Block's chief executive Jack Dorsey made the case in March that AI can replace the manager whose main job is passing information along, a few weeks after Block cut about 4,000 jobs. The case is half right. AI is very good at passing information along. But new research on Americans who use AI at work shows that the manager's other job matters more to whether people end up better off with AI.
What AI takes over, and what it leaves
Take a warehouse shift manager. Every shift she collects numbers from team leads, checks what is behind, and writes a handover note. An assistant that reads the scanner logs and drafts that note could give her back about half an hour a day. Over a month of working days, that is roughly ten hours.
Dorsey's argument stops there: the ten hours are savings, so one manager can cover more people, or the job can go. The research points to a different use for them. Of the 14 factors most closely tied to people feeling that AI benefits them, 11 are things a manager controls: whether they feel safe trying it, whether good use gets noticed, whether the manager shows how they use it. The single strongest one is feeling involved in how AI fits their job.
The people who hand out a say already feel they have one
Managers are almost twice as likely as their staff to feel they have a say in how AI fits their work: 67% against 37%. A second gap has the same shape. More than half of managers believe they make their own AI use visible to their team, and about three in ten staff say they see it.
In both cases the manager thinks the job is done and the team does not see it. That makes the freed ten hours valuable only if they are spent where staff can watch: sitting with two pickers, trying the assistant on a real shift, and showing the attempts that failed as well as the ones that worked. That is slow, in-person work, and it is the part of management AI cannot do for her.
Why cutting layers on the routing argument backfires
Google's 2002 experiment shows what happens without managers: the questions did not disappear, they travelled upward until they landed on one person. A company that cuts a third of its managers for AI reasons will see the same thing in a new form. Each remaining manager has more people, less time to sit with any of them, and the questions about whether their job is changing go unanswered. Staff will not feel they have a say, and by the numbers above, that is the factor that most separates people who gain from AI from people who do not.
My forecast is that the firms that cut first will report the routing savings quickly, and slower, less visible losses afterwards: staff who use AI less, or use it for the wrong things, because nobody helped them decide. The companies that redraw the work first, as we argued earlier, and only then the org chart, avoid that.
If you run a team, there is a cheap test. Have the team decide together where AI fits its work and where it does not. A month later, ask each person one question: did your manager involve you in how AI fits your job? Compare the answers with what the managers believe. If the two columns disagree the way these did, the problem is not the tool.