A consultant is preparing to enable AI-driven lead scoring for a sales org whose Dynamics 365 instance has accumulated years of manually entered leads, many of which are duplicates of the same contact from repeated marketing campaigns. Before turning on scoring, what should the consultant prioritize?

The correct answer is the deduplication and data-cleansing pass on the lead records so the model trains on distinct, accurate records. AI-driven lead scoring learns patterns from the historical data it is trained on, so duplicate contact records skew those patterns by overrepresenting certain attributes and outcomes; cleansing and deduplicating the data before enabling scoring ensures the model learns from distinct, accurate records and produces reliable scores. Enabling scoring immediately and relying on the model to learn to ignore duplicate records over time is incorrect because scoring models do not inherently detect or discount duplicate records as training noise; the duplicates would continue to distort the training data indefinitely rather than being learned away. Excluding marketing-sourced leads from the scoring model entirely is incorrect because marketing-sourced leads are a normal and often valuable source of pipeline, and excluding them wholesale discards legitimate leads instead of addressing the actual problem, which is duplication rather than the source itself. Increasing the minimum score threshold required before a lead is routed to a rep is incorrect because raising the routing threshold only filters which scored leads reach a rep; it does nothing to fix the underlying data quality issue feeding inaccurate scores in the first place. Addressing data quality before enabling the model is the standard prerequisite for reliable AI-driven scoring.