Common Product Sense Mistakes

The most common mistake is hearing 'design an AI feature' and immediately jumping to a solution without understanding the user or problem space. Interviewers see candidates who, within 30 seconds of hearing the question, start describing a recommendation algorithm. This signals that you are a technologist looking for a problem, not a PM who starts with the user.

The fix: Force yourself to spend the first 3-4 minutes on Audience (who is this for?) and Problem (what specific pain point are we solving?). State your user segment and problem statement explicitly before proposing any solution. 'Before I design a solution, let me make sure I understand the user and the problem.' This single sentence resets your approach and signals structured thinking.

The second mistake is proposing 'we will use AI to...' without specifying the approach, data requirements, or expected performance. This is equivalent to saying 'we will use code to build it.' It communicates nothing about your understanding of how AI systems work.

The fix: Every time you propose an AI-powered feature, immediately specify three things: the type of AI approach (classification, generation, recommendation, etc.), the data it requires, and the expected accuracy level and whether that is sufficient. 'We would use a transformer-based embedding model for semantic search, trained on our product catalog and click-through data. We would need at least 6 months of click data, and we should target NDCG@10 of 0.65+ to outperform the current keyword search.' This level of specificity is what separates a 3 from a 4.

AI products fail. Models make wrong predictions. Latency spikes happen. Data pipelines break. If your design only covers the happy path, you are telling the interviewer that you have never shipped an AI feature in production. Real AI PMs spend as much time designing the failure experience as the success experience.

The fix: For every AI feature you propose, explicitly describe what happens when the model is wrong, not confident, or unavailable. 'When the recommendation model returns results below our confidence threshold, we fall back to popularity-based ranking and show a subtle indicator that these are trending items rather than personalized recommendations.' Also discuss what happens during cold-start (new users with no data) and model degradation (performance drops over time due to data drift).

Saying 'we would measure engagement and retention' is not an evaluation plan. It is a vague aspiration. Interviewers want to see specific metrics tied to the specific AI feature, with clear success thresholds and a testing methodology.

The fix: Name the exact metrics, the baseline you are comparing against, and the threshold for success. 'We would measure the search-to-purchase conversion rate, comparing AI-ranked results against the current keyword-ranked results. Our success threshold is a 10% relative improvement in conversion. We would run a 2-week A/B test at 10% traffic with a guardrail on page load latency (must stay under 200ms P95).' Specificity demonstrates experience.

Mistake 5: Boiling the ocean. Candidates propose a comprehensive AI platform when the interviewer asked for a single feature improvement. The fix: start with one focused feature, execute it thoroughly, and then mention future extensions briefly at the end. 'For V1, I would focus specifically on [X]. The natural extensions for V2 would be [Y] and [Z], but I want to make sure V1 is solid first.'

Mistake 6: Ignoring personalization tradeoffs. Every AI feature that uses personal data creates privacy implications. Candidates who do not mention this signal a blind spot that worries interviewers. The fix: whenever you propose a feature that uses personal data, spend 30 seconds on the privacy consideration. 'This requires access to browsing history, so we need clear consent, a data retention policy, and an opt-out mechanism. For users who opt out, we fall back to non-personalized results.'

Mistake 7: Running out of time. Product sense rounds are typically 35 minutes. Candidates who spend 15 minutes on Audience and Problem have 20 minutes left for Intelligence, Design, and Evaluation. This is not enough. The fix: use a time budget. Audience: 4 minutes. Intelligence: 7 minutes. Design: 7 minutes. Evaluation: 5 minutes. Questions and discussion: 12 minutes. Practice with a timer until this becomes natural.