AI saves you an hour. A colleague spends it fixing your work

AI saves the person using it time. New US research shows a good share of that time then moves to a colleague who has to fix what they were sent, while the boss expects more. Teams that don’t agree three rules end up no better off.

Most AI users do save time

A survey asked 1,068 full-time US employees in computer-based jobs about their hours, their workload and how they use AI. Of the AI users who put a figure on it, 76% save at least an hour in a typical week and 32% save four hours or more.

Four hours a week across ten people is 40 hours, a full working week for one person. Few people check where that time went.

Where the hour goes

So you did save the hour. Your colleague spent it. Say you wrote the monthly report in forty minutes instead of two hours and sent it on. The person who received it spent fifty minutes checking the figures and rewriting the paragraphs that made no sense. You saved eighty minutes. They lost fifty. Across the two of you the team gained half an hour, and the only number anyone wrote down was yours.

Expectations go up too. Only 26% of AI users say they take on more work with the time they save. 39% finish earlier, and 15% say nothing really changes. Yet 46% say more is now expected of them.

Why the cost stays hidden

The saving and the cost sit in two different places. The person who saved the hour knows about it and will probably mention it. The person who spent the hour fixing the work often says nothing, because it feels like criticising a colleague. So the saving gets counted and the fixing never does. If you run the team, you hear the first number in the weekly update and never the second, because nobody writes “fixed the report” down as a task.

UK financial services firms are heavy AI users, and they keep a person in charge. A survey of AI across the sector found 80% of firms have AI agents deployed, and the same share run them for decision support, where the agent recommends and a person decides and acts. Only 2% run agents with broad autonomy, which the report defines as acting with limited real-time human involvement.

A team can copy that in ten minutes. Here are the rules I’d start with.

Three rules that keep the hour

Rule one works because the sender has the source open, knows what they asked the tool for and can spot a wrong figure in seconds. The reader has to work it out from scratch.

Rule two cuts the guessing. A lot of fixing time goes on working out which bits to trust. One line at the top answers that, and it makes rule one visible: if you say you checked the figures, you checked the figures.

Rule three takes two minutes and saves the next round of fixing. Quiet fixing teaches the sender nothing, so the same errors come back next week. A two-line note with your version attached is cheap, and it stops the same mistake coming back.

What to do this week

Take the three rules to your next team meeting and agree them out loud. Pick the one kind of work your team sends most, a weekly report, a client email, a board paper, and run the rules on that for a fortnight.

Then ask your team one question: how much time did you spend fixing other people’s AI work this week? If the answer is still an hour, check whether the rules are being followed. If it drops, that is the time AI was meant to give you. Ask it the same way each week, so the answers compare.

Sources: Hubstaff, State of the Workweek 2026 (1,068 full-time US employees in computer-based roles). Bank of England and FCA, Artificial intelligence in UK financial services 2026.