The prediction that gets adopted has three properties: it predicts something the user cares about, it is accurate enough to bet on (above 70 percent), and it shows its work through Top Predictors. The prediction that gets ignored fails on at least one. Plan for all three from day one.
- Pick a target the user has a decision around
The prediction must inform a specific decision (which lead to call first, which case to escalate, which customer to renew). A prediction the user does not act on is academic and gets ignored.
- Pick the right source for the model
Built-in Einstein if the target is one of the canned ones (Lead Scoring, Opportunity Scoring). Prediction Builder for custom targets. Einstein Discovery if you also need narrative explanations. Tableau if a data team will own model maintenance.
- Audit training data for leakage and labels
Exclude any feature filled in after the outcome was decided. Confirm labels are consistent. Both audits matter; both are skipped more often than they should be.
- Train and review accuracy on a holdout set
Above 70 percent accuracy is a deployable threshold for most decisions. Below 60 percent is rarely worth shipping. Between 60 and 70 is a judgment call; pilot with one user team to see if the lift is real.
- Add the prediction field to list views and record pages
Without the field on relevant list views, users cannot sort by prediction. Surface it on the record page near where the decision is made.
- Add Top Predictors to the record page
Top Predictors is the explainability surface. A prediction without it reads as a magic number users distrust; with it, the prediction becomes a conversation about the underlying signals.
- Schedule monitoring and a retraining cadence
Weekly accuracy review for new predictions, monthly for stable ones. Retrain when accuracy drops more than 5 percentage points from baseline. Without monitoring, drift goes unnoticed for quarters.
Probability, score, classification, or regression. Drives downstream consumption pattern (threshold, branching, arithmetic).
Built-in Einstein, Prediction Builder, Einstein Discovery, Tableau, or external connector. Drives setup work and maintenance burden.
The explainability component. Surface on the record page; without it, predictions read as magic.
How often the model retrains. Weekly or monthly typical; depends on how fast the underlying business shifts.
Production-quality threshold per use case. 70 percent is a reasonable floor for most decisions; safety-critical decisions need higher.
- Data leakage in training features makes a model look brilliant in evaluation and fail in production. Audit features for "was this knowable at prediction time" before evaluating accuracy.
- Predictions without Top Predictors read as magic. Users distrust magic; the small layout change to surface Top Predictors has outsized adoption impact.
- Predictions stale as the model drifts and the underlying world changes. Monitoring catches drift; without monitoring, the model degrades quietly.
- A prediction with 80 percent accuracy is wrong 20 percent of the time. Downstream automation that treats predictions as certain creates incidents.
- Picking the wrong AI shape (using an LLM to score, or a predictive model to draft text) is slower, more expensive, and less accurate than picking the right tool.