A concrete starting sample
For an initial exercise, prepare 10 questions: 5 with known answers, 2 with missing information, 2 with conflicting versions and 1 about a source that user cannot access. This is a proposed test design, not certification or sufficient evidence for every risk.
What to record for each answer
Keep the question, expected answer, source, result, review time and available cost. Agree on acceptance thresholds before testing. An information leak should block expansion of the pilot.
Separate the operating commitments
Distinguish setup, connected sources, usage and support. Require a spending limit and a document owner. If AI will send, approve or change records, evaluate those actions separately with permissions and human confirmation.