AI's 'Variance Wedge' Can Slow Workflows Despite Faster Individual Tasks

A new study from Bartolucci and Vivo, posted on arXiv, challenges the prevailing wisdom that generative AI always boosts workplace productivity. Using a queueing model, the researchers demonstrate that while AI tools can cut average per-task completion times—a metric widely celebrated in industry reports—they can simultaneously degrade overall workflow performance. The culprit: a 'variance wedge' where a subset of AI outputs containing errors slips past human review and returns as rework, clogging the system under congestion. The paper draws on examples from customer service, writing, and software development to illustrate how faster drafts can paradoxically slow down the entire process.

The authors derive two key implications analytically. First, when workloads are high, reviewers rationally increase the risk threshold for checking AI outputs, reducing scrutiny precisely when it matters most. Second, AI assistance only stabilizes an overloaded workflow if two conditions hold: the fraction of tasks handled by AI exceeds a critical threshold, and the total human attention required for review plus expected rework is lower than that for manual completion. This requirement is far more stringent than simply delivering faster draft speeds. The findings urge organizations to evaluate AI deployment not just by speed but by its systemic effects on congestion, rework, and the robustness of human oversight under load.