5 days to 35 minutes.
Fraud catch up 3.4×.

$4.2M a year leaking to fraudulent returns.

A D2C apparel brand was processing returns on a 5-day refund cycle and losing $4.2M annually to fraudulent or abusive returns. Their existing rule-based fraud system caught the obvious cases (duplicate orders, address mismatches) but missed the patterns: address-cycling, multi-account fraud, and worn-and-returned scams. Tightening the rules would burn legitimate customers and tank NPS. The CEO wanted both: faster legitimate refunds and higher fraud catch, without losing the customers who were not abusing the policy.

Two agents, one routing decision.

We built a return-classifier that scores every inbound RMA against fraud patterns, return-reason taxonomy, and customer history. High-confidence legitimate returns ship the refund instantly. High-confidence fraud routes to human review with the evidence pre-assembled. The middle gray zone gets a confidence-weighted decision: refund as store credit (recoverable) or request photos before approving cash refund.

The fraud patterns I'd never have written rules for.

• The agent caught a multi-account refund-cycling pattern across 47 accounts in month two. No human reviewer would have spotted it.
• Store-credit-instead-of-cash for gray-zone returns was the trust unlock. Customers who were not fraudsters used the credit. The fraudsters never came back.
• CSAT held flat-positive despite the crackdown because the legitimate returns shipped faster. NPS only moves when the median customer experience moves.