In PMI-aligned AI data management practices, handling missing data is approached from a risk, quality, and fitness-for-use perspective. Before model development, the project manager must ensure that the dataset is not only complete enough, but also representative and unbiased for the intended AI use case. When the portion of missing data is minimal and not systematically biased, a common, acceptable mitigation is to remove those records so that the remaining dataset maintains integrity and consistency while avoiding the introduction of artificial or misleading values.
Options B and C (duplicating data or blindly filling zeros) can create serious distortions in the underlying data distribution, leading to biased model behavior, degraded performance, and weaker generalization, which contradicts responsible AI practices highlighted in PMI-style guidance. Simply ignoring missing data (option A) without a structured strategy or analysis is also discouraged, as it hides potential data quality issues and can propagate errors downstream.
Therefore, in line with good AI data preparation practice, when missingness is genuinely limited and not concentrated in critical attributes, removing records with missing values if minimal (option D) is the most effective and responsible approach among the given choices.