Salesforce's own team says the model was never the problem: 0.83 told a seller nothing

Read the three inputs closely and one of them is doing less work than its billing suggests. Of signal, business logic and knowledge, only the signal comes out of a model ^[7]. Thresholds, priorities, operating rules and current business motions are things a revenue organisation decides and keeps rewriting. Institutional and field expertise about what a good response looks like sits with people who were never in the training pipeline ^[7]. Two of the three ingredients in the Next Best Action layer have owners outside the data science team ^[12].

That is what makes the 0.83 example useful rather than cute. The churn model was not wrong; by the authors' account it did its job, and the open questions were whether that number was high enough to act on, why the account was at risk, whether it outranked another renewal, and what move to make ^[4]. The team had years of XGBoost and gradient boosting work behind them on seller performance, pipeline health, skill gaps, attrition risk, product adoption and onboarding completion, all good at assessment and stopping there ^[5]. Retraining does not produce a priority order. Somebody in sales operations has to write one down.

The dashboard count deserves arithmetic the post does not do. Roughly 12,000 dashboards ^[2] at one minute each, without pauses, is 200 hours, or 25 eight-hour days of looking ^[13]. No consolidation programme survives contact with that number, which is presumably why consolidating them was listed as an option and then set aside ^[11].

Note what the fix does to the inventory. The stated destination is a single conversation the seller already has open ^[9]. Nothing in the writeup says the 12,000 dashboards were retired. The delivery surface changed; the estate behind it stays, which means two versions of the truth to keep aligned, and the newer one now asserts an expected impact it can be held to ^[8].

The supplied text ends mid-sentence, in the middle of a list of diagnostic questions, and carries no adoption figure and no measured outcome ^[14]. So read it as a design argument rather than a result. The part that transfers is cheap and awkward: take any prediction in production and ask what decision a user can make from that output alone ^[6]. Applied honestly, the answer is usually that somebody downstream has been supplying the missing interpretation for free, and their time never appeared in the model's cost line. Salesforce's engineers call the reusable pattern the separation of assessment from action, not the layer they happened to build ^[10].