A manufacturing distributor recently changed its Dynamics 365 Sales data entry process, adding new required fields for deal-stage justification and removing an old free-text notes field that sellers used heavily. Three months later, a consultant reviewing the AI-driven opportunity scoring model finds its accuracy has dropped compared to launch. Which action should the consultant recommend first?
Select an answer to reveal the explanation.
Short Explanation
Think of a scoring model like someone who learned to judge deals by reading a particular kind of note sellers used to write. Take that note away and hand them a different form to read instead, and their judgment gets shakier, not because they got worse at judging, but because the information they built their instincts on changed shape. That is exactly what happens here: the process change swapped out a signal-rich field for new structured ones, so the model's old patterns stopped lining up with reality. The fix is not to undo the improvement, since the new fields were added for good reason, and it is not to just feed the model more history in the old shape, since that keeps mixing signals that no longer behave the same way. It is also not to give up on the feature altogether over something this fixable. The right move is to let the model relearn from data that reflects how deals are actually recorded now. Once it retrains on the current structure, its scores should realign with outcomes again.
Full Explanation
The correct answer is B. The data entry process change altered which fields carry signal for the model: a heavily used free-text field disappeared and new structured fields appeared, so the patterns the model originally learned no longer match how deals are recorded. Retraining on recent data that reflects the new structure lets the model relearn which fields now predict outcomes, restoring accuracy. Option A is incorrect because the new fields were a legitimate process improvement; reversing them disrupts the business to chase a symptom instead of fixing the model's alignment with current data. Option C is incorrect because simply widening the historical window without updating which fields are used still trains on a mix of old and new field structures, which does not resolve the mismatch causing the drift. Option D is incorrect because disabling the feature discards a working capability over a fixable, well-understood cause and leaves sellers without prioritization support for months. The right response treats this as an expected consequence of changing input data, not a reason to abandon or delay the model.