Designing a strategy is less about Strategy Builder mechanics and more about deciding what the recommendations should mean. The steps below cover the design and the build.
- Define the audience and surface
Pick the user role and the page where the strategy will run (sales console opportunity page, service console case page, community portal). Different audiences need different strategies; do not try to serve all in one.
- Inventory the relevant recommendations
List the Recommendation records that could matter for this audience. Group them into thematic buckets (cross-sell, retention, escalation, education). The bucketing helps later when designing branches.
- Draft the branch logic on paper first
Before opening Strategy Builder, sketch the decision tree. When does each bucket qualify? What scores or fields drive the decision? Reviewing the sketch with the business owner avoids rework once the canvas is open.
- Build, version, test on the canvas
Translate the sketch into Strategy Builder nodes. Save as a new version. Use the Test tab against five to ten representative record IDs. Iterate until the output matches the intent.
- Activate and build the reaction report
Activate the strategy version. Build a Recommendation Reaction report grouped by recommendation. Review acceptance rate weekly and refine.
Brings recommendations or related records into the strategy. Most strategies start with a Load Recommendations node.
Routes the stream based on a condition. The if-then-else of Strategy Builder.
Sets or transforms per-recommendation values. Useful for computing dynamic acceptance labels or priority scores.
Calls a CRM Analytics or Einstein Discovery model and appends the score. Used for inline predictive branching.
Caps the output at the top N. Three is the practical maximum for a single component.
- Strategies do not log execution by default. Use the Test tab on the canvas; do not expect debug logs in production.
- Ties on the Prioritize node fall back to entry order, which is often non-obvious. Make priority expressions tie-breaking when it matters.
- Inline Predict nodes add latency at page load. Pre-compute the score on the record when the score is stable; reserve inline predicts for context-sensitive scoring.
- Strategies need version bumps for non-trivial changes. Editing the active version directly is technically possible and operationally risky.
- Stale branches surface stale recommendations. Schedule a quarterly walk-through of every active strategy.