A next best action is only useful if the person receiving it can see why it was suggested. A score with no reasons gets doubted and ignored; a recommendation that carries its evidence gets actioned. Designing for explainability is therefore not a presentation choice, it is what makes the whole system usable, and in regulated markets it is what makes it defensible. This guide sets out the design, from the signals behind a suggestion to the suppression rule that can withhold it.
Executive summary
An explainable recommendation attaches its reasons to the suggestion: how recently the player was active, whether their net position is trending, how they behave after a loss, and the day, hour and format most likely to suit them. Those same signals indicate when not to act, so suppression is built into the recommendation rather than hoped for downstream. The design keeps a person in the loop, a recommendation is a reviewable, prepared proposal, not an automatic send, and records the decision either way. These are reference-design recommendations, not a description of an audited Bounty AI deployment.
Step 1: a score is a starting point, not an answer
Ranking players by a single number is easy and rarely enough. Two players can share a churn score for opposite reasons, one gone quiet after a long history, one who never really started, and the right next step differs completely. The score hides that; the reasons reveal it. So the first design decision is that a recommendation is a short, legible case, not a number: the signals a manager would have asked for anyway, travelling with the suggestion.
Getting this foundation right is what makes the later steps safe. The behavioural signal set, the timing and format logic, and above all the suppression rule that routes a concerning pattern to protective review instead of a reload offer, all build on the principle that a recommendation must be able to explain itself before it asks to be trusted. Each step below is shown with the product screens for the view it describes.