Review of 19 studies finds AI agents lift retail and supply chain results on average, but effects vary widely

The average hides enormous differences between studies. Heterogeneity was near-total (I² = 99.5%), and the 95% prediction interval of -4.20 to 7.36 means that in a new setting the true effect could range from substantial harm to exceptional benefit. The authors say the pooled figure should not be read as a general implementation expectation.

Evidence for differences between operational domains was weak. Retail, supply chain, e-commerce and inventory differed under a common-effect model, but not significantly under the random-effects model (p = 0.081). A limit meta-analysis that adjusts for selective reporting lowered the estimate to g = 1.26 (p = 0.066).

The authors say the empirical evidence on agentic AI in retail and supply chains had not been systematically quantified before. Their meta-analysis covers 19 studies and 5,292 participants in total, using a random-effects model with restricted maximum likelihood estimation and Hartung-Knapp confidence intervals.

The pooled gain of g = 1.58 is statistically significant, but the spread between studies is extreme. With I² of 99.5% and a prediction interval of -4.20 to 7.36, a new deployment could see either substantial harm or exceptional benefit.

Domain comparisons proved inconclusive. A common-effect model showed differences between retail, supply chain, e-commerce and inventory (Q = 935.45, p < 0.001), but the random-effects model, which the authors consider more appropriate, did not (p = 0.081).

After adjusting for selective reporting, the estimate fell to g = 1.26 (p = 0.066). The authors call for real-world field experiments, architecture-specific evaluations, publication of null results and validated quality frameworks.