Choose the decision
State what the pilot should help decide: whether a specific task can be supported, what quality is acceptable and what needs human review.
Prepare representative examples
Collect realistic inputs, difficult cases and expected output characteristics. Separate evaluation examples from any examples used to tune the workflow.
Test more than accuracy
Review answer usefulness, source handling, permissions, failure paths, latency and operating costs. Document the setup so results can be interpreted.
Set a go/no-go review
End the pilot with a recommendation, unresolved limitations and a scoped next step. A successful demo is not automatically a production release.
Make it specific to your situation.
Use this guide as a starting point. The next useful step is to record the real context, identify the decisions that need an owner and agree what evidence would support the decision.