Managing scope, effort estimates and delivery schedules for a machine learning product
Company: Google
Role: Product Manager
Category: Product / Decision Making
Difficulty: medium
Interview Round: Other
Think of an AI product or AI-powered feature you worked on. Walk the interviewer through how you defined its scope, how you estimated the effort and cost involved, and how you built and managed the timeline from the first prototype to launch. Be specific about the decisions you made at each stage and how you kept the plan on track as the work progressed.
```hint Plan around what you cannot predict
Some of the work on an AI product is ordinary engineering that estimates normally, while some depends on how fast model quality improves. Show how your plan treated those two kinds of work differently.
```
### Clarifying Questions
- Should the example be a feature built on an existing foundation model, a model the team trained, or is either acceptable?
- Does estimation here mean engineering effort and running cost, or also sizing the opportunity that justified the scope?
- Is the interviewer more interested in the initial plan or in how the timeline was managed once work began?
### What a Strong Answer Covers
- Scoping to a minimum useful version with an explicit quality bar, a fallback for model errors, and named non-goals
- Estimates that separate predictable engineering work from uncertain model and data work, using ranges and time-boxed experiments
- A timeline with milestones and go or no-go gates (offline evaluation, internal testing, limited rollout) and a clear rule for whether the date or the scope flexes
- Cost and dependency planning: data access and labeling, inference cost, and privacy, safety and legal review
- Honest reflection on what was misestimated and what the candidate now does differently
### Follow-up Questions
- Halfway through, the model is stuck below the quality bar and the launch date is fixed. What do you do?
- How do you give leadership a date when the critical path is an experiment whose outcome you cannot predict?
- How would your plan differ for a feature built on a hosted model's API versus one that needs a custom-trained model?
Overview: A product manager execution question asking you to walk through how you scoped, estimated and scheduled an AI product you worked on. It tests planning around uncertain model quality, setting a quality bar and milestone gates, estimating effort and running cost, and communicating timeline risk to stakeholders.