A retailer clusters customers with k-means on Recency (days, 1–700), Frequency (orders, 1–40), and Monetary (dollars, 5–12,000) using the raw columns.
What will go wrong, and what is the fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
k-means minimizes Euclidean distance, and a $10,000 difference in spend is numerically ten thousand times a 1-day difference in recency. Unscaled, the algorithm effectively clusters on one column. Standardizing (or taking logs of the skewed dollar and count columns, then standardizing) puts the three RFM dimensions on equal footing so the segments reflect all three behaviors.
k-means works in any number of dimensions. PCA is for visualization or noise reduction, not a requirement.
Distance-based methods — k-means, kNN, SVM — are among the most scale-sensitive algorithms. Trees are the ones that ignore scale.
k-means has no notion of business relevance. Numeric range alone decides which feature drives the distances.