What changed
Online fashion returns are a familiar consumer habit and an expensive retailer headache. Many shoppers order multiple sizes or colors expecting to send some back, and many merchants tolerate that pattern because stricter return fees can backfire. This paper tested a softer intervention: remind shoppers that returns carry environmental costs, then check whether machine learning can decide who should see that reminder.
Researchers partnered with a leading European fashion retailer and ran a large randomized field experiment covering 117,304 shopping sessions. The green nudge appeared in two places: inside the cart and again near the end of the shopping journey. The baseline finding was practical and modest rather than magical. Showing the nudge to everyone reduced return shipments by 2.6%, with statistical significance at p < 0.03, and did so without adversely affecting sales.
The AI part came in the second step. The authors used digital-footprint and cart data to train a causal machine-learning model that estimated which shoppers were most likely to reduce returns when nudged. Their off-policy evaluation found strong heterogeneity: about 60% of customers appeared responsive, while the rest did not. Under that targeted policy, smart green nudging was estimated to lower return shipments by 5.19% versus no nudging and by 3.15% beyond the naive strategy of nudging everyone.
The mechanism the paper points to is narrower than broad consumer virtue. The authors say the cart prompt likely reduced bracketing: buying several variants with the intention of returning some. That is useful because it ties the result to a specific everyday shopping behavior instead of claiming a general transformation in ecommerce.
What this could change for you
For shoppers, the most believable near-term change is not an AI agent that shops on your behalf. It is more subtle checkout design that nudges you to pause before ordering several near-duplicate items with the expectation that returns are free and painless. If that works, you may see fewer manipulative pressure tactics and more low-friction reminders aimed at reducing buyer regret and retail waste.
For retailers, the practical lesson is sharper. The paper suggests there is a middle ground between doing nothing and charging punitive return fees. A lightweight reminder reduced returns without a sales penalty in this setting, and a targeted version could potentially do better by focusing on customers who are actually responsive instead of annoying everyone.
The broader AI lesson is that personalization can matter even when the front-end experience looks simple. The machine-learning model did not generate prose, negotiate prices, or replace human support. It sorted a narrow decision problem: which shoppers are worth nudging at all.
What it does not prove
This was one European fashion retailer in one short-run experimental setting. The paper does not prove the same effects would transfer unchanged to electronics, groceries, furniture, luxury goods, marketplaces, or retailers with different return policies and customer expectations.
The stronger AI result comes from off-policy evaluation, not a live randomized deployment of the targeted model itself. The cleanest proven field claim is the 2.6% reduction from showing the green nudge broadly. The larger 5.19% reduction is an evidence-backed estimate of what targeted deployment could achieve, not a realized production rollout in this paper.
The study also does not settle the long-term customer side of the story. It did not directly measure customer lifetime value, lasting satisfaction, or whether repeated exposure to the nudge would wear out, backfire, or change brand perception over time. The authors also say they cannot isolate the cart prompt's mechanism with absolute certainty.
The bottom line
AI did not solve ecommerce. It solved a narrower and more believable problem: too many avoidable online fashion returns. In this large randomized retailer experiment, a simple returns reminder cut return shipments without hurting sales, and a causal machine-learning model suggested the same intervention could work better when aimed only at the shoppers likely to respond. The honest takeaway is that some of the best shopping-AI gains may come from smarter restraint, not from pushing people to buy more.
Primary research
Smart Green Nudging: Reducing Product Returns Through Digital Footprints and Causal Machine Learning
Marketing Science · 2024 · DOI 10.1287/mksc.2022.0393
