Discover and scope AI product opportunities
Identify where models add value, assess feasibility/data readiness, write an AI PRD with success metrics.
Tools: Notion/Confluence (PRD docs), Figma (clickable flows), Google/Claude AI Studio (feasibility spikes), Amplitude/Mixpanel (usage signal)
You should be able to…
- Run customer discovery sessions and translate business problems into scoped AI product capabilities
- Assess feasibility and data readiness before committing a roadmap slot to an AI feature
- Write a PRD with explicit success metrics, not just a feature description
- Prioritize which AI capability to build first given engineering cost and business impact
- Evaluate build-vs-buy: off-the-shelf LLM API vs fine-tuned model vs vendor product
- Define the launch bar — what 'good enough to ship' means for a probabilistic feature
- Say no to an AI feature request and explain the reasoning to stakeholders
Needs first: Explain how LLMs work and where they fail
The few resources that matter
Product requirements: how to write a great PRD
What is an AI Product Manager?
Introduction to Generative AI
Generative AI for Everyone
AI feature PRD: automated GST-invoice data extraction for an Indian SMB accounting tool
Interview (or role-play interviewing) 3 target users — small-business owners who manually re-key GST invoices into accounting software — and write a full PRD proposing an AI feature that extracts structured fields (GSTIN, amount, tax breakup) from photographed invoices. Include a feasibility assessment (data availability, expected accuracy, failure modes), success metrics (extraction accuracy, time saved per invoice), and a phased rollout plan with a clear launch bar.
- PRD includes problem statement, at least 3 real or role-played user quotes, and explicit success metrics
- Feasibility section names specific data sources and a realistic accuracy target with justification
- A launch bar is defined (e.g. '95% field-level accuracy on a held-out set of 200 invoices before GA')
- Rollout plan includes a fallback for low-confidence extractions (human review queue)
Evidence a recruiter can check
- Public PRD document (Notion/Google Doc, view access) linked from a portfolio site
- Notes or recording from the 3 user interviews (anonymized if needed)
- A one-slide feasibility summary a recruiter can skim in 60 seconds
Questions you'll get asked
- Walk me through how you'd scope an AI feature from a vague 'use AI for X' request.
- How do you assess whether we have the data to make an AI feature work before writing a single line of code?
- What goes into your AI feature PRD that wouldn't be in a normal PRD?
- Tell me about a time you said no to an AI feature — what was the reasoning?
- How would you prioritize between an AI feature that saves internal ops time vs one that's customer-facing?
- What's your process for deciding RAG vs fine-tuning vs a third-party API for a new capability?