Predictive Routing
Predictive routing is a Service Cloud pattern where Einstein machine learning sets the case fields that decide an assignment, and your existing routing rules then act on those values.
Definition
Predictive routing is a Service Cloud pattern where Einstein machine learning sets the case fields that decide an assignment, and your existing routing rules then act on those values. Salesforce sells no product under this exact name. It assembles the behavior from two features running back to back.
Einstein Case Classification reads the text a customer wrote and predicts values for fields such as Type, Reason, or Priority. Einstein Case Routing runs your case assignment rules or skills-based routing against the saved values, so the prediction changes where work lands.
Both halves sit inside Einstein Classification Apps, configured from Setup under Einstein Classification. A no-cost Try Einstein tier gives you one model per app. The Einstein for Service add-on raises that to five and enables the automation setting routing depends on.
In plain English
“Think of a help desk inbox where every message has to be sorted before anyone works on it. A computer reads thousands of old messages and remembers how each one was sorted. When a new one arrives it guesses the label, and the normal sorting rules send it to the right desk.”
Worked example
A regional broadband provider trains a model on 18 months of closed cases, using Subject and Description as inputs and predicting Type and Priority. A customer files a case with the subject "Charged twice for August service" and a description about a duplicate debit. Einstein predicts Type as Billing and Priority as High, and both clear the confidence threshold the admin configured for automated values. The case saves with those values before any human opens it. Einstein Case Routing then fires the existing assignment rule for High priority Billing cases, dropping the record into the Billing Escalations queue, where Omni-Channel hands it to the next free rep holding the Billing skill.
How a prediction turns into an assignment
Two features, one handoff
Neither half earns its keep alone. A model nobody acts on just decorates records, and assignment rules reading a blank Type field dump everything in one catch-all queue. Order of operations is the point: the value lands during the save, so the rule sees a filled field. The handoff has a switch, and missing it is the most common way a rollout quietly does nothing. In the Einstein Classification settings you choose whether Einstein runs case assignment rules, skills-based routing, or both. Leave that blank and predictions still populate fields while every case sits where it started. The corollary bites as well: pick an option, define no underlying rules, and nothing moves.
What the model is allowed to learn from
Einstein trains on your own closed cases, and the eligibility bar is narrower than most admins expect. A case counts only if it was created in the past six months and carries at least one string field with content in it. The floor is 400 qualifying closed cases, with 10,000 given as the ideal. Per field the arithmetic runs separately. Predicting one field wants 400 closed cases holding a value there, spread across two or more distinct values. Picklists over 100 options train badly, because no single option collects enough examples. Thirty-three languages are supported. Sloppy history is the real disqualifier: duplicate cases, blank fields, values that overlap in meaning.
Three automation levels, and why only one routes
Every prediction field gets its own level of trust. Recommend Top Values shows the rep three candidates. Select Best Value pre-selects the winner when confidence clears your threshold, though a person still saves. Automate Value writes the field with no review, and needs the paid add-on. That last setting is the pivot point. Routing fires off the automatic save, so a field left on Recommend or Select never reaches Einstein Case Routing however accurate it is. Sensible teams start at Recommend Top Values, compare suggestions against what reps actually choose for a few weeks, then promote a field once agreement holds. The staging costs nothing and produces accuracy numbers from your own data.
Where it beats a hand-written rule, and where it does not
A rule is auditable. You can read it, point at the line that fired, and explain an assignment in one sentence. A model offers nothing comparable. It offers categorization drawn from how thousands of real cases resolved, rather than an admin's guess about which keywords signal urgency. So keep the rules and improve their inputs. Assignment rules and skills-based routing stay as the deterministic backbone. Einstein takes over the judgment call that used to happen when a rep read the case and set Type by hand. If a prediction goes badly, retrain or drop that field to a lower level. Your routing architecture stays untouched.
Licenses, ceilings, and model drift
None of this ships with the base platform, and five models per app becomes a real ceiling once you segment by region or product line. Admins also need View Setup and Configuration, Customize Application, Manage Profiles and Permission Sets, plus Edit on cases or Modify All Data. The longer risk is drift. A model reflects what your team did six months ago, so launching a product or adding a case Reason dents accuracy until fresh closed cases accumulate. Worse is feedback poisoning, where reps accept whatever Einstein guessed because correcting it is tedious. Review accuracy on a schedule and retire models that no longer match the business.
Setting up predictive routing without disrupting live assignment
Nothing routes until three things line up: enough clean history, a field predicted automatically, and a routing option selected in settings. Build them in that order.
- Audit your closed cases first
Count closed cases from the last six months carrying a Subject or Description and a value in the field you want predicted. Under 400 and you fix data before opening Setup.
- Create the model
From Setup, enter Einstein Classification in the Quick Find box, select Einstein Classification, then start a new model. Pick up to 30 input fields, at least one a string field, and keep prediction fields to one or two at first.
- Start every field at Recommend Top Values
Activate with predictions visible to reps but never automatic. Run it for a few weeks, then compare Einstein's suggestions against the value each case actually closed with.
- Promote a field, then turn on the handoff
Once agreement holds, move the field to Automate Value. Back in settings, choose to run case assignment rules, skills-based routing, or both, and confirm those rules are active.
Three candidate values appear and the rep picks one. Safest starting point, and it never routes on its own.
The highest-confidence value arrives pre-selected, but a person still saves the record. A reasonable middle ground.
Einstein writes the field with no review once confidence clears your threshold. Paid add-on only.
- Turn on both routing options and the two rule sets have to be mutually exclusive. Overlapping rules produce assignments nobody can explain.
- Select a routing option with no rules defined and every case stays exactly where it was.
- Trying this out in a sandbox is fine, but Einstein configuration does not copy between sandbox and production orgs. You rebuild it by hand.
Prefer this walkthrough as its own page? How to Predictive Routing in Salesforce, step by step
How organizations use Predictive Routing
Predicts Type and Priority on inbound web cases so billing disputes reach the escalations queue without a triage tier reading each one.
Runs a separate model per product line, letting classification set the Reason field that skills-based routing matches to specialists.
Drops back to Recommend Top Values during storm season, when case language shifts faster than the model retrains.
Trust & references
Cross-checked against the following references.
- Autofill Case Fields with Einstein Case Classification Apps (opens in new tab)Salesforce
- Run Assignment Rules for Einstein Case Classification (opens in new tab)Salesforce
- Prepare for Einstein Classification Apps (opens in new tab)Salesforce
- Salesforce Past Product & Feature Retirements (opens in new tab)Salesforce
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 writes and edits salesforcedictionary.com, published by KineticBit Inc., to help admins, developers, architects, and cert/interview candidates sharpen their fundamentals. More about Dipojjal.
Test your knowledge
+5 pts / dayQ1. Which Einstein Case Classification automation level actually triggers Einstein Case Routing?
Q2. An admin sets a field to Automate Value, confirms predictions are saving correctly, yet no case is being reassigned. What is the most likely cause?
Q3. What does Einstein need before it can predict values for one specific case field?
Discussion
Loading discussion…