Test the ordinary and the difficult
A demo shows that something can work. A production evaluation asks when it fails. Use representative tasks, incomplete inputs, unusual cases, and examples that require the system to decline or ask for help. Agree on acceptance criteria with the people who will use it.
Design the handoffs
Define which actions need approval, what happens when a system is unavailable, and who receives exceptions. Avoid silent failures. A workflow needs clear ownership at every step, including the moments when automation cannot continue.
Make operation visible
Capture the information needed to understand quality, latency, cost, and errors without collecting unnecessary sensitive data. Assign an owner for incidents and a process for evaluating changes to prompts, models, and data sources.
Treat adoption as delivery
Introduce the system in the workflow where it belongs. Explain its boundaries, train the people who use it, and keep a channel for feedback. Roll out gradually enough to learn, with a practical fallback if the new process does not behave as expected.