Predictive Routing
Predictive routing in Salesforce is a Service Cloud capability that uses Einstein machine learning to send each case to the agent most likely to resolve it well, based on what happened with similar cases in the past.
Definition
Predictive routing in Salesforce is a Service Cloud capability that uses Einstein machine learning to send each case to the agent most likely to resolve it well, based on what happened with similar cases in the past. Instead of fixed if-then rules alone, the model learns from closed cases and predicts the values that drive assignment.
In practice it works through two linked features. Einstein Case Classification reads a case Subject and Description, then predicts fields like Priority, Reason, or Type. Einstein Case Routing takes those predicted values and feeds them into your existing assignment rules or skills-based routing so the case lands with the right person or queue.
How predictive routing decides where a case goes
Classification first, routing second
Predictive routing is not a single switch. It is the combination of Einstein Case Classification and Einstein Case Routing working in sequence. Classification is the prediction engine. It studies the text a customer writes in the Subject and Description, compares it to patterns in your closed cases, and recommends values for fields such as Priority, Reason, Type, or a custom picklist. Routing is the action layer. Once a field has a predicted value, Salesforce runs that case through the assignment rules or skills-based routing you already use, so the prediction actually changes where the work ends up. This split matters for how you plan a rollout. You can turn on classification, watch the recommendations against what agents actually pick, and only later let those predictions feed routing. The model improves as more cases close, since Einstein keeps learning from fresh outcomes. Treating the two pieces separately gives you a safe way to measure accuracy before predictions start moving live work.
What the model learns from
Einstein builds its model from your own historical cases, not a generic dataset. You point it at a set of closed cases and tell it which field to predict. The model then looks for statistical links between the words customers use and the value that field ended up holding when the case closed. Because the training data is yours, the predictions reflect how your team has actually categorized and prioritized work over time. Data volume drives quality here. Salesforce guidance is that Einstein needs at least 400 closed cases to build a model, and 1,000 or more is better. To predict a specific field, it needs at least 400 closed cases that have a value in that field. If your historical data is sparse or inconsistent, the model has little to learn from and predictions suffer. This is why data preparation, cleaning up blank fields and fixing mislabeled cases, comes before any setup work. Garbage history produces shaky predictions, and routing built on shaky predictions sends cases to the wrong place.
Where it differs from rule-based routing
Standard Omni-Channel routing relies on conditions you write by hand. You define queues, capacity limits, and skills, then cases flow to whoever matches and has room. That model is predictable and easy to audit, but it depends on assumptions. You are guessing which agents handle billing questions best, or which keywords signal a high-priority issue. Predictive routing changes the input, not the whole machine. The prediction fills in a field value that your rules then act on. So if Einstein predicts Priority as High for an angry-sounding case, your existing high-priority assignment rule does the rest. The benefit is that the categorization comes from real resolution history instead of a person's best guess. The tradeoff is transparency. A hand-written rule tells you exactly why a case routed somewhere. A model is harder to explain line by line. Many teams keep rule-based routing as the backbone and let predictions improve the fields that feed it, rather than replacing their logic wholesale.
Licensing and edition requirements
Predictive routing is not part of the base platform. The Einstein Case Classification and Einstein Case Routing features come with the Service Cloud Einstein add-on, which you buy on top of Enterprise or Unlimited edition. Without that license, you can still route cases with standard assignment rules and Omni-Channel, but the machine learning predictions are not available. There is a lighter entry point. A Try Einstein option lets you build a single classification model per app so you can evaluate the feature before committing to the full add-on. The paid Einstein for Service license raises that limit, allowing up to five models per app, which matters if you want separate models for different record types or business units. Plan your model count against this limit early. If you run a global support operation with very different case mixes by region, five models may go quickly, and you will need to decide which prediction problems are worth a dedicated model.
A worked example
Picture a support center that handles billing, technical, and account questions. Historically, agents manually set the Type field, and a senior team handles anything marked Critical. The team enables Einstein Case Classification and trains a model on two years of closed cases, predicting the Type and Priority fields. A new case arrives reading: "I was charged twice this month and need this fixed today." Classification reads that text and predicts Type as Billing and Priority as High, because similar wording in closed cases nearly always resolved that way. Those predicted values populate the case. The existing assignment rule, which sends High-priority Billing cases to the senior billing queue, now fires correctly without an agent touching the fields first. The case reaches a qualified person in seconds. Over the following weeks, every newly closed billing case feeds back into the model, so the next prediction is informed by the latest outcomes. The admin watches a report comparing predicted Priority against the value at close, and only widens automation once accuracy holds steady.
Measuring whether it actually helps
Predictive routing earns its keep only if it beats what you had. Before flipping it on for live assignment, capture a baseline: current resolution times, reassignment rates, and how often cases land in the wrong queue. Then run classification in recommendation mode, where agents see predictions but apply them manually, and compare the model's suggestions to the values agents choose. That comparison tells you the real accuracy on your data. Salesforce surfaces prediction performance so you can see how often the model matched the value the field held when the case closed. Use that to decide how much to automate. High agreement on a field means you can trust it to drive routing. Low agreement means keep it as a hint or retrain with cleaner data. Watch downstream metrics too, because a model can be accurate yet still route in a way that overloads one team. The honest test is whether resolution outcomes improve against your baseline, not whether the AI looks impressive in a demo.
Keeping the model healthy over time
A predictive model is not set-and-forget. Your case mix shifts as products change, new issues appear, and seasonal spikes come and go. Because Einstein retrains on closed cases, the model naturally absorbs new patterns, but only if the data it learns from stays clean. If agents start leaving the predicted fields blank or overriding them with sloppy values, the model learns from that noise and drifts. Build a light governance habit. Review prediction accuracy on a regular cadence, not just at launch. When you add a new product line or a new case Reason, expect a dip until enough closed cases accumulate for the model to learn the new category. Keep an eye on fields you depend on for routing, since a drop there has the most operational impact. If you run multiple models, retire ones that no longer reflect how the business works rather than letting stale predictions quietly misroute cases. Treating model maintenance as an ongoing operational task, similar to tuning assignment rules, keeps predictive routing trustworthy.
How to set up predictive case routing
Predictive routing is enabled by setting up Einstein Case Classification, then letting its predicted fields feed your existing routing. This is a high-level path; check current data requirements and your Service Cloud Einstein license before starting.
- Confirm the license and prep your data
Verify you have the Service Cloud Einstein add-on (or Try Einstein for one model). Then clean your closed cases so the field you want to predict has consistent, accurate values across a large enough sample.
- Create a classification model
In Setup, create an Einstein Case Classification model. Choose which closed cases it learns from and which field it predicts, such as Priority, Reason, or Type. Build the model so Einstein can train on your history.
- Review predictions before automating
Run the model in recommendation mode so agents see suggested values without them being applied automatically. Compare predictions against what agents and closed cases actually show to gauge accuracy.
- Let predictions drive routing
Once accuracy holds, allow the predicted field values to populate automatically. Your existing assignment rules or skills-based routing then act on those values to send each case to the right agent or queue.
The case field the model fills in, like Priority, Reason, Type, or a custom picklist. Pick fields your routing rules already depend on.
The set of historical closed cases the model learns from. Needs roughly 400 cases minimum (1,000+ ideal), with values present in the target field.
Whether predictions are shown as recommendations or applied automatically. Start with recommendations, then widen automation as accuracy proves out.
How many models you run per app. Try Einstein allows one; the full add-on allows up to five, useful for different record types or regions.
- Predictive routing is not a standalone setting. It is Einstein Case Classification feeding values into your existing assignment or skills-based routing.
- Thin or messy closed-case history produces weak predictions. Fix data quality before training, not after.
- Below roughly 400 closed cases with a value in the target field, Einstein cannot build a reliable model for that field.
- Accuracy can drift as your case mix changes. Review prediction performance on a regular cadence, not just at launch.
Prefer this walkthrough as its own page? How to Predictive Routing in Salesforce, step by step
Trust & references
Cross-checked against the following references.
Straight from the source - Salesforce's reference material on Predictive Routing.
Hands-on resources to go deeper on Predictive Routing.
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 powers Predictive Routing's decision about which agent should get a Case?
Q2. How does Predictive Routing differ from standard rule-based Omni-Channel routing?
Q3. Predictive Routing is a feature within which Salesforce product area?
Discussion
Loading discussion…