A retailer says returns fraud is costing them millions. They ask what AI can do. Where do you start?
Think first
Quantify, define fraud precisely, and think about the cost of false positives on honest customers.
Strong answer
Ask how they know the size of the loss and how fraud is defined today (wardrobing, receipt fraud, return of stolen goods, employee collusion). Each has different signals and different interventions. Find what labelled data exists: confirmed fraud cases, chargebacks, store investigations. Then frame the decision the AI supports: flag a return for manager review, require ID, or refuse. The cost of a false positive is an honest customer treated as a thief, so precision at the review threshold matters more than overall accuracy. Start with a scoring model plus explanations on historical returns, set the threshold with loss-prevention and customer-experience leads together, and pilot in a few stores with a control group. Generative AI is a supporting tool here, for summarising case history for investigators, not the core model.
Weak-answer indicators
- Reaches for an LLM for a tabular classification problem
- Ignores false positives and customer experience
- No control group or way to prove savings
Likely follow-up questions
- How do you prove the savings were caused by the model?
- Fraudsters adapt. How do you keep the model current?