You review the AI to be safe. In one study, your review is what made…
Twelve accountants review an AI: the work takes fifteen times longer and two right answers turn wrong. A three column review grid, and the test that says whether your review is actually a check.
You have a human review everything the AI produces, and you consider that the careful thing to do. A study published in early October put twelve accountants up against an AI on rigged month-end closes, and the result is uncomfortable: the pair takes fifteen times longer than the AI alone, without being better. In two cases, the human's correction turned a right answer into a wrong one. What you walk away with: a three column review grid, and a five minute test to find out whether your review checks anything. What changes (PO, PM, designer, design system lead): "reviewed by a human" is no longer enough as an acceptance criterion. Why this week: Mercor published the study on 2 October 2026, with an AI that ticks every criterion where the accountants meet 37 %. · · · <!-- order:freeform…