A polished demonstration tells you little about how an assistant will handle your recurring work. A small trial with representative tasks gives you a better basis for comparison.
Select one routine task, one messy task and one case where the right response is to ask a question. Remove confidential details and write down the expected result before running the test. Give each candidate the same inputs and time budget.
Record factual errors, manual fixes and total time to an acceptable result. A fast answer that takes ten minutes to correct should be counted as ten minutes of work. Keep the examples so you can repeat the trial after a meaningful change.