Data minimization is difficult in legacy systems because teams cannot identify which fields are actually used. A code scanning approach can identify unused data fields by tracing their paths through the codebase, creating an evidence-based shortlist for removal decisions rather than relying on room consensus.
PostHog built a semantic layer—a governed catalog of definitions sitting atop existing data—to ensure AI agents and humans consistently interpret metrics like MRR the same way. The semantic layer doesn't copy or move data; it describes what metrics mean, which tables to trust, and how data sources connect, solving the problem of tribal knowledge that causes different agents to produce conflicting answers from the same question.