Most machine learning projects fail not from poor models but from wrong problem framing, according to 2024 RAND Corporation research. A new framework argues teams waste months optimizing metrics that don't map to business value.

Andrew Ng's data-centric AI approach emphasizes that better algorithms mean teams should spend more time on problem definition before training code. The article presents a five-step protocol to catch framing failures early through stakeholder conversations rather than GPU cycles.

Real-world examples demonstrate the cost of skipping this step. Zillow lost over $500 million when its home-buying algorithm assumed stable market relationships that didn't hold. A medical imaging model learned to detect rulers instead of cancer because training data showed rulers correlated with malignancy. These weren't modeling failures but upstream framing mistakes that no hyperparameter tuning could fix.