AI Product Strategy
Decide where AI creates real value: AI-first vs AI-added products, moats and data flywheels, build vs buy vs partner, model choice, opportunity assessment and AI roadmaps.
AI Product Strategy
Decide where AI creates real value: AI-first vs AI-added products, moats and data flywheels, build vs buy vs partner, model choice, opportunity assessment and AI roadmaps.
In one line
AI product strategy is deciding which problems AI should solve for your users and business, how you'll win (data, distribution, workflow, trust), and what to build, buy or partner for, before writing a single PRD.
The 6 classes
Where AI creates value
Separate real AI opportunities from hype.
- Start from the user job, not the model.
- Plot value vs feasibility.
- Error tolerance decides how much autonomy AI gets.
Moats and data flywheels
Understand what makes an AI product hard to copy.
- Models are commodities; data, workflow and trust are moats.
- Design feedback loops from day one.
- Measure the flywheel: does quality improve with usage?
Build vs buy vs partner, and choosing models
Make sourcing and model decisions with clear trade-offs.
- Buy non-core, build core, partner for missing assets.
- Choose models with your own evals and cost per task.
- Avoid lock-in with an abstraction layer.
Assessing and prioritising AI opportunities
Score AI ideas on value, feasibility and risk.
- Prototype + eval before roadmap commitment.
- Compare cost per task with value per task.
- Price in the cost of errors.
AI roadmaps and portfolio of bets
Plan AI work under uncertainty.
- Balance quick wins with bigger bets.
- Include evals, data and safety work in the plan.
- Set kill criteria for each bet.
Case studies and final project
Apply the strategy tools end to end.
- Strategy = where to play + how to win + what to build.
- Tie every bet to evals and metrics.
- Revisit as models change.
Interview questions
How do you decide where to use AI?
Start with frequent, painful jobs; check value, feasibility (prototype + eval), risk of errors and cost per task; plot on a value-feasibility matrix.
What is a moat for an AI product?
Proprietary data and context, workflow ownership, distribution, trust/quality in a domain, and fast learning loops; not the model itself.
Build vs buy for AI?
Buy non-core needs, build the core experience on model APIs, partner for missing data or distribution; consider lock-in, data policies and cost.
How do you choose a model?
Your own evals, cost per successful task, latency, context, data policies and residency, multilingual quality, with an abstraction layer to switch.
What is a data flywheel?
A loop where usage creates data and feedback that improves quality, attracting more usage.
How do you roadmap AI under uncertainty?
Outcome roadmaps with confidence levels, three horizons, quarterly review, and kill criteria tied to evals and adoption.
