Your status report now writes itself, your meeting notes arrive as you leave the room, and someone senior has started to ask what your middle managers do all day. Several chief executives have already given their answer:
- Jack Dorsey and Roelof Botha wrote, of Dorsey’s Block: There is no need for a permanent middle management layer.
- Brian Armstrong laid off 14% of Coinbase and replaced what he calls pure managers with player-coaches, in a hierarchy capped at five layers.
- Meta set up an applied AI engineering team at 50 people per manager, double the 25-to-1 ratio that is usually seen as the outer limit.
By September, Business Insider reported that Meta was asking some individual contributors in its Applied AI division to move back into management. Four people with knowledge of the matter told the publication so, the programme is opt-in, and Meta declined to comment. Nobody there has said why, so we must leave the verdict open.
Google ran a similar experiment in 2002 with no AI anywhere near it. David Garvin records that it lasted only a few months: They relented when too many people went directly to Page with questions about expense reports, interpersonal conflicts, and other nitty-gritty issues.
AI can answer the expense reports today. The interpersonal conflicts are still waiting for someone.
Three jobs, one title
The title covers three jobs:
- Routing: moving status up, decisions down, and context across.
- Judging: deciding whether the work is good enough, who is ready for more, and what stops.
- Answering: being the named person who takes the rating, the conflict, and the missed launch.
AI takes the first, adds to the second, and cannot do the third.
The case for flattening rests on the first job, and on that job I agree with it. Kai-Fu Lee writes in the FT (syndicated by CNA): An organisation is, in effect, an information system. Each layer that passes a message along can degrade it, and now an agent can read the source directly. Scholars predicted in 1958 that computers would eliminate middle management by the 1980s, and Lee’s view of the new attempt is This time it may work. Roles that only ever routed information should shrink.
I wrote in The IC Renaissance that the management ladder was a workaround for limited execution capacity, and I stand by that. It is an argument about the ladder: nobody should have to climb it to have impact, and nobody should climb it because they hit a ceiling. The job underneath the ladder is still there once the ladder goes. Even Lee keeps it: managers who can shoulder the consequences of what agents do remain central, in his telling.
To review, to people, and to read
Count the routing hours AI saved and you have half the sum. Luca Rossi argues that every hour saved gets spent on new work that AI itself creates: more output to review, slop in docs and code, and running the AI transition itself.
In my teams the hours went to review, to people, and to read.
Review is the queue I described in The AI Acceleration Trap. People is the job that was always there.
AI creates a lot of documentation, and a culture of writing has been an aspiration of ours for a long time. So the documents look like progress, but they create a bottleneck: who is going to read all of that, to do the other two jobs? You cannot judge work you have not read, and you cannot answer for a decision you only know from the summary. Every agent that writes a document adds to the reading of someone who already has a rating and a conflict on their desk.
I widened mine
I have widened my span of control and those of my top leaders, so I have no quarrel with the ratio moving. I am following outcomes to see if it’s working, and AI helps with that: data is much more accessible. That is the first job shrinking in my favour.
The outcomes tell me whether the bet is paying. They do not tell me who picks up the work that a wider span pushes onto fewer people. Block’s essay admits that parts of it will likely break before they work, and the question is who is holding those parts when they do. The chief executive decides. The remaining managers and senior individual contributors pay, with reviews they cannot finish and conflicts that land on whoever is nearest. A player-coach owes the same ratings and the same conflicts as a manager, with fewer hours to give them.
Google in 2002 had no AI to take the expense reports, so its experiment proves nothing about routing. It shows where people go when nobody is there to answer.
Before you change a ratio
Four questions, in the order I would ask them:
- Which hours has AI actually removed from a manager’s week? Measure them.
- Where did those hours go: to review, to slop, to reading, or to running the AI transition?
- Who answers for each outcome: the missed launch, the rating, the conflict?
- What would make you reverse the decision? Write it down before you cut.
Your status report will keep writing itself. Who picks up the conflict it doesn’t mention?
Leave a Reply