By Shravan Prasad

At the end of my last post, I promised to investigate returns. I teased that the “hassle-free returns” promise wasn’t as neutral as it sounded. Having spent the last few weeks reading through academic research, patent filings, consumer protection literature, and industry reports, I can confirm: it isn’t. Not even close.

Here is what AI-powered returns systems actually do — and what that means for you as a shopper.

The Promise vs. The Reality

Every major e-commerce platform advertises some version of easy, free, no-questions-asked returns. It is one of the most powerful conversion tools in online retail — removing purchase anxiety by assuring shoppers they can change their mind. The promise feels unconditional.

The reality is that returns processing has become one of the most sophisticated applications of machine learning in e-commerce. Behind every return request is an automated decision system evaluating whether to approve, delay, scrutinise, or quietly flag your account. The outcome depends not just on the item or the reason you give — it depends on who the algorithm thinks you are.

What Is a Return Risk Score?

This is the concept at the centre of modern returns management: every shopper has a risk profile, built and maintained by the platform, that influences how their return requests are handled.

The inputs that typically feed into this profile include your historical return rate, the monetary value of items you tend to return, how often your stated return reason matches platform-detected patterns, how long you tend to keep items before returning them, your overall purchase volume and account tenure, and signals that suggest whether returned items are genuinely defective or simply unwanted.

Shoppers with low return rates and long purchase histories are generally processed quickly with minimal friction. Shoppers with higher return rates — even if every individual return was entirely legitimate — can find themselves routed into slower, more scrutinised workflows. In some documented cases, accounts are flagged or restricted without any direct communication to the shopper explaining why.

This is not a conspiracy theory. It is documented in industry literature, discussed openly in retail operations circles, and referenced in consumer protection research across multiple markets.

The Economics Behind It

To understand why this exists, you need to understand the scale of the returns problem in e-commerce. Industry estimates consistently place global e-commerce return rates between 20% and 30%, with categories like apparel and footwear reaching 40% or higher. The cost of processing a return — logistics, inspection, restocking or disposal, customer service — frequently exceeds the margin on the original sale.

Returns fraud and abuse add further pressure. Wardrobing — buying an item with the intention of using it once and returning it — is estimated to cost the retail industry billions annually. Bracket buying, where shoppers purchase multiple sizes or colours intending to keep only one, has become normalised behaviour encouraged by free returns policies.

From a pure business logic standpoint, using AI to differentiate between high-value, low-risk customers and high-cost, high-risk ones is rational. The problem is that this logic operates invisibly, affects real people in ways they cannot anticipate or contest, and uses proxies that can inadvertently penalise entirely legitimate shopping behaviour.

How the Decisions Actually Get Made

Returns AI systems typically operate across several decision points.

Instant refund or wait. Some shoppers receive refunds before they have even shipped the item back. Others must wait for the item to be received and inspected before any refund is issued. This differential treatment is driven by risk scoring — trusted accounts get speed, flagged accounts get process.

Return label or not. Platforms are increasingly experimenting with “keep it” returns — telling certain shoppers to keep the item and issuing a refund anyway, because the cost of processing the physical return exceeds its value. Predictably, this courtesy is extended to accounts the system has already assessed as trustworthy. It is invisible to everyone else.

Approval or rejection. In cases where the system flags high return risk, a return request may be declined outright, or routed to a human review queue that can take days. The rejection is rarely explained in terms that acknowledge the algorithmic assessment behind it.

Account restriction or suspension. At the extreme end, accounts that exceed certain return thresholds — even without any individual fraudulent transaction — can be restricted or closed. Several consumer protection bodies in the EU and UK have documented cases where shoppers lost account access and the purchase history, loyalty points, and stored payment methods that came with it, without meaningful recourse.

The Fairness Problem

Here is where I think the ethical issue becomes genuinely serious.

High return rates are not always a signal of abuse. People with certain disabilities may need to order multiple sizes or variants of an item because trying things in-store is not accessible to them. Shoppers in regions with poor size standardisation across brands may need to bracket-buy simply to find something that fits. First-time buyers in a new product category may over-order while they figure out what they need. None of this is fraud. All of it can generate a return profile that looks suspicious to an algorithm trained primarily on aggregate behavioural patterns.

The algorithm does not know why you return things. It knows that you return things, at what rate, and how that rate compares to the population it was trained on. If your legitimate behaviour falls outside the statistical norm, you can be disadvantaged by a system that was never designed to evaluate your specific circumstances.

This is not a hypothetical concern. Academic research into algorithmic decision-making in retail has documented consistent disparities in returns outcomes across different demographic groups — disparities that persist even after controlling for return rate differences, and that trace back to proxy variables that correlate with characteristics the system was never supposed to consider.

What You Can Do

I want to be careful here not to suggest that shoppers should game the system or that legitimate returns should be avoided. Returns are a consumer right. The advice below is simply about understanding how to interact with these systems in a way that reflects how they actually work.

Be specific about return reasons. Vague reasons like “changed my mind” are treated differently from specific ones like “size does not match the product description” or “item arrived damaged.” Specificity — especially when accurate — gives the system more signal to work with and typically routes your return more smoothly.

Return promptly. Items returned close to the deadline, or after extended hold periods, generate more scrutiny than items returned quickly. If you know you are returning something, initiating the process early works in your favour.

Know your consumer rights. In many jurisdictions, particularly across the EU under the Consumer Rights Directive, you have statutory return rights that platforms cannot override with their own policy language. A platform policy that is more restrictive than statutory law is not enforceable. If a return is rejected and you believe you have a legal right to return the item, escalate to your payment provider or the relevant consumer authority rather than accepting the platform’s decision as final.

Document everything. If you receive a defective or misdescribed item, photograph it before initiating a return. This documentation can be critical if your return is disputed or if you need to raise a chargeback through your bank or card provider.

The Bigger Pattern

If you have been reading The Neurals from the beginning, the shape of this series is probably becoming clear. We have looked at personalisation, chatbots, data profiling, dynamic pricing, search ranking, and now returns. In every case, the story is structurally similar: a system that presents itself as neutral and consistent is, underneath, making differentiated decisions about individual users based on behavioural profiles those users cannot see, did not consent to in any meaningful sense, and have very limited ability to contest.

The sophistication of these systems is genuinely impressive. The opacity is genuinely troubling. And the gap between how these systems are described to consumers and how they actually operate is, I think, one of the defining consumer issues of this decade.

Returns felt like the end of a transaction. They are actually the beginning of a new data point — one that quietly shapes every interaction you have with that platform going forward.

Next, I want to investigate something that sits upstream of all of this: how AI is being used to decide which customers are worth acquiring in the first place, and what that means for who gets shown ads, offered promotions, and targeted for growth. The discrimination starts before you even open the app.

If this resonated with you, share it with someone who shops online. That is everyone. See you next time.

— Shravan