Case evidence
HarbourHR wants an AI assistant to answer staff benefit questions from an approved knowledge base. It can draft answers but the pilot has no authority to change employee records.
AI-01: 40 synthetic test questions, 6 wrong answers, 2 answers exposing synthetic sensitive records.
AI-02: the knowledge base includes outdated policy pages.
AI-03: business acceptance thresholds proposed for the exercise are at most 2 wrong answers in 40 and zero sensitive-record exposures..
Every record here is synthetic. Day 0 is the exercise baseline. Translate relative dates consistently if you choose calendar dates. These observations support only the stated fictional scope.