How should an evaluation team choose the first handover boundary to model?

Start with a boundary where the receiving team's next decision is clear. A utility connection request moving from application review to field scheduling is a useful shape: the task has an identifiable sender, recipient, entry condition and next action. The pilot does not need to reproduce the whole organisation.

Specify what the evaluated agent controls. It might prepare the transfer, decide whether to accept it or resolve a returned case. Combining all three roles immediately can make a failure hard to attribute because the agent is effectively creating and judging its own work.

Agree the boundary of the simulation as well. If scheduling capacity is represented by fixed slots, say so. If customer replies are outside scope, define how unanswered questions affect completion. These limits are part of interpreting the evaluation, not incidental implementation notes.

A useful first request is a generated case specification with visible inputs, permitted actions, an expected end state and one contrasting failure case. Review whether those elements support a meaningful judgement before requesting broader variation or more connected workflows.