ai-product-strategy

Start with the user problem, not the AI capability. Identify minimal-impact use cases to learn current model capabilities, then gradually increase agency as reliability confidence grows. Avoid adding AI features without a clear user problem.

Build proprietary AI only where you have a data advantage or it is core to the business. Partner with best-in-class providers for commodity needs like foundation models, and design flexible architectures that let you swap models as they improve.

It is the framework defining what the algorithm decides versus what humans decide. The core PM task is assigning responsibility: give machines optimization within constraints while humans set strategic intent and handle judgment calls.

Because the 1% of failures can severely damage user trust if the UX assumes perfect outputs. Design for squishiness by building interfaces that gracefully handle incorrect or non-deterministic AI responses.

Use specialized models when tasks have different requirements, such as reasoning versus speed versus coding. Deterministic engines should handle logic and calculation while LLMs translate results into natural language.

Common mistakes include AI for AI's sake, single-model thinking, ignoring failure cases, static architectures that cannot evolve with model improvements, skipping evals and observability, and over-automating away valuable human judgment.