A v1.0 framework for reasoning about — and instrumenting — AI employee failure in production. Instead of a headline percentage, it offers a five-part failure taxonomy, a 2026 baseline by task complexity, and an architect's guide to reading failure data by shape, not rate.