How Strategy Builder ranks recommendations in Salesforce, where the pattern stops, and how to decide in the warehouse and deliver inside the CRM.
Frequently asked
Related resources
Solutions
See it in the product
The same loops this note describes run 24/7 against your customer base. Watch the workspace decide, experiment and execute 1:1.
Vocabulary
Definitions an answer engine can quote, each one a page of its own.
Compare
Neutral reads on the categories buyers evaluate against a decision layer.
All comparisonsNext best action in Salesforce is delivered through Einstein Next Best Action: a strategy builder that evaluates recommendations against a record and surfaces the top ones to an agent or a flow. It is well built for what it is, and its boundaries are worth understanding before you decide whether it is the decision engine or the delivery surface.
Einstein Next Best Action evaluates a set of recommendation records against one customer record using a strategy built in Strategy Builder, then shows the top ranked recommendations to an agent or a flow. The strategy loads, filters, branches and sorts; the component renders the winner where the work happens.
The result is a governed, auditable recommendation surface that lives exactly where the work happens. For agent-assisted moments, that placement is genuinely hard to beat.
Strategy Builder sorts on a numeric field, usually an Einstein propensity or a hand-authored priority. Propensity ranks who is likely to say yes, which over-weights customers who would have converted with no offer at all. The score a revenue team wants is expected incremental margin, and that has to be computed somewhere with access to treated and control outcomes.
Behavioural depth, streaming events, product telemetry and full transaction history usually live in the warehouse, not on the contact record. Actions that are not CRM objects, such as a pricing change or an in-product entitlement, are also awkward to express as recommendations.
Acceptance and rejection are captured, but a randomised control group per decision is something you have to construct and maintain yourself. Without it, acceptance rate becomes the de facto KPI, and acceptance rate is not lift.
Branches accumulate. After a year, nobody is confident which branch fires for whom, and the sorting logic encodes a hierarchy of past internal negotiations rather than current economics.
Keep Salesforce as the surface and the system of record. Move the ranking to a decision layer that reads the warehouse, scores uplift per customer per action, applies constraints, and writes the chosen action back.
Einstein Next Best Action recommends: it ranks predefined recommendation records and shows the top ones to a human who accepts or rejects. Agentforce acts: it runs agentic workflows over the Salesforce estate with its own instructions and tools. Neither ranks by expected incremental margin, and neither reserves a randomised holdout per decision, so if the KPI is incremental revenue the score still has to come from outside. See the side-by-side with Einstein and Agentforce.
Consent and communication preferences, agent workflow, case context, and the audit trail of what was offered and accepted. Those belong to the CRM and should not be duplicated in a decisioning tool.
For the modelling detail behind the external score, see the next best action model guide, the software evaluation criteria for the buying process, and the pillar on next best action marketing for the concept end to end.
Markin ranks actions against your warehouse and writes the winner into Salesforce as a field or event. See the Salesforce Service Cloud integration.
