Start with the decision
Before choosing a model, define the decision or task the system supports. Identify who uses the result, what a useful answer looks like, and the consequence of an incorrect one. A narrow, measurable objective makes evaluation and ownership much clearer.
Evaluate the whole system
Model quality is only one part of performance. Test retrieval, permissions, integrations, latency and cost against representative tasks. Keep an evaluation set that includes difficult inputs and failure cases, and review it as the product changes.
Design for uncertainty
Give users a clear route to review, correct or reject output. Establish fallback behavior when evidence is missing or an upstream service is unavailable. For consequential actions, decide explicitly where a human must remain in control.
Make ownership operational
Assign responsibility for monitoring, changes and incident response. Record model versions, configuration and release decisions. Production readiness is the ability to operate the system on an ordinary Tuesday, as well as on launch day.