NRF and Happy Returns forecast about $850 billion in returned merchandise for the year. Retailers put the overall return rate at 15.8% of sales and the online rate at 19.3%. They estimate that 9% of all returns are fraudulent. (NRF)

Those figures come from a survey of consumers who returned an online purchase and of ecommerce professionals at large US merchants. They describe a cost most businesses can see on the income statement and rarely see in detail: the margin that walks out the door in a box.

What the fraud looks like

Among retailers that track fraud, overstated quantities show up most often (71%), followed by empty boxes (65%) and decoy returns such as counterfeit items (64%). Eighty-five percent say they already use AI to detect or prevent it. Retailers also expect around 17% of holiday sales to come back.

Fraud isn't the only leak. Honest returns carry costs too: shipping both ways, handling, inspection, repackaging, and the markdown on goods that can't be resold at full price. A return rate of one in six means one in six sales passes through that process.

The blunt response

Many retailers have started charging for returns, citing rising operations costs and carrier costs, 40% each. Others tighten windows or restrict who can return what. These measures work on the cost line. They also land on the shopper who sent back one wrong size, and she pays the same as the person boxing up bricks. A rule that treats everyone alike taxes the customers you most want to keep.

It can also change behavior in ways that hurt. Shoppers who expect to pay for returns buy less of the things they are unsure about, which are often the products with the best margin.

Score the return, not the customer

A return score looks at the shape of a return: the item, the account's history, how the returns cluster, what came back compared with what was sent. It produces a number. Your team picks a threshold, and everything above it gets a second look from a person. Nothing is refused automatically.

The signals are ordinary and most of them are already in your systems.

• How the account's return rate compares with others who bought the same products.
• Whether returns arrive in a pattern: the same item type, the same time after delivery, the same address.
• Whether the weight or contents of the returned parcel match the order.
• How long the account has existed and what it bought before.
• Whether the refund request or the claim of a missing item looks like earlier claims.

A language model can help by reading the free-text reasons shoppers give and the notes your team writes, and by drafting the case summary. It shouldn't decide.

A worked example

Take 10,000 returns in a month. Suppose 9% are fraudulent, as the retailers surveyed estimate for the industry. That is 900 returns worth chasing and 9,100 from honest customers.

A threshold set low flags many accounts. Say it catches 700 of the 900 fraudulent returns and also flags 1,400 honest ones. Your team reviews 2,100 cases to stop 700, and 1,400 honest customers get a slower, more suspicious experience.

Raise the threshold and the picture changes. Say it catches 500 of the 900 and flags 200 honest ones. Your team reviews 700 cases, and 200 honest customers are inconvenienced. You miss more fraud, and you spend far less review time and annoy far fewer good customers.

Neither setting is right in the abstract. The right one depends on what a fraudulent return costs you, what a review costs, and what it costs to irritate a good customer. That is a business decision, and your team should be able to move the dial and see what happens. The return fraud field guide has a slider that lets you feel the trade.

Read the pattern both ways

When a regular's returns change, the cause can be abuse, or the account can be starting to leave. The same signals feed both questions, so the second look is worth having a person do, with the account history in front of them. A sudden rise in returns from a loyal customer might mean a sizing problem with a new line, a product defect, or a shopper who has stopped trusting you.

Fix the cause when you can

The cheapest return is the one that never happens. Look at the products with the highest return rates and ask why. Size guides that say what the garment is actually like, photos that show scale, attributes that answer the question a shopper had, and honest descriptions all reduce returns that were never fraud in the first place. Better product data pays here as well.

Where to start

Pull three months of returns and join them to orders and accounts. Look at the cases your team already thinks are abuse and ask what they had in common. Build a simple score from those signals, set a threshold that gives your reviewers a workload they can manage, and keep a control group so you can tell whether the score is earning its keep. Our piece on the holdout rule covers how.