AI use is growing fast. Measure what comes back.

Pick one repeatable task for two weeks: preparing a customer follow-up, cleaning a spreadsheet, assembling a research brief, or drafting a proposal. Keep the before-and-after comparison ordinary.

For every task, record whether the result was accepted as-is, edited, sent back for correction, reopened later, or completed manually after the assistant stopped. Add the minutes spent checking it and the minutes spent fixing it. If the task affects another person, note whether they had to ask a follow-up question or repeat information.

This is not an anti-AI scorecard. A thoughtful review is often the right price for useful work. The problem is claiming that review does not exist.

“More work delegated” is a product-use measure. “Fewer loose ends after the handoff” is closer to an outcome. The difference shows up quickly.

A sales follow-up is not done when a draft exists; it is done when the right person receives a correct message and the next step is clear. A research brief is not done when it has citations; it is done when someone can use it without spending an hour rebuilding the argument. A spreadsheet is not done when the rows are filled; it is done when the numbers reconcile and the person downstream does not discover a quiet mistake.

Choose a finish line before the experiment begins. Then keep the difficult cases in the sample. If the assistant only looks good after the messy work is quietly removed, the team has learned something useful—but not the thing the dashboard promised.

There is a second trap in AI reporting: a team may save time on the first pass and immediately spend it on more volume. That can be a business choice, but it is not the same as making people’s days lighter.

Ask one plain question at the end of the pilot: what stopped happening? Maybe it is the weekly status chase. Maybe it is the late spreadsheet cleanup. Maybe it is the repeated context dump when a coworker takes over. If nothing disappears and the same people are still doing the final check at night, the AI may be generating capacity without returning any margin.

The broader enterprise numbers suggest that more work will be handed to AI this year. The teams that learn fastest will not confuse a bigger activity graph with a cleaner Tuesday.