Guides/Working with ML teams
Working with ML and data science teams
AI product work runs through a different partner than typical engineering collaboration — model iteration is empirical, timelines are probabilistic, and “done” is a moving target. Shared vocabulary, early alignment on evaluation, and explicit ownership handoffs keep the partnership from breaking down.
Ownership across the build lifecycle
Roles vary by org size — a startup may compress several of these into one person, but the phases and the question of who signs off on each one don't go away.
Frame problemPM
Assess data & feasibilityDS + PM
Build & iterateML
Evaluate & sign offML + PM
Ship & integrateEng + MLOps
Monitor & retrainShared
Where the partnership breaks down
| Friction point | Why it happens | Fix |
|---|---|---|
| Success metric defined after the model is built | ML optimizes a proxy metric with no product input up front | Define the eval rubric and business metric together before training starts |
| Model updates treated like feature releases | Model iteration is empirical and unpredictable, not fixed scope | Plan in confidence intervals and experiment budgets, not committed dates |
| No shared vocabulary for “good enough” | ML thinks in offline metrics (F1, perplexity); PM thinks in user outcomes | Translate offline metrics into a shared, product-level acceptance threshold |
| Monitoring ownership unclear post-launch | Data science moves to the next model; nobody owns drift | Assign an explicit monitoring owner as part of the launch checklist |