafting, summarizing, and explaining. Stacking them is fine. A predictive model scores the lead, a generative model drafts the outreach email.
How to set up a predictive model in Salesforce
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.
Trust & references
Cross-checked against the following references.
- Einstein Prediction BuilderSalesforce Help
- Einstein for SalesSalesforce Help
Straight from the source - Salesforce's reference material on Predictive Model.
- Einstein platform overviewSalesforce Help
- Einstein StudioSalesforce Help
About the Author
Dipojjal Chakrabarti is a B2C Solution Architect with 29 Salesforce certifications and over 13 years in the Salesforce ecosystem. He runs salesforcedictionary.com to help admins, developers, architects, and cert/interview candidates sharpen their fundamentals. More about Dipojjal.
Test your knowledge
Q1. What is a Predictive Model in Einstein?
Q2. What tools deploy predictive models?
Q3. When are predictions most valuable?
Discussion
Loading discussion…