The feature ships in minutes. The Knowledge audit that makes the recommendations useful takes one or two days. Skipping the audit is the rollout failure pattern. The teams that ship the audit first hit 30 to 50 percent acceptance in week one and never look back; the teams that skip it hit 15 percent and quietly stop using the feature.
- Audit Knowledge before turning the feature on
Pull all Knowledge articles by view count. Archive articles older than 18 months unless they are timeless. Rewrite titles on the top 50 most-viewed articles to be plain-language descriptions of the question they answer, not internal codes.
- Enable Einstein Article Recommendations
Setup, Einstein Service, Article Recommendations. Toggle on. Pick the case record types it applies to. Service Cloud edition with Einstein is required.
- Add the component to the case record page
Open Lightning App Builder, edit the Case Record Page. Drag the Article Recommendations component into the sidebar. Save and activate.
- Verify agent access to Knowledge
Agents need Read on Knowledge and access to the data categories the case maps to. Without both, the recommendations list renders empty even when articles match.
- Pilot with one team for two weeks
Pick five to ten agents. Watch acceptance rate daily. Pull samples of accepted and ignored recommendations to spot Knowledge issues that affect quality.
- Train agents to accept even when they only reference the article
The acceptance signal feeds back into the model. An agent who reads an article but does not formally accept it gives the model no signal. Brief them on this loop in the rollout meeting.
- Schedule the monthly Einstein for Service Insights review
Pull the acceptance rate per article report monthly. Rewrite or retire the bottom 10 articles by acceptance rate. The cadence is what compounds value.
Which case record types Article Recommendations applies to. Default is all; can be restricted to specific record types during pilot.
How many recommended articles to show. Default 5; tune to placement space available on the case page.
Which actions (Attach, Email, Insert) are exposed per recommendation. Default all three.
Which Knowledge categories are eligible for recommendations. Defaults to all accessible to the agent.
Sidebar, tab, or main column on the case record page. Sidebar is most common but the choice affects agent visibility.
- Stale Knowledge produces stale recommendations. The model cannot make a bad Knowledge base sharp; it can only return what is there.
- Empty recommendation lists almost always trace to permissions, record type, or empty categories rather than the model. Walk through those three before investigating model behavior.
- Acceptance rate is the feedback loop. Agents who reference articles without accepting starve the model of signal and the recommendations degrade over time.
- Article titles drive ranking more than agents expect. Renaming the top 50 most-viewed articles to clear plain-language titles often produces visible accuracy improvements within a week.
- The retraining cadence is on Salesforce's schedule. A new article appears in recommendations within hours; a rewritten title takes effect on the next retrain, which can be days.