# Return Fraud Risk Scoring

Companion to **AI Agents for Return Fraud and Abuse Detection**.

**AI identifies risk and supports investigation. It does not prove fraud and must not automatically accuse customers.**

Scores are **internal**. Customer-facing language is policy outcome (approve / reject / review), never “you are fraudulent.”

## Signals (examples — calibrate on **your** base rates)

| Signal | Direction | Not proof because |
| ------ | --------- | ----------------- |
| High RMA rate vs category baseline | Higher risk | New SKU quality issue |
| Delivered-not-received cluster on one address | Higher risk | Carrier failure in a zip |
| Serial returns of high-resale SKUs | Higher risk | Legitimate collectors |
| New account + high-value + immediate RMA | Higher risk | Gift / wrong size |
| Mismatched identity signals | Higher risk | Shared household |

## Score → action

| Band | Agent does | Agent does not |
| ---- | ---------- | -------------- |
| Low | Continue normal refund matrix | Extra interrogation copy |
| Medium | Attach score + evidence to HITL | Auto-reject |
| High | Flag for review; freeze **auto**-refund | Publish accusation; ban without human |

Never Execute `createReturn` solely because the score is low. Eligibility still follows [`refund-investigation-decision-matrix.md`](refund-investigation-decision-matrix.md).

## Related

- Refunds: [`refund-investigation-decision-matrix.md`](refund-investigation-decision-matrix.md)
- Security: [`ecommerce-agent-security-checklist.md`](ecommerce-agent-security-checklist.md)