Setting up Einstein Case Routing is a guided build inside Service Cloud Einstein. The work upstream (clean historical data, stable queue structure) matters more than the build itself.
- Confirm volume against the threshold
Pull a count of closed cases over the past 12 months grouped by intended routing destination. Confirm at least 400 per destination and 1,000 total. Without the volume, the feature will refuse to build.
- Stabilize the routing destination set
Audit the queue or skill list. Merge duplicates, retire unused destinations, document the final list. Training a model on a queue structure that is about to change wastes the build.
- Run the Setup wizard
Setup, Service Setup, Einstein Case Routing. The wizard prompts for the target field (the case field the prediction writes to) and the feature fields (Subject, Description, and any custom fields).
- Build the model and review the model card
Launch the build. After completion, open the model card. Check overall accuracy and per-destination accuracy. If one destination has very low accuracy, the model may have insufficient examples for it.
- Wire the prediction into Omni-Channel or flow
Configure Omni-Channel skills-based routing to read the Einstein prediction field, or build a flow that branches on the field and assigns the case. Test with sample cases covering rich, sparse, and edge contexts.
The case field where the prediction is written. Typically a custom Recommended Queue or Recommended Skill picklist.
The case fields the model uses as inputs. Subject and Description by default, plus optional custom fields with predictive signal.
The minimum prediction confidence required to write the recommendation. Default starts conservative; tune based on misroute tolerance.
The historical period the model trains on. Default 12 months. Shorter windows react faster to change; longer windows give the model more data.
Default weekly. Sufficient for most teams; high-volume orgs may want daily refreshes if the data shifts quickly.
- Queue structure changes invalidate the model. Plan to rebuild after any queue merge, rename, or new addition.
- Per-destination accuracy can vary widely. A model that is 85 percent overall may be 95 percent on common destinations and 50 percent on rare ones.
- Confidence threshold set too low produces misroutes that erode agent trust. Start conservative; loosen only after observing the first round of accuracy data.
- Predictions are written at case creation or update. A case that already routed via standard rules will not retroactively re-route when Einstein later predicts a different destination.
- Long, descriptive Descriptions improve accuracy more than threshold tuning. Encourage agents and customers to provide context-rich case bodies.