A consultant ran a two-week proof-of-concept of Dynamics 365 Sales with a sales team, during which several sample and duplicate leads were entered purely to demonstrate features to stakeholders. Before turning on predictive lead scoring for the production rollout, the consultant reviews the historical data and finds those proof-of-concept leads are still present alongside real customer leads. What should the consultant do before enabling scoring?
Select an answer to reveal the explanation.
Short Explanation
A model only learns what you show it, and if part of what you show it is a handful of records that were only ever created to look good in a demo, it will happily learn the wrong lessons from them. Those records don't represent anything a real buyer did, so they need to come out of the historical data before the model starts training on it. There's no built-in filter that quietly strips out sample or duplicate records for you, turning on scoring later doesn't make old records disappear, and turning those demo leads into fake won deals only adds more noise instead of canceling it out. The safest move is a manual cleanup pass before flipping the switch.
Full Explanation
The correct answer is D. Predictive lead scoring learns from historical outcomes, so leaving demo and duplicate leads created purely to showcase the system during the proof-of-concept mixed in with genuine customer leads would teach the model from artificial patterns that have nothing to do with how real buyers behave, weakening its predictions once it goes live. Option A is incorrect because there is no automatic filter that recognizes and excludes duplicate-flagged leads from model training; duplicate detection and predictive scoring are separate features. Option B is incorrect because the model trains on historical data regardless of when scoring is enabled, so leads created earlier, including the proof-of-concept records, remain in scope unless someone removes them. Option C is incorrect because converting the demo leads into opportunities would not offset bias, it would compound it by feeding the model fabricated won deals that do not reflect real buying behavior, further distorting what the model learns.