A pilot is a decision tool. Start by naming the uncertainty you need to resolve: can a workflow become more useful, more reliable or easier to operate?
Choose a bounded task and a small, representative evaluation set. Record the baseline, expected failure modes and who will review the output. Make the stop criteria as explicit as the success criteria.
Build enough to learn. Review the evidence with the people who will use and operate the system. The best next step may be to scale, adapt or stop. All three can be useful outcomes.