Rank fields by population rate and last-modified date, then check each candidate against every consumer — reports, flows, integrations, list views — before deprecating. Hide first, delete later: a field with no data can still be the one an integration writes to nightly.
An AI assistant is genuinely good at the inventory-and-cross-reference pass here, and poor at the judgement call about what a field means to a team. Use it for the audit, keep a human on the deprecation decision, and stage the removal so anything you got wrong surfaces while it is still reversible.
Every claim above comes from work LeanScale published. These are the sources.