Clarify roles

Who owns prompt changes, who approves tool additions, who curates golden sets? Ambiguity here creates either bottlenecks or silent edits. Write it down as RACI aligned to your org—not generic “AI centre of excellence” slides.

Align product and platform on what “done” means for AI features: quality thresholds, latency budgets, and rollback triggers.

Rituals that match risk

High-risk domains get stronger review; low-risk internal tools can move faster with automated checks. Match ceremony to consequence.

Runbooks and handover

Operations needs playbooks: how to roll back a model version, how to drain a queue, how to escalate when evaluation fails. I include those in delivery work—not as optional docs.

Learning loops for the team

Schedule regular retros on incidents and near-misses involving AI outputs—not only infra outages. Capture prompt changes that fixed or worsened behaviour so knowledge compounds instead of living in chat threads.