Tune which fields Einstein predicts and the data it learns from for the team's case types.
Einstein for Service automatically classifies incoming cases by predicting fields such as category, product, department, and priority. Classification accuracy depends on which fields Einstein predicts and the quality of the historical case data it learns from.
This guide walks you through selecting which fields to predict, auditing the training data for consistency, configuring the classification models, testing accuracy on recent cases, and scheduling retraining as case patterns change.
Show faster, more consistent service inside Salesforce with automation metrics.
See where the AI reduces effort and where classification or replies need tuning.
Resolve common requests end to end with a scoped AI agent that escalates the rest.
Identify the fields to predict. List the case fields your team uses for routing and reporting—category, product, department, and priority are typical choices.
Confirm the historical cases are labelled consistently for training. Review a sample of closed cases to check that each field is populated and uses consistent naming conventions. Einstein learns from historical data; inconsistent labels reduce prediction reliability.
Standardise labels if needed. If you find inconsistencies in historical case labels, update cases to use consistent values before training.
Configure and train the classification models on those fields. Set up a classification model for each field you identified, selecting the historical case data as the training dataset.
Test predictions on recent cases and review accuracy. Once training completes, test the model against recent cases and review the predicted values.
Retrain periodically as case patterns change. Schedule retraining to run as new labelled cases accumulate, keeping predictions aligned with current case patterns.
A support team at [Company Name] wants Einstein to predict category, product, and priority on incoming cases. 1. The admin reviews closed cases and identifies category, product, and priority as the fields to predict. 2. She reviews historical cases and finds inconsistencies: category appears as both [Value A] and [Value B]. She standardises the labels. 3. She configures three classification models: one for category, one for product, one for priority. 4. She trains each model on historical closed cases. 5. After training completes, she tests the models on recent cases and reviews accuracy. 6. She schedules periodic retraining and enables the models for classification.
Configure case classification to predict category, product, department, and priority, trained on the labelled case history, and set it to retrain as new labelled cases accumulate.
Get a complete, on-brand draft grounded in the team's knowledge to review and send.
Send each case to the correct queue automatically, removing manual redirection.