After generating 4,200 AI test cases, reducing to 3,100 through expert review, the automated nightly test runs caught three regression bugs in the first month.