Check three things separately: whether the output met the requirements, what the full attempt cost and whether anyone used it to do something useful. Count the failed drafts and reviews too. Compare with the current way of doing the same job. You can keep work that earns its place and stop producing a report simply because AI can.
Count setup, finding inputs, corrections, retries, review and tool costs, not just the seconds the final answer took. Use the same quality checks when comparing with the old method. You see whether work was actually saved or just moved into a less visible part of someone’s afternoon.
Ask which decision or action the output was meant to support. If there is no useful use, stop the task or choose a better assignment. If evidence is thin, agree what one further test must show and what it may cost. You protect time and money from an experiment that has become a habit.