AI Product Design Questions: Structure

Product sense questions are where AI PM interviews are won or lost. At Google, the product sense round carries more weight than any other round. At Meta, it is tied with execution. The AIDE framework gives you a repeatable structure for answering any AI product design question. AIDE stands for Audience, Intelligence, Design, and Evaluation.

The framework works because it mirrors how real AI products get built. You start by understanding who the user is and what problem they have (Audience). Then you determine what kind of AI/ML approach fits the problem (Intelligence). Then you design the user experience around the AI capability, including how to handle failure cases (Design). Finally, you define how you will measure success both for the model and for the product (Evaluation).

The key difference between AIDE and generic product frameworks is the Intelligence and Evaluation steps. These are where you demonstrate AI fluency. A traditional PM might jump from 'user has this problem' to 'here is the feature.' An AI PM asks: What data do we need? What model approach fits? What does 'good enough' look like for the model? How do we handle the cases where the model is wrong?

Spend 3 to 4 minutes on Audience. Identify 2 to 3 user segments, state which one you will focus on, and explain why. The 'why' is critical. Interviewers are evaluating your prioritization logic, not just your ability to list segments. A strong rationale references impact ('this segment is 60% of the user base'), strategic importance ('this segment has the highest churn rate'), or AI applicability ('this segment generates the most behavioral data, which enables better model performance').

For AI products specifically, consider how different user segments will react to AI-driven experiences. Power users may want more control and transparency. Casual users may prefer fully automated recommendations. Enterprise users may require explainability for compliance reasons. Selecting the right segment shapes every downstream decision, including the AI approach.

This is the step that separates AI PM answers from traditional PM answers. Spend 5 to 7 minutes here. Define the AI/ML approach at the right level of abstraction: you are not designing the model architecture, but you are choosing the general approach and justifying it.

For each feature you propose, answer three questions: What type of AI/ML approach fits? (Classification, recommendation, generation, detection, etc.) What data is required and is it available? What is the expected accuracy, and is that good enough for the use case? A search ranking feature with 80% relevance might be acceptable. A medical diagnosis feature with 80% accuracy is not.

Always discuss at least one alternative approach and why you rejected it. 'I would use a collaborative filtering approach for recommendations rather than content-based filtering because we have strong implicit signal data from user behavior. Content-based would require us to build a metadata taxonomy from scratch, which is a 3-month prerequisite.' This shows you are making a deliberate choice, not just naming the first thing that comes to mind.

Design (5 to 7 minutes): Sketch the user experience. For AI products, design must account for three states: the happy path (model is correct), the failure path (model is wrong), and the uncertainty path (model is not confident). Most candidates only design the happy path. Strong candidates design all three. How does the user recover when Smart Reply suggests something inappropriate? How does the recommendation system behave when it has low confidence? What does the UI look like when the model is loading or processing?

Evaluation (3 to 5 minutes): Define success at two levels. Model-level metrics: precision, recall, F1, BLEU, ROUGE, or whatever is appropriate for the task. Product-level metrics: engagement, conversion, retention, task completion, NPS. Explain how you would run an A/B test to validate the feature, including what guardrail metrics you would watch (e.g., false positive rate for fraud detection, user complaint rate for content moderation).

A common mistake is treating evaluation as an afterthought. At Google and Meta, evaluation is often where the interviewer probes deepest because it reveals whether you truly understand how AI products work in production. If you define clear evaluation criteria upfront, you demonstrate that you know shipping an AI feature is just the beginning.