Public-health leadership asks where data preparation sits relative to training and deployment in the AI lifecycle. What is the correct placement?

In the AI lifecycle, data preparation sits before training and deployment: raw municipal public-health data is cleaned, labeled or featurized, quality-checked, and made usable so training can proceed and a later deployment has a trustworthy basis. Leadership should place preparation upstream of model fitting and production release. Data preparation does not occur before data collection and training; collection must supply the raw municipal data first, and placing preparation before collection would leave nothing to clean, label, or quality-check before training, inviting silent quality failures. It also does not occur before data labeling and training—labeling is a core task inside preparation alongside cleaning and featurizing, so preparation cannot sit upstream of its own component before training. It does not occur before data cleaning and training—cleaning raw data is part of preparation itself, so preparation cannot sit upstream of its own cleaning work before training. Exam caveat: preparation may iterate after monitoring finds issues (new labels, cleaned features, corrected schemas), but those loops still treat preparation as work that includes labeling and cleaning to feed retraining, not as a stage that occurs before data labeling and training or before data cleaning and training. Operational check: ensure public-health datasets clear preparation gates (schema, quality, privacy, lineage) before training jobs consume GPU time, and keep preparation artifacts versioned for audit before any production deployment.