A junior employee drafts a document with AI. Their manager has AI read it and review it. The junior pastes the returned red ink right back into AI.
It's a hypothetical. Still, I wouldn't be surprised if it were already an ordinary scene in some meeting room. I'm no exception; I have AI helping with my work nearly every day.
Watching that loop, a thought creeps in. What are the humans in the middle actually doing? If they're questioning assumptions and adding their own judgment, there's a reason for them to be standing there. But if they're just passing output to the next person, they're no different from someone standing in a bucket brigade. Why not wire the steps directly together?
Thinking about that, I remembered an old book. The Mythical Man-Month.
Adding people doesn't make it faster
What Frederick P. Brooks wrote in The Mythical Man-Month was that adding people to a late software project makes it later. The book is half a century old and still quoted everywhere. Which suggests we haven't learned the lesson very well.
A job that takes one person 20 days should take four people 5. I understand the temptation. But for that, the work has to split cleanly and everyone has to be able to start running at once. In reality, some work can't move until the design is settled, and once you split things up, aligning and stitching them back together takes effort of its own. Brooks himself said in a later interview that more people means repartitioning the work, teaching the newcomers, and more people to talk to.
Of course, adding people doesn't always backfire. It depends on the shape of the work and when people join. But we do seem overly hopeful about how much the extra hands will help. A study by Staats, Milkman, and Fox named the tendency to underestimate the labor required as teams grow the "team scaling fallacy." It's a grand name, but it's really just an academic name tag pinned on the optimism that "more people will sort it out."
Nor is division of labor itself the problem. As Becker and Murphy's work argues, splitting by specialty lets each person play to their strengths. In exchange, bringing those strengths together takes coordination. From sales to planning, from planning to development. The finer you slice the roles, the more times the bucket changes hands. If everyone talks to everyone, n people have n(n−1)/2 paths between them. Then again, that's a count of roads, not proof that someone walks every one of them each day.
AI agents, too, are handed goals, have their work checked, and get fitted into the whole. They likely need less teaching than people, so it isn't quite the same story. Even so, the number and skill of the workers alone don't decide how long completion takes. That much, I think, stays the same.
Delivered, understood, nodded
When more people get involved, things get harder to decide. Where does that time go?
I suspect three separate events are tangled together there. The information arriving. The meaning being understood. And people nodding along.
In "Grounding in Communication," Clark and Brennan discuss the process of not just conveying words but mutually confirming that the other person has understood. Handing out documents doesn't mean anyone has understood each other. A spec may say "real-time," and one person pictures under 1 second while another thinks every minute is plenty. You explain, you ask, you rephrase, and only then do you start seeing the same landscape.
And understanding each other doesn't mean you'll nod. Sales cares about the deadline; development cares about quality. Both can fully grasp the other's situation and still disagree. Veto player theory in political science holds that the number of people whose consent a change requires, and the distance between their positions, shape how easily change happens. Applying it to a company is only an analogy, but the idea of asking "whose nod is required" rather than "how many people are there" makes a strange amount of sense to me.
Hearing every voice, understanding each other, and not deciding until everyone nods. Bundle those three together, and I suspect decisions turn heavy as lead.
Faster hands, the same old line
AI makes the hands carrying the buckets faster. A document takes 10 minutes, and the code is written before you notice.
But if the approver's calendar doesn't open until next week, that document sleeps on a desk until next week. If the request was misread, the faster you built it, the faster you get to rebuild it. The faster our hands get, the more sharply the time spent waiting, understanding, and nodding stands out. I can't shake that feeling. And with more output, there won't be enough eyes to read it all.
So rather than how much AI we used, I want to look at how long it took from being asked to reaching the other person's hands.
Some people are trying to change the shape of the line itself. In "From Hierarchy to Intelligence," Jack Dorsey and Roelof Botha describe a plan to move the information-gathering and coordination that Block's hierarchy used to handle into a system. They flatly say permanent middle management is unnecessary, while also admitting, interestingly, that the transition has only just begun. NRI's "AI and the Future of Organizations" likewise ties hierarchy and specialization to the limits of human cognition, and sees organizations possibly changing shape if AI loosens those limits.
What snagged my attention was that Block's plan assumes decisions and progress inside the company are recorded. Context that lives only in someone's head can't be retrieved without going through that person. Deleting a box from the org chart is easy. But if whatever lived only inside that box disappears with it, you've lost the point.
As for how far AI can stand in for a group of people, we're slowly learning a few things. In "The Cybernetic Teammate," an experiment with 791 people at P&G, one person with AI produced work of about the same quality as a two-person team without it. Their proposals also leaned less toward their own specialty. AI may be able to take over part of the reason we brought people together. But it was a one-day product proposal experiment. Long-term operations, developing people, and negotiating between departments at odds weren't part of it. The "Work Chart" in Microsoft's report, pairing people and AI around goals, is likewise still a vision.
An approve button you only press
One role often said to remain for humans is checking at the end and taking responsibility.
But if more output arrives every day than anyone could read, and the job becomes pressing an approve button, how much meaning is left in that check? The finger pressing the button may get stronger, but the ability to judge won't.
You can have another AI check what one AI made. But if something important is missing from the source material, both may miss it together. A 2023 study of 758 BCG consultants, presented by HBS, argued that AI's strengths and weaknesses are jagged across tasks. Today's models are surely smarter. Even so, the question of whether similar work can be delegated the same way hasn't aged.
Rather than adding more checks, I want to decide first what each check is for. Calculations get recalculated. Specs get asked about with the customer. Code gets run. And if you hand someone responsibility, I want to hand over the authority to stop things, the material to judge with, and the time to check, too. Handing over responsibility alone is like giving someone an empty bucket and sending them to the fire.
Rereading The Mythical Man-Month in the age of AI, the question looks a little different. It isn't only how much we can increase our capacity to work. It's what we need to confirm with each other, and whose nod we need, to get that work where it's going.
No matter how fast the hands passing the buckets get, if the line is arranged the way it always was, the water won't reach the fire much sooner. When both juniors and managers use AI, where in that line are humans judging what, and learning what? I want to work somewhere that can actually explain that.
References
This article is based on sources as of October 30, 2026. The business examples are hypothetical. Applying prior research to AI is my own view.
- Frederick P. Brooks, Jr., The Mythical Man-Month: Essays on Software Engineering, first edition 1975, anniversary edition 1995. On the costs of adding people, see also his interview Quoted Often, Followed Rarely (Fortune, December 12, 2005, reprinted on a university site).
- Bradley R. Staats / Katherine L. Milkman / Craig R. Fox, The team scaling fallacy: Underestimating the declining efficiency of larger teams, Organizational Behavior and Human Decision Processes, 2012. Public PDF.
- Gary S. Becker / Kevin M. Murphy, The Division of Labor, Coordination Costs, and Knowledge, The Quarterly Journal of Economics, 1992.
- Herbert H. Clark / Susan E. Brennan, Grounding in Communication, Perspectives on Socially Shared Cognition, 1991.
- George Tsebelis, Decision Making in Political Systems: Veto Players in Presidentialism, Parliamentarism, Multicameralism and Multipartyism, British Journal of Political Science, 1995. Applying it to corporate decision-making is an analogy made in this article.
- Roelof Botha / Jack Dorsey, From Hierarchy to Intelligence, March 31, 2026. An essay by the people involved describing Block's planned transition.
- Kazunari Dobashi, Tomoki Takemura, Ken Mori, AI and the Future of Organizations (in Japanese), Chiteki Shisan Sozo, March 2026. This article refers to the summary on the public page.
- Fabrizio Dell'Acqua et al., The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork, Organization Science, published online June 12, 2026. Refers to the journal version's text and abstract, not the 2025 working paper.
- Microsoft, 2025: The year the Frontier Firm is born, April 23, 2025. A report drawing on a survey of 31,000 people across 31 markets, among other sources. I treat its organizational proposals separately from causal evidence for them.
- Harvard Business School AI Institute, Navigating the Jagged Technological Frontier, September 21, 2023. Refers to the research institute's summary of the study.