How I think about AI product management

Probabilistic outputs, trust, and why the hardest work is defining “good” before you touch the model.

Most AI features fail in the specification layer, not the API layer. Teams jump to prompts and retrieval when they haven’t agreed what a successful session looks like for the user — or for compliance.

I start from friction: what manual loop is expensive, error-prone, or slow? Then I ask what humans must still own (approvals, escalations, audit logs) and design those paths as first-class product behavior, not exceptions.

Evaluation is a product surface. For campus RAG and internal agents alike, I care about golden questions, regression sets, and clear fallback when confidence is low — because “the model said so” is not a release strategy.