Design an ML Trust System Against Fraudulent Third-Party Sellers on Amazon
Company: Amazon
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
You are a product manager at Amazon. Some of the third-party sellers who list products in Amazon's online store commit fraud. Explain how Amazon could use AI and machine learning both to detect this seller fraud and to prevent it.
```hint Map fraud to the seller's journey
Seller fraud takes different forms at different moments: signing up, creating listings, setting prices, fulfilling orders and collecting reviews. Decide which forms to tackle first, then choose signals and responses that fit each moment instead of one model for everything.
```
### Clarifying Questions
- Which kinds of seller fraud are in scope: counterfeit or misrepresented products, fake or incentivized reviews, deceptive prices and discounts, or orders that never ship?
- Should the approach start with one Amazon marketplace or cover every country's store from the outset?
- How costly is it to wrongly block an honest seller compared with letting a fraudulent listing through?
- Should I build on the seller-protection programs Amazon already runs, or design the approach from scratch?
### What a Strong Answer Covers
- A taxonomy of seller fraud and the harm each type does to shoppers, honest sellers, brand owners and Amazon, with a reasoned choice of where to focus first.
- Concrete signals and ML approaches matched to each fraud type, including where training labels come from and their weaknesses.
- Prevention and response: actions tiered by risk, human review, appeals, and stopping banned sellers from returning.
- Success metrics that balance fraud caught against harm to honest sellers, with guardrails.
- Risks such as adversaries adapting, delayed labels, fairness to small or new sellers, and regulation.
### Follow-up Questions
- Fraudsters learn your thresholds and change tactics within weeks. How do you keep the system effective?
- A small, honest seller is suspended by mistake during the holiday peak. What should their experience and the process look like?
- How would you estimate how much fraud you are not detecting at all?
Overview: A product sense question on how Amazon could apply AI and machine learning to find and stop fraud by third-party sellers on its marketplace. It tests mapping fraud types and their victims, choosing signals and models, designing risk-tiered enforcement with appeals, and balancing fraud caught against harm to honest sellers.
