Why AI making tasks faster can actually slow your whole team down

When AI tools handle customer service or coding tasks, they produce fast first drafts—boosting average completion time per task. But in workflows where tasks queue up, undetected AI errors bounce back as rework, creating congestion that slows the entire pipeline. The researchers prove this "variance wedge" using queueing theory: under pressure, reviewers lower scrutiny thresholds, and AI only helps if error-correction overhead stays below manual-work effort. Measuring AI by task speed alone misses system-level damage.