Two hours a week is a quarter of a workday. In one large experiment, office workers got those hours back, and the shape of their week didn't change.
Researchers led by Eleanor Dillon gave an AI writing assistant, built into the software people already use every day, to randomly chosen workers: 7,137 people across 66 companies, each one picked individually rather than as part of a team. The ones who used it spent about two fewer hours a week on email and did less work after hours. But the app data, logged automatically, showed no real change in their meetings or shared documents.
I've used this study before to argue that saved time doesn't automatically become different work. The obvious follow-up is what would make it.
Changing meetings, or who owns a shared document, requires "coordinating with colleagues and agreeing on new norms," the researchers write. A weekly meeting is an agreement among, say, eight people. Hand one of them a better tool and the agreement doesn't budge. Email got faster because email is mostly something you do alone.
The study wasn't built to test whole teams changing together. Even in the groups with the most access, only 39 percent of people had the tool on average. Numbers in the appendix hint that people with more tool-using coworkers cut email and meetings further, but simpler comparisons didn't find anything clear.
The economist Paul David's study of factory electrification offers a parallel. Electric motors were around for decades before factories ripped out the overhead shafts tying every machine to one engine and laid out their floors around how the work actually flowed. The big productivity gains of the early 1920s came with that redesign.
Software went through a smaller version of this. For years, developers handed code over a wall to a separate operations team, usually as a ticket plus a deploy window. When it fell over at 3 AM, the pager went to someone who hadn't written it and couldn't fix it without sending a request back over the wall. The developers heard about it the next morning, if at all.
Better deploy tools kept arriving. Mostly they let us throw code over the wall faster.
What stuck was moving ownership. You build it, you run it: the team that writes the service carries its pager. Once people were getting woken up by their own code, they started writing logs they could read, alarms they believed, and changes small enough to roll back. Nobody held a meeting to agree on those habits; being on the hook produced them. And once ownership moved, faster tools had something useful to speed up.
Apply that to AI at work. Redistributing tasks is easy to do on a Monday and quietly undo by next quarter. Moving who owns an outcome is harder to reverse, because the consequences keep enforcing it.
One possibility is that durable reorganization comes from handing a whole result to a whole group: everyone who shares the meetings gets the tool, and together they own something end to end, a launch or a customer account, that used to be split across functions. A lot of recurring meetings and shared documents exist to shuttle information between people who each hold one piece. Put all the pieces in one team and those handoffs lose their reason to exist. (My hunch, not a finding: status meetings go first.)
That also gives the saved hours somewhere to go. Two hours per person across eight people is about two full workdays a week. Left unclaimed, they turn into fewer late-night emails, which is a real benefit and worth keeping. They turn into different work when a team owns a result that needs them.
Someone still has to decide which results a team owns. The assistant won't make that decision for them.