Constrained override policy boosts AI-human teamwork: 1.28% inventory cut, no sales loss

A new arXiv paper from researchers at a major Chinese smart vending retailer (59,000+ machines, 4,000 SKUs) tackles a classic AI supervision dilemma: how much override authority to give human workers. They propose a simple but effective solution—a constrained override policy that limits the number of overrides per decision episode (here, 2 downward overrides per machine). In a randomized field experiment with 553 workers, they compared no overrides, free overrides, and the constrained policy.

Results were striking. Free overrides reduced inventory by 1.95% but also cut sales by 1.19%, suggesting human bias and noise outweighed private information. The constrained policy, however, reduced inventory by 1.28% with no negative sales impact, as workers selectively overrode only the most impactful SKUs. Gains were largest for experienced workers, high-incentive SKUs, and growth-stage products. A follow-up simulation showed a personalized version of the policy could further increase sales probability by 9.1%. The approach is scalable and requires no algorithm redesign, extra training, or information customization—just smarter override limits.