Einstein Discovery
Einstein Discovery is the AutoML platform inside Salesforce that builds predictive models on the org's data without requiring data science expertise.
Definition
Einstein Discovery is the AutoML platform inside Salesforce that builds predictive models on the org's data without requiring data science expertise. The user supplies a dataset (a Salesforce object, a CSV upload, or a CRM Analytics dataset), picks a target variable to predict (Will the deal close? Will the case violate SLA? What is the expected lifetime value?), and Einstein Discovery automatically builds, evaluates, and deploys a regression or classification model. The output is a Story: a queryable model with explanations of which variables drive the prediction.
Einstein Discovery is the AI surface that bridges raw Salesforce data and operational decisions. Predictions can be embedded in record pages (showing the predicted score directly on the Account or Opportunity), referenced from Flow and Apex (using the prediction to drive automation), and called from Einstein Next Best Action strategies (ranking recommendations by predicted acceptance). The platform handles model retraining, drift detection, and explanation generation. Discovery sits inside the broader Einstein suite alongside Einstein Prediction Builder, Einstein Bot, Einstein Search, and Agentforce.
How Einstein Discovery builds and deploys predictive models
Stories: the unit of Einstein Discovery output
A Story is the central artifact in Einstein Discovery. It contains the trained model, the dataset used for training, the variable explanations, the model performance metrics, and the deployment configuration. Users open a Story to see Insights (what drives the predicted variable), Predictions (per-record predicted values), and Model Performance (accuracy, R-squared, confusion matrix for classification). Stories can be cloned, retrained on fresh data, and versioned. A Story is the AutoML output that users actually consume.
Building a Story: the AutoML workflow
Building a Story is a guided workflow. The user picks a dataset (from CRM Analytics, an uploaded CSV, or a Salesforce object), names the Story, picks the goal (maximize or minimize a target variable), excludes variables that should not be used as predictors (Account.Owner.Name is usually excluded as a leak), and clicks Build. Einstein Discovery runs a hyperparameter sweep across multiple model families (linear regression, gradient boosting, random forest), evaluates each with cross-validation, picks the winner, and produces the model. The whole process takes 5 to 30 minutes depending on dataset size.
Explainability: which variables drive predictions
Einstein Discovery is built with explainability in mind. Each Story exposes Insights: for each predictor variable, the model shows how it affects the predicted variable, in plain language. "Cases with Priority = High have an SLA violation rate 3.2x higher than average". "Accounts in Industry = Manufacturing close 45 percent of opportunities, compared to 28 percent across all industries". This is the difference between a black-box ML model and a Salesforce-usable model: the explanations let business users understand and trust the predictions.
Model deployment and runtime prediction
After a Story is built and approved, the user deploys it. Deployment creates a Prediction Definition: a referenceable model that can be called from Flow (Get Predicted Values from Story), Apex (sObject.predict()), Process Builder, or Einstein Next Best Action strategies. Each call sends record data to the deployed model, receives a prediction score, and optionally writes the score back to the record. Predictions can also surface inline on record pages through the Einstein Predictions Lightning component, which queries the deployed model on page load.
Model retraining and drift detection
Models degrade over time as the underlying data changes (sales cycles shift, customer behavior evolves, product mix changes). Einstein Discovery supports scheduled retraining: pick a frequency (weekly, monthly, quarterly), and the platform retrains the Story on the latest data on schedule. Drift detection compares the current model's predictions against actual outcomes, flagging when performance drops. The combination keeps deployed models honest without manual data science intervention.
Use cases: sales, service, marketing, custom
Einstein Discovery is general-purpose. Common deployments include: sales opportunity scoring (likelihood to close), case escalation prediction (likelihood to violate SLA), customer churn prediction (likelihood to churn next quarter), lead conversion scoring (likelihood to convert). Beyond these, any record-level outcome is a candidate. The same platform that predicts opportunity close probability can predict student dropout risk in an education org, employee attrition in an HR org, or claim fraud in an insurance org. The model family is the same; the use case varies.
Licensing and CRM Analytics dependency
Einstein Discovery requires CRM Analytics (formerly Tableau CRM) licensing. The two products share the dataset infrastructure: CRM Analytics datasets are the primary input to Einstein Discovery Stories. Customers buying Einstein Discovery typically buy a CRM Analytics-Growth or CRM Analytics-Plus license that bundles both. Standalone Einstein Discovery is not sold. The licensing cost is meaningful (tens of thousands per year for enterprise deployments) and should be planned alongside the broader AI strategy.
Building a predictive model with Einstein Discovery
Building an Einstein Discovery model is a guided workflow inside CRM Analytics. Pick a dataset, define a goal, build the Story, evaluate the model, deploy as a Prediction Definition, surface in Flow or Lightning page.
- Confirm licensing and prepare the dataset
Einstein Discovery requires CRM Analytics licensing. Prepare the dataset: a Salesforce object, an uploaded CSV, or an existing CRM Analytics dataset.
- Open Einstein Discovery in CRM Analytics
App Launcher, CRM Analytics, Einstein Discovery. Click Create Story.
- Pick the dataset and goal
Select the dataset. Pick the goal variable (Closed Won percentage, Time to Resolution, Customer Churn). Set the optimization direction (Maximize or Minimize).
- Configure predictors and exclusions
Exclude variables that leak the target (variables only known after the outcome) or that should not be used for business reasons (Owner name, gender, race).
- Build the Story
Click Build Story. Wait for the AutoML run. The platform evaluates multiple model families and picks the best performer.
- Review insights and performance metrics
Open the Story. Review the Insights (which variables drive the prediction) and the Model Performance (accuracy, R-squared). Confirm the model is trustworthy before deployment.
- Deploy the model as a Prediction Definition
Click Deploy. The platform creates a Prediction Definition referenceable from Flow, Apex, and Einstein Next Best Action.
- Surface predictions in the Salesforce UI
Lightning App Builder, edit the record page, drag the Einstein Predictions component, configure the Prediction Definition reference. Predicted scores render on the record.
The central artifact containing the trained model, dataset, insights, and deployment configuration.
The target the model predicts. Numeric (regression) or categorical (classification).
The input features used to make predictions. Excluded variables are not used.
The deployed model referenceable from Flow, Apex, and other Salesforce surfaces.
Plain-language explanations of which variables drive the prediction.
Configurable cadence (weekly, monthly, quarterly) for refreshing the model on new data.
- Einstein Discovery requires CRM Analytics licensing. Standalone licensing is not available. Plan the cost as part of the broader AI investment.
- Leaky predictors (variables only known after the outcome) generate models that look great in training but fail in production. Exclude them carefully.
- Bias concerns matter. Variables like gender, race, or zip code (a proxy for race in many US markets) should be excluded unless explicitly justified. Audit predictor lists for fairness.
- Model drift is real. Schedule retraining and review the drift detection alerts. A model that was 85 percent accurate at deployment may degrade to 65 percent in six months.
- The Einstein Predictions Lightning component queries the deployed model on page load. Heavy use across many records adds page-load latency. Consider batch prediction with stored scores for high-volume scenarios.
Trust & references
Straight from the source - Salesforce's reference material on Einstein Discovery.
- Einstein Discovery OverviewSalesforce Help
- Create a StorySalesforce Help
- Deploy a StorySalesforce Help
Hands-on resources to go deeper on Einstein Discovery.
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 does Einstein Discovery do?
Q2. Where does Einstein Discovery live?
Q3. Who is Discovery aimed at?
Discussion
Loading discussion…