A meta-analysis of 420 studies on human-AI collaboration finds that humans working with AI outperform humans working alone, but fall short of the best solo human performer. That gap is the central design fact for anyone building products where AI handles part of the work. The finding is not a footnote. It sets the ceiling for what handing work to an agent can achieve, and it tells you where routing decisions matter most. The analysis is self-updating, meaning it adds new studies as they appear and revises its conclusions. For a design team, that is a live calibration instrument. You can track whether the gap between human-AI pairs and the best solo performer is narrowing, holding, or widening as models change. That kind of longitudinal signal is rare in a field that moves this fast. The practical implication points at two patterns: how you hand work to an agent, and how you route that work back when the agent's ceiling is not high enough. A product team that treats handing work over as a binary switch, on or off, will leave performance on the table. The evidence here suggests you need to know, before handing work over, whether the task is one where human-AI pairs close the gap or one where a skilled human alone still wins.