When this article first published on June 8, it opened on a statistic we cannot stand behind. It said a study out of Harvard Business School found that roughly six in ten managers spend more than half their week on administrative work. Going back for the source, we could not find that finding in any Harvard Business School research, and the HBS work that does exist on managers and AI does not contain it. We should not have published a number we could not point to. The error is ours.
The underlying idea, that managers spend a great deal of their week on administrative and other non-managing work, does hold up, on different evidence. McKinsey's research on middle managers found they spend close to half their time on nonmanagerial work, including roughly one full day a week on administration, with much of the rest going to individual-contributor tasks rather than managing. That is the sourced version of the point the article was making.
A second claim needs correcting. The article said Shopify, Klarna, and Duolingo had thinned their management ranks on this logic. That is not quite what those companies did. Their moves were AI-first headcount and contractor decisions rather than cuts aimed at management layers, and Klarna has since walked part of its reduction back after service quality fell. The flattening of middle management is a real trend, but these three are better described as AI-first staffing examples than as management cuts, and we conflated the two.
What stands is the spine of the piece. Transformation efforts do fail at around seventy percent, a long-running McKinsey finding, and the more important point, that the failures live in the human layer rather than the strategy, is made directly in a March 2026 Harvard Business Review article by Jenny Fernandez on why transformations stall when senior leaders misread silence as agreement. The argument that an organization automates the visible paperwork while losing the judgment underneath it does not depend on the two claims we got wrong. But we did get them wrong, and the honest thing is to leave the original below and say so here.
The slide had a row of boxes sitting between the executive team and the people who shipped the product, and the senior director running the deck called that row a layer. Not names. Not roles. A layer, the way you would describe insulation you were about to pull out of a wall. By the time someone talks about part of an organization that way, the decision has usually already been made, and the meeting is only the place where it gets said out loud.
The case for cutting it was clean. A study out of Harvard Business School found that roughly six in ten managers spend more than half their week on administrative work, the status updates and the scheduling and the report assembly that an AI agent can now produce faster and without complaint. Shopify, Klarna, and Duolingo had already thinned their management ranks on something close to that logic. If most of what a manager does in a given week is paperwork, and the paperwork is now automatable, the math looks like it finishes itself.
It finishes itself only if you believe the paperwork was the job.
The function hiding inside the activity
The administrative work that fills a middle manager's calendar is the visible residue of something that never appears on a calendar at all. When a manager assembles a status report, the report is not the point. The point is that, to assemble it, she had to notice that two teams were quietly solving the same problem twice, that a deadline everyone had signed off on was being held together by one engineer working weekends, that the process the executives were proud of was being routed around on the floor because it added a step nobody could defend. A little of that surfaces in the report. Most of it does not. It lives in the judgment of the person who sits close enough to the work to feel it move.
McKinsey's State of Organizations 2026 puts the failure rate of transformation efforts at around seventy percent, and it locates the cause not in the strategy but in the human layer that was supposed to carry the strategy into the work. HBR made a harder version of the same point in March, finding that transformations stall when leaders read silence as agreement and cannot detect resistance until it has already hardened. The layer that catches that resistance early, that turns an abstract priority into a plan a team can actually run on Monday, is the same layer being described with a laser pointer as overhead.
You can automate the report a manager produces. You cannot automate the reason she was the one who noticed there was something to report.
What you are actually deciding when you cut
When an organization removes that layer because a model can generate the artifacts the layer used to produce, it is making a quieter decision than it thinks. It is deciding that the translation of strategy into operational reality will now happen somewhere else, or will not happen at all. It is deciding that the early warning of a transformation going wrong, the part that used to arrive as a manager's uneasy aside in a one on one, will have to be caught by someone with less time and more distance from the work. The model can write the summary. The model cannot sit in a room and feel that a team has stopped believing in the plan. Somebody still has to own that, and a reorg almost never says who.
The cost does not show up in the quarter you make the cut. It shows up two or three quarters later, when a program that looked on track in every dashboard turns out to have lost the people who would have told you it was drifting. The dashboards still read green. They read green because the work of keeping them honest had been spread across a layer that somebody had decided was clerical.
In Agile Sucks! (When You Do It Wrong), James Wright and I open with a pattern we call successful failure, organizations where close to seventy percent report healthy metrics while producing no real business outcome. The mechanism is the one running underneath this whole question. A system measures the activity it can see and loses track of the function it cannot, then congratulates itself on the efficiency right up until the outcome fails to arrive.
So before the next headcount review turns a row of people into a layer, it is worth asking a plainer question. If you ran a time-study on your own middle managers tomorrow, how much of what actually makes them valuable would even show up in it? And once the administrative half is automated, who in your organization is going to own the half that was never on the calendar to begin with?
In Agile Sucks! (When You Do It Wrong), James Wright and I open with a pattern we call successful failure: organizations where close to seventy percent report healthy metrics while producing no real business outcome. The same mechanism is at work when a company automates the work it can see and loses the function it cannot.
Read Agile Sucks! (When You Do It Wrong) →