AI for Returns Processing: Predicting Which Returns Are Fraudulent

Return fraud cost US retailers an estimated $24.5 billion in 2025, according to the National Retail Federation. For every $100 in returned merchandise, about $15.14 was fraudulent. The problem is growing faster than overall ecommerce because fraudsters have gotten sophisticated, and most retailers still rely on blanket return policies that treat every customer identically.

A sporting goods retailer running about $80 million in annual online revenue shared that they were losing roughly $1.9 million per year to return fraud before implementing an AI-based detection system. After 14 months with the system running, confirmed fraud losses dropped to about $620,000. The model did not catch everything, but it flagged enough to cut losses by two-thirds.

The Common Fraud Patterns

Wardrobing is the most prevalent form, where customers buy clothing or accessories, wear them once (often for a social media photo or event), and return them. Signals include returns of high-value items within 2-3 days, return of items with tags reattached (sometimes detectable through tag condition data if your warehouse tracks it), and repeat patterns from the same customer.

Bracketing fraud is when customers intentionally order multiple sizes or colors planning to return most of them. This is not always fraudulent (some retailers encourage it), but it becomes a problem when combined with wardrobing or when the returned items show signs of use. The distinguishing signal is the return rate: legitimate bracketers return 60-70% of multi-item orders, while fraudulent bracketers often return 90%+ with items that cannot be resold as new.

Receipt fraud and price switching involve returning items bought at a discount for full-price credit, or returning a cheaper item in the box of an expensive one. These are harder to catch without physical inspection, but AI can flag suspicious patterns like customers who consistently return items that were purchased during promotions.

Organized retail crime accounts for a smaller percentage by volume but a larger share of dollar losses. These are professional operations that buy goods and return them for store credit or refunds. The behavioral signals are different from individual fraud: multiple returns across different accounts using similar addresses or devices, returns of high-demand items in bulk, and new accounts with no purchase history making large returns.

What the AI Models Actually Look At

The fraud detection models that work best in ecommerce returns are not looking at any single factor. They score returns based on a weighted combination of signals. Customer return history is the strongest predictor. A customer with a 65% return rate across 30+ orders has a fundamentally different risk profile than someone returning their second item ever. The model tracks not just return rate, but return velocity (how quickly items come back), return value distribution (always returning the most expensive item in a multi-item order), and seasonal patterns.

Item-level data matters too. Certain product categories have naturally higher return rates (apparel averages 25-30%, while electronics sit around 8-10%). The model needs category-specific baselines to avoid flagging legitimate returns in high-return categories. Within categories, specific items with unusually high return rates from specific customer segments can indicate a targeted wardrobing pattern.

Device and session data adds another layer. If the same device fingerprint is associated with multiple accounts that all show high-return behavior, that is a strong signal of organized fraud. IP address patterns, browser fingerprints, and account creation timelines all feed into the risk score.

Timing patterns are surprisingly informative. Fraudulent returns cluster around specific windows: right before the return policy deadline, immediately after major events or holidays (suggesting wardrobing for the event), and suspiciously quickly after delivery (ordered, received, returned within 48 hours with no apparent reason).

Building vs. Buying a Returns Fraud System

Off-the-shelf solutions from companies like Riskified, Signifyd, and Forter offer return fraud scoring as part of their broader fraud prevention platforms. These work well for mid-market retailers because they have been trained on data from hundreds of merchants, giving their models a broader view of fraud patterns than any single retailer could develop alone. Typical costs run 0.5-1.5% of the return value processed through the system.

Building in-house makes sense for large retailers with unique fraud patterns or those processing enough returns (50,000+ per month) to train their own models effectively. The build typically involves a feature engineering pipeline that extracts the signals described above, a gradient boosting or neural network model trained on historical returns labeled as fraudulent or legitimate, and a scoring API that evaluates each return request in real time.

The labeling problem is the hardest part of building in-house. Most retailers do not have clean historical data on which returns were actually fraudulent. You often need to start with a rule-based system (flagging returns that meet certain criteria for manual review), use the manual review outcomes to build a labeled dataset, and then train a model on that labeled data. This bootstrap process takes 6-12 months to generate enough labeled examples for a reliable model.

What To Do With the Fraud Scores

The model outputs a risk score, typically 0-100. The question is what to do with it. Most retailers implement a three-tier approach. Low-risk returns (score below 30) get processed automatically with no friction. Medium-risk returns (30-70) go through the normal process but get flagged for post-processing review. High-risk returns (above 70) trigger additional verification steps before the refund is issued.

Those additional steps might include requiring photos of the item before a return label is generated, routing the return to a dedicated inspection team at the warehouse, delaying the refund until the item is received and inspected, or in extreme cases, politely informing the customer that the return does not qualify under the current policy.

The key is calibrating the thresholds. Set the high-risk threshold too low and you add friction for legitimate customers, damaging satisfaction and lifetime value. Set it too high and you miss fraud. Most retailers start conservative (threshold at 80-85) and gradually lower it as they build confidence in the model accuracy.

For ecommerce and retail businesses dealing with growing return volumes, even a basic fraud scoring system that catches the most obvious patterns can recover 1-3% of return-related losses. The more sophisticated your data collection and model training become, the higher that recovery rate climbs. The retailers seeing the best results are those that treat return fraud detection not as a one-time project but as an ongoing optimization, regularly retraining their models as fraud patterns shift and evolve over each season.