Return fraud: the returns nobody reads
the returns you argue about are not the ones costing you money. the expensive ones go through quietly, on the same morning as forty others, because nobody had the time to read them.
the shape of it
a survey of two hundred retailers by radial put fraud and abuse of returns at the top of the list of concerns for 2026 — ahead of shipping cost, ahead of demand. the national retail federation puts what it costs the us market at around a hundred billion dollars a year.
what changed recently is the tooling on the other side. riskified reports that close to half of shoppers now use ai somewhere in a return claim: a generated photo of damage that never happened, a complaint written by a language model in exactly the register that gets refunds approved. the claim arrives fluent, specific, and plausible, and it is indistinguishable from a real one to somebody skimming.
and the volume is not the issue. the national retail federation counts around nine per cent of returns as confirmed fraud — and that is only what somebody proved, not the grey area of claims nobody had time to look at. damage, disputes, outside the window, wrong item back in the box: those are the ones that need a decision rather than a label, and they are where the money goes.
what the portals cover, and what they do not
return portals are genuinely good products. self-service, printed labels, exchange offers, tracking, a customer who never has to email you. for the majority of returns, which are honest and ordinary, they are the right purchase and cheaper than anything built.
their design goal, though, is to avoid needing a human decision. that is why they work. and it is why the ones that need one fall straight through: the portal issues the label because issuing the label is what it does.
so the leak is not a gap in the product. it is the product working as designed, on the cases it was never meant to handle.
the part that actually needs building
what closes it is not another portal and not a fraud score. it is three things in a row.
read the claim, not the reason code. the reason code is a dropdown the customer picked. the wording is where the information is: what they say happened, when, in what order, and whether it matches what the order and the tracking say. a model reads that in a second and does not get bored on the fortieth one.
apply your rules, identically, every time. not the model’s opinion of fairness — yours, written down. which claims go through without question, which need a photo, which need the item back before anything is refunded, what happens on a third claim from the same address in two months. once that is written, the standard is the same at nine in the morning and at six in the evening, which it currently is not.
hand back one batch, not a hundred interruptions. everything ordinary settles itself. everything that needs a person arrives once, in one place, with the reason it was flagged and what the rules suggest. signing off a morning batch takes minutes. being interrupted forty times takes the day.
that is what i run in my own business, on real returns, every day. the ai reads the wording; my rules make the call. i can explain any decision to a customer, which matters more than the recovered margin.
the honest limits
this is worth saying plainly, because the opposite gets sold a lot.
it does not catch a first-time abuser who writes a careful, ordinary-looking claim. nothing does. what it catches is patterns and inconsistency — the same address, the same story, a timeline that does not fit the tracking, a claim that contradicts the order.
it will also be wrong sometimes, in both directions, which is exactly why the batch goes to a person. a returns process that never refuses anything is not watching, and one that refuses on its own authority is a customer-service problem waiting to happen.
and it is not worth building at small volume. under about a hundred returns a month, read them yourself. the case for building starts where the volume guarantees that nobody reads them properly.
when did your returns process last refuse a claim?
when did your returns process last refuse a claim and say why?
if the answer is never, that is not evidence your customers are honest. it means the ones that need a decision are being decided by default, in the customer’s favour, every single day, and it does not show up anywhere you look.
faq
what counts as return fraud?
anything from wardrobing and worn-item returns to claiming damage that was not there, keeping the item and reporting it missing, or returning a different product in the box. most of it is not organised crime, it is ordinary customers who learned which claims go through unchallenged.
do return portals stop it?
no, and they do not claim to. a portal is very good at the majority of returns that are honest and ordinary: it issues the label, tracks the parcel, offers the exchange. fraud lives in the minority that needs a decision, and a portal’s job is to avoid needing one.
can ai read a return claim and decide?
it can read the claim and apply your rules to it consistently, which is the part a tired person does badly at four in the afternoon. the decision itself should stay yours, written down as rules, because the standard has to be the same for everyone and you have to be able to explain it.
how do i know if anything is checking my returns?
ask when your process last refused a claim and named the reason. if it never has, either your customers are unusually honest or nothing is looking.
is it worth automating if my return volume is small?
probably not. under about a hundred returns a month a person reading them is cheaper and better. the arithmetic changes when the volume is high enough that nobody reads them properly and the ones nobody reads are the expensive ones.
if any of this sounds like your mornings, bring me a month of returns and i will tell you where it is leaking and whether it is worth building anything at all. the first conversation is an hour and it is free
Created with AI assistance.