Process for Building a Generative AI Product from Problem Selection to Launch
Company: Amazon
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
How would you go about building a product that uses generative AI? Walk through your process from idea to launch; you may pick any product to illustrate it.
```hint Answer the "how"
The interviewer wants a process more than a product idea. Think about which steps differ from building ordinary software when the core component produces probabilistic output that is sometimes wrong.
```
### Clarifying Questions
- Are we a startup without models of our own, or a team that already has proprietary data and machine learning infrastructure?
- Should the process be illustrated with one concrete product, or kept general with examples only where they help?
- Is the product for consumers or for businesses, given that the tolerance for wrong or unsafe output differs?
- Are there limits on inference cost or response time that would rule out the largest hosted models?
### What a Strong Answer Covers
- Starting from a user problem where generated output adds value, and checking whether generative AI beats rules or conventional models for it.
- Constraints specific to generative AI: tolerance for wrong or invented output, privacy and data rights, safety guardrails, cost per request and latency.
- Choosing a technical approach (prompting a hosted model, retrieval-augmented generation, fine-tuning, agents) and the trade-offs behind the choice.
- An evaluation plan and release bar set before launch, pairing model-quality metrics with user and business metrics.
- A staged rollout with monitoring, feedback loops and a plan for model updates and cost changes after launch.
### Follow-up Questions
- Offline evaluation looks strong, but pilot users rewrite most outputs heavily. What do you do?
- Inference cost per active user turns out higher than revenue per user. Which levers do you pull?
- How do you decide when a task can be fully automated and when a person should review every output?
Overview: A product management prompt asking how you would build a product with generative AI, from choosing the problem to launch. It tests whether a candidate can lay out a repeatable process covering error tolerance, privacy and cost constraints, model approach, evaluation, staged rollout and monitoring.
