Production asks questions a demo never does
What happens when the user asks something outside the scope? Who is allowed to see which information? Where does the audit trail live? What does the system do when the upstream API is slow, or when the document it needs has been edited since it was indexed? How do you know it got worse last Tuesday?
These are not edge cases. They are the normal operating conditions of an organization, and each one needs a decision made deliberately rather than discovered later.
Governance is engineering, not paperwork
It is tempting to treat governance as the documentation you write after the build. In practice, the controls that matter are structural: confidence thresholds that route uncertain cases to a person, permission boundaries that follow the role of who is asking, approval gates before anything with a financial or contractual effect, and an audit trail that answers "why did it do that" months later.
Built in, these are ordinary engineering decisions. Bolted on afterwards, they are a rebuild.
The measure has to exist before launch
You cannot tell whether a deployment worked if nobody agreed what "worked" meant beforehand. Completion rate, escalation accuracy, time to first response, groundedness against a source set, cost per completed task, pick the two or three that map to the outcome the sponsor actually cares about, and instrument them before the first user touches the system.
A system with a baseline and a measure survives scrutiny. One without either becomes an argument about impressions.
The point is completion
An answer is not a result. The step that separates a useful system from an interesting one is whether the work moves forward: the booking is made, the record is written, the request is routed, the person is notified. If the system stops at producing text, a human still has to do everything that matters.
Start with the task that finishes, not the conversation that impresses.
