A Dynamics 365 Sales admin is configuring the Sales Copilot forecasting assistant. Sellers report that the AI-generated deal risk scores feel disconnected from reality because the model was trained mostly on historical opportunities from a product line the company discontinued last year. Which action should the consultant recommend to improve the relevance of the AI-driven insights going forward?
Select an answer to reveal the explanation.
Short Explanation
Think of a forecasting model like an employee who learned the job years ago and never got retrained after the company changed direction. If most of what it learned came from a product line that no longer exists, its instincts about what a risky deal looks like are going to be off, no matter how confident it sounds. The fix here is not to fire the assistant or bury it in extra paperwork, and it is not to make it flag more deals just in case. The real fix is to give it fresh, relevant experience: recent opportunity data that actually reflects what the company sells and how deals move today. Once the model is learning from the current reality instead of an outdated one, its risk scoring starts to line up with what sellers are actually seeing in the field. Retraining on current data is the direct fix; everything else either avoids the problem or papers over it without changing what the model actually knows.
Full Explanation
The correct answer is B. When an AI scoring model was trained on data that no longer reflects the business, such as a discontinued product line, the fix is to refresh the training data with recent, representative opportunities so the model's patterns match current reality. Option A is incorrect because disabling the assistant discards a capability the organization wants to keep rather than fixing the root cause, and abandons the investment without addressing the data quality issue. Option C is incorrect because adding more required fields increases data entry burden and may improve some inputs, but it does not correct a model trained on stale, irrelevant history; the problem is training data recency, not field count. Option D is incorrect because adjusting the confidence threshold changes how sensitive the output is, not whether the model's learned patterns are accurate, so it would surface more false positives rather than better ones. Retraining with current data directly addresses why the risk scores feel disconnected: the model needs examples that reflect today's business, not a product line the company no longer sells.