A junior employee drafts a document with AI, and their manager reviews it with AI. The junior then feeds the manager's revision notes right back into AI. It's a hypothetical, but I'd bet some meeting room somewhere is already living it.
Watching that loop, I can't help wondering what the humans standing in the middle are doing. If they're checking assumptions and adding their own judgment, they've earned their spot. But if they're just taking output and passing it to the next person, they're not doing much more than standing in a bucket brigade. Why not wire the steps directly to each other?
Somewhere in that line of thought, The Mythical Man-Month came to mind. Adding people doesn't shrink time to completion in proportion to headcount; that's the whole point of the book. Even if AI makes each person's hands faster, I suspect a similar wall is still standing.
What I want to think about is how to build an organization once work capacity has grown. If handing off information, building shared understanding, and reaching agreement all stay the way they've always been, how much shorter does the time to deliver results actually get?
Person-months and calendar time don't trade freely
Brooks's law, made famous by Frederick P. Brooks's The Mythical Man-Month, says that adding people to a late software project makes it later. The book is half a century old and still quoted constantly, which suggests we haven't quite learned the lesson.
The back-of-the-envelope math of "one person takes 20 days, so four people take 5" only works if the job splits into independent pieces that can run in parallel. In reality, some work can't start until the design settles, and splitting the work creates its own alignment and integration chores. Total effort and time to completion belong in separate ledgers.
Brooks himself said in a later interview that adding people means repartitioning the work, spending existing members' time training newcomers, and multiplying the people everyone has to talk to. That said, it isn't a universal law that every addition backfires. How well it works depends on the dependencies in the work and when people join.
AI agents, too, need to be handed goals and context, have their output checked, and get folded into the whole. Of course they don't need a human's onboarding period, so assuming the same cost structure would be sloppy. Still, they share one thing: the number or capability of the workers alone can't fully explain how long completion takes.
The gains of division and the burden of coordination
Nor is this just folk wisdom from the software industry.
A 2012 study by Staats, Milkman, and Fox drew on two experiments and project data from a software company to show that people tend to underestimate the total labor required as teams grow. They can't fully account for the coordination and integration burden relative to the gains of dividing the work. It goes by the name "team scaling fallacy." As someone prone to optimistic estimates, I find that a little too close to home.
That said, the study doesn't claim that more people always means slower work. Adding people has real advantages. We just tend to overestimate how big they'll be.
Becker and Murphy's 1992 theoretical work likewise weighs the productivity gains of specialization against the cost of coordinating specialists who complement each other. Split the work among sales, planning, development, and operations, and each group can make full use of its expertise. On the other hand, you create touchpoints for conveying customer requests, reconciling interpretations of the spec, and checking the impact of changes. The finer you slice the roles, the more you have to design the bucket handoffs themselves.
If everyone talks to everyone, n people have n(n−1)/2 possible channels between them. But that's strictly a count of potential channels. It doesn't prove that actual communication volume or cost always grows quadratically. This is also exactly where organizing areas of responsibility and points of contact earns its keep.
Sharing, understanding, and agreeing are different things
When more stakeholders mean decisions take longer, where is that time actually going? I want to split it three ways.
| Process | What to check | Possible problem |
|---|---|---|
| Information sharing | Has the necessary information arrived? | Documents or changes didn't reach people |
| Shared understanding | Do people understand the meaning and assumptions? | The same term is used with different meanings |
| Agreement | Which option will people accept? | Priorities differ even with understanding |
In "Grounding in Communication," Clark and Brennan discuss the process of not merely transmitting an utterance but mutually confirming that the other party has understood it well enough for current purposes. Handing out documents and arriving at shared understanding are entirely different events.
Say a spec contains the word "real-time." One person pictures updates within 1 second; another thinks an update every minute is perfectly fine. You explain, ask, rephrase, and confirm, and only through that back-and-forth does the needed understanding line up. The theory itself doesn't measure headcount against time. But it's more than enough to explain why "I sent everyone the document, so it's shared" is a risky belief.
And understanding doesn't guarantee agreement. Sales cares about the deadline; development cares about quality. Both can grasp the situation perfectly and still calmly reach different conclusions.
Veto player theory in political science holds that the number of actors whose consent is required to change the status quo, and the distance between their positions, govern how easily change happens. Applying it to a company is only an analogy. Still, looking beyond the number of stakeholders to "whose consent is actually required" strikes me as a handy lens.
Hearing everyone's opinion. Sharing the understanding that's needed. Requiring unanimity to decide. Lump those three together and, I suspect, decision-making gets heavy the moment you do.
The waiting that remains after work speeds up
Bring this back to AI adoption, and problems surface that individual processing speed can't solve.
Sales summarizes a customer request, a planner drafts the spec, a manager approves it, and an engineer builds it. Speed up any step with AI you like; if the request was misread or approval is stuck, delivery time won't shrink by the same proportion. The document may be done in 10 minutes, but if the approver's calendar doesn't open until next week, that document will be napping peacefully on a desk until next week.
So here's the hypothesis I'm working with. The shorter creation time gets, the more the time spent waiting on understanding, judgment, and coordination stands out. More output may also pile more onto reviewers. No matter how fast the hands carrying the buckets get, if the line jams somewhere, the water just pools there.
That's why staring at AI usage rates or lines of generated code doesn't get you far. Time from request to something customers can use, rework, and the quality of the result: those are the things worth measuring.
Handing the hierarchy to AI
In "From Hierarchy to Intelligence" from March 2026, Jack Dorsey and Roelof Botha lay out a plan to rebuild Block's organization. It doesn't stop at handing everyone AI. The idea is to move the information-gathering and coordination that hierarchy used to do over to a system.
The roles they describe are three: specialists who do the hands-on work, DRIs who own cross-functional outcomes, and player-coaches who do the work themselves while developing people. The authors flatly call a permanent layer of middle management unnecessary, while also writing that the transition is at an early stage. Treating it as an organizational model with proven success would be premature.
What caught my eye was less the roles than the premise that decisions and progress inside the company get recorded and the system can see them. If information lives only in certain people's heads, it's hard to move work forward without routing it through them.
NRI's "AI and the Future of Organizations" also ties hierarchy and specialization to the limits of human cognition. The view is that if AI loosens those limits, the shape of organizations may change too.
Reading the two side by side, I think that before erasing positions from the org chart, you need a state where anyone can reach on-the-ground information and the reasons behind decisions. Deleting a box from a chart is easy. Losing everything that only lived in that person's head along with it defeats the purpose.
How far AI can stand in for collaboration
Apart from corporate plans, experiments have begun to tell us some things.
Dell'Acqua and colleagues' "The Cybernetic Teammate" is a preregistered experiment with 791 P&G professionals working on real product innovation tasks. Participants were randomly assigned to four conditions: individual or two-person team, with or without AI.
The result: individuals using AI produced work of similar quality to two-person teams without AI. On top of that, AI users' proposals were less skewed toward their functional area, whether R&D-leaning or commercial-leaning.
AI might be able to take over part of the reason we brought people together. The result certainly invites that thought. But this was a one-day product proposal experiment at one specific company. It doesn't show AI taking on long-term operations, developing junior staff, or negotiating between departments at odds. The paper itself carefully separates improved idea generation from the human role of evaluation and selection.
Microsoft's 2025 report proposes a "Work Chart" that pairs people and AI around goals rather than departments. It's an interesting direction. But it's a report mixing surveys with visions of the future, not a randomized experiment proving that organizational form is superior.
Experiments, corporate initiatives, forecasts. Each needs to be read as evidence of a different strength.
The weight of an approve button
A role often assigned to humans is to check things at the end and take responsibility. But if more output arrives than anyone could possibly read, and the job turns into pressing an approve button, how much meaning is left in that check? It may strengthen the finger that presses the button, but not the judgment behind it.
Have one AI check another AI's draft, and if the source material is missing key information, both may miss it together. I want to treat the number of reviews and verification against independent evidence as two separate things.
A study of 758 BCG consultants, presented by HBS, argues that AI's capability is jagged across tasks. It's from 2023, so it doesn't reflect current models as-is. Even so, the design question it raises hasn't aged: similar work can't necessarily be delegated the same way.
Rather than adding more human checks across the board, I'd first decide what each check is meant to confirm. Calculations get recalculated. Specs get confirmed with real customers. Code gets run. For high-stakes decisions, also consult evidence separate from the explanation used to generate the output.
If you hold people responsible, you also have to hand them the authority to stop things, the material to judge with, and the time to check. Handing over responsibility alone is like giving someone an empty bucket and sending them to the fire.
So that's where I've landed, using The Mythical Man-Month as a lens on organizations after AI speeds up the work. Read for the AI era, "how much can we increase work capacity?" isn't enough of a question. It needs a companion: "what do we have to coordinate to turn that capacity into results?" 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, at which touchpoints do humans judge what, and what do they learn? And who gets to decide how much? I want to be part of an organization that can 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.