The successful pattern: audit historical labels first, enable in suggested mode, pilot for four weeks, validate accept rate per field, switch high-accuracy fields to auto-set. The feature pays back fast on orgs with clean labels and never pays back on orgs that skip the audit.
- Audit historical case labels
Pull 500 historical cases per Type or Reason value. Confirm each is correctly classified. Fix mislabeled cases. The audit takes a week and is the highest-leverage step in the rollout.
- Confirm training data volume per field
Setup, Einstein Service, Case Classification. Check the training record count per field. Fields with fewer than 1,000 historical cases per value will produce noisy predictions; consider consolidating values or excluding the field from autoclassification.
- Enable Classification in suggested mode
Toggle classification on for the pilot fields. Mode Suggested. Wait for the model to train (usually under an hour).
- Pilot for four weeks and watch Insights
Pull the Insights report weekly. Watch accept rate per field per value. Catch values where predictions are consistently wrong; usually those are values whose historical labels were inconsistent.
- Switch high-accuracy fields to auto-set
Above 85 percent accept rate, auto-set is safe and unlocks the routing and SLA benefits. Below 85, leave in suggested or revisit the historical labels.
- Wire routing rules to use the predicted values
Once auto-set is live, update assignment rules and SLA branching to depend on the predicted Type, Reason, or Priority. The downstream value compounds when routing matches the predicted intake.
- Schedule the monthly Insights review
Pull Insights monthly. Track accuracy drift, especially on values whose underlying patterns shift seasonally. Trigger retraining when drift exceeds 5 percentage points.
Per-field toggle for which picklists Classification predicts. Common picks: Type, Reason, Priority, Sub-Type.
Suggested shows the value for agent acceptance; auto-set writes it on case create. Start with Suggested.
Record type, age, owner filters that scope the historical cases the model trains on. Useful when older cases used different labels.
Weekly by default; can be triggered immediately after a label audit or data shift.
Einstein Case Classification permission that gates agent visibility into predicted values.
- Inconsistent historical labels produce inconsistent predictions. The audit before activation is the highest-leverage step; skipping it produces a feature that never earns trust.
- Imbalanced training data (one value dominates) produces a model that defaults to the majority and rarely predicts the minority. Consider consolidating values or excluding the field.
- Auto-set mode without piloting suggested produces wrong values that flow into routing and SLAs. Validate accept rate before enabling auto-set.
- Predicted values write as the running user, not "Einstein". Record history does not flag Classification-written values explicitly.
- The 85 percent accept rate threshold is the right gate for auto-set. Below it, the feature creates downstream noise faster than it saves time.