Open a manager's calendar and separate the work into two piles.
One pile moves information. It routes the question, summarizes the meeting, chases the missing input, translates the update, reconciles status, packages the exception, and formats the answer for the level above. AI is very good at that pile, and it is getting better.
The other pile commits the organization. It decides what matters, develops the person, resolves the conflict, sponsors the choice, protects the trade-off, and answers for the result. AI can inform that work. It cannot be accountable for it.
The mistake is treating the first pile as the manager.
That is the narrow claim of this essay. AI is coming first for the coordination work bundled into management, not necessarily for the manager as a whole. The "not necessarily" matters. This is not a defense of existing layers, not a prediction that managers are safe, and not a prediction that management disappears. The honest question is smaller and more useful: which managerial tasks compress, where do decision rights move, and who remains answerable for the outcome?
The management bundle
Management is often discussed as if it were a single substance. It is not. A role called manager usually contains several different kinds of work that got bundled together because, historically, moving information required a person in the middle.
Some of that bundle is routing: getting work, questions, exceptions, and decisions to the right person or system. Some of it is synthesis: turning many inputs into a coherent view for someone who cannot read everything underneath. Some of it is monitoring: maintaining visibility into status, risk, progress, and variance.
That is real work, not ceremony and not waste. They became management work because somebody had to notice what was happening, translate it into the next context, and decide whether it needed attention. Under expensive information, that was a substantive function.
But management also contains coaching: improving a person's capability, confidence, judgment, and future performance. It contains conflict work: resolving disagreement among goals, priorities, functions, and people. It contains judgment: choosing under ambiguity when the facts, values, risks, and timing do not settle the answer. It contains sponsorship: spending political capital, securing resources, and protecting a decision after the meeting ends. And it contains accountability: being answerable for the result and its consequences.
That eight-part decomposition is mine, and it is inferred. No source in the evidence base cuts management this way. The categories are chosen because they behave differently when intelligence gets cheap, not because anyone measured where the boundaries fall, and a reader is entitled to redraw them.
Those activities are not all equally exposed to cheap intelligence. Routing compresses. Synthesis often compresses, with verification cost attached. Monitoring is technically compressible and politically dangerous. Coaching can be assisted. Conflict can be prepared. Judgment can be better informed. Sponsorship can be drafted around. Accountability does not compress as information does, because it is not a flow of data. It is an assignment of consequence.
That distinction is not a compliment to managers and it is not an insult to coordination work. Both matter. The point is that they behave differently when intelligence gets cheap.
Why coordination is exposed
The strongest evidence for coordination compression comes from the work that looks least like an org chart.
Dell'Acqua and colleagues' Cybernetic Teammate study, published in Organization Science in 2026, was a pre-registered field experiment with 791 professionals at Procter & Gamble, a global consumer packaged-goods company, on new-product-development tasks. Individuals working with AI matched teams working without AI, and AI helped bridge functional silos by producing more balanced cross-functional solutions regardless of professional background.
That is not a staffing formula. It does not say half the people are unnecessary, and it does not permit headcount arithmetic. It says something narrower and more interesting: some of what organizations previously got by assembling people from different functions can now be partially replicated by a person plus a model. The model supplies candidate knowledge, translation, and cross-functional context that used to require coordination among humans.
That matters because coordination is one of the reasons layers exist. Part 1 of this series used Luis Garicano's model of hierarchy to make the point: expertise is costly, so organizations route routine problems away from scarce attention and escalate exceptions upward. I have used similar language in a different setting, writing about agent systems that "only escalate to humans on exception." The operating pattern is familiar because the economic problem is familiar. Attention is scarce, so the system filters.
AI changes the cost of that filtering. It can classify, summarize, route, draft the exception, retrieve the missing context, and turn a messy work state into something another role can consume. If a large share of a manager's day is spent moving information to the right altitude, that share is exposed.
But exposed is not the same as eliminated. The same exception still has to be owned when it arrives. The same conflict still has to be resolved when the facts are incomplete. The same person still has to tell a team why their priority lost, or tell a sponsor why the attractive answer is not defensible.
That is the unbundling. The coordination work compresses first. The management question comes after.
The fork: information or communication
The mistake in most flattening arguments is that they say "technology" when they mean one of two different things.
Bloom, Garicano, Sadun, and Van Reenen found that information technology and communication technology have opposite effects on the allocation of decision rights. Better information for the person doing the work is associated with more autonomy and wider spans. Better communication upward is associated with less autonomy and narrower spans, because the principal can monitor and direct more cheaply.
This is the load-bearing citation for the essay, and the application to generative AI is my inference rather than the paper's finding. Generative AI can be information technology for the operator: retrieval, synthesis, local decision support, and exception packaging near the work. It can also be communication technology for the principal: continuous status synthesis, performance visibility, risk reporting, and standardized escalation upward.
Those are not the same deployment. They may use the same model, the same interface, and the same vendor invoice. Organizationally, they point in different directions.
I have described the operating topology before in the three postures of AI work: "Chat is one human asking. Build is one human directing one agent. Automate is agents working with each other and pinging us when they need us." That sentence is useful here because it names who initiates, who acts, who reviews, and who is paged on exception. Those choices are structure choices wearing product language.
Which effect dominates is not a property of the model. It follows from what gets built and who it reports to.
The historical warning
The historical record makes the simple version of flattening hard to sustain.
Rajan and Wulf documented that large US firms did flatten between the mid-1980s and the late 1990s: fewer levels between the CEO and division heads, and more direct reports to the CEO. That part happened, and the series does not deny it.
The caution is what flattening meant. Guadalupe, Li, and Wulf later showed that as firms removed layers, top management teams grew through functional C-suite roles, and the pattern was consistent with more centralization around the CEO. Wulf's own synthesis carried the useful title, "The Flattened Firm: Not as Advertised." Fewer layers did not simply mean more distributed authority.
That history does not predict what AI will do. It does something more modest: it removes the permission to treat flat and decentralized as synonyms. A flatter structure can still pull decisions upward, and it can make central control easier if the middle layer's translation work is replaced by instrumentation reporting directly to the top.
This is where the surveillance risk enters.
Monitoring is a legitimate management task, and it is how leaders see variance, find risk, and intervene before small problems get expensive. But monitoring becomes a different thing when the same system that summarizes work also records, scores, compares, and routes consequences back to people.
Kellogg, Valentine, and Christin describe algorithmic control through six mechanisms: restricting, recommending, recording, rating, replacing, and rewarding. I am not claiming every AI-enabled management system does all six, or that monitoring is inherently surveillance. The risk is drift. A tool introduced to reduce coordination cost becomes a tool for continuous visibility. Continuous visibility becomes rating. Rating becomes central direction. The org chart looks flatter while managerial control expands.
That is the steelman for centralization and it deserves taking seriously. Pulling visibility upward can improve consistency, reduce variance, and give senior leaders earlier warning on material risk. In some businesses and some regulatory contexts that is not a pathology. It is the job.
So the issue is not whether centralization is wrong. It is whether centralization was chosen, named, bounded, and owned. If AI is being deployed as communication to the principal, say that. Say what autonomy is being traded for what assurance, which telemetry will not be used for individual performance management, and who can challenge the reading. The danger is not surveillance so much as surveillance by default, justified afterward as efficiency.
What the labor evidence does and does not say
The best current labor-market evidence gives neither side permission to overclaim.
Humlum and Vestergaard's March 2026 revision is titled Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. That title matters because the paper is not saying nothing happened. Their setting is Danish administrative data linked to two adoption surveys, covering roughly 25,000 responses per survey round across about 7,000 workplaces in 11 exposed occupations. There, the authors estimate precise null effects on earnings and recorded hours at both the worker and the workplace level, ruling out effects larger than 2 percent two years after the launch of ChatGPT.
They also report, in the body of the paper, that adopters in their sample saved about 3 percent of work hours. Handle that carefully: it is body-only rather than the abstract headline, and self-reported, in an early and largely pre-agentic window. The more important point here is qualitative. The work reorganized. Employers absorbed AI through task reorganization, including new tasks in content generation, AI oversight, and AI integration, and the authors frame that reallocation as suggestive rather than causal.
That is the level at which this essay should use it. Time savings appeared, earnings and recorded hours did not move detectably in that window, and some of the savings went into reorganized work. None of that says managers are safe or that future reductions cannot happen. The conversion from task savings to labor reduction is an empirical claim, not arithmetic.
Why the gains lag, and what that licenses
There is a known mechanism for the gap between task-level gains and firm-level effects, and it predates this technology by a quarter century.
Brynjolfsson and Hitt, Bresnahan with Brynjolfsson and Hitt, and Brynjolfsson with Hitt and Yang established across the information-technology era that returns depend on complementary organizational change rather than the technology alone, and that the complementary investment is a large share of the total asset. The market-value estimates there are specification-dependent, so take them as an order of magnitude rather than a coefficient: the organizational asset is worth several times the computer capital, not some decimal multiple of it. Brynjolfsson, Rock and Syverson later named the timing. Their productivity J-curve describes measured productivity understating true growth while the intangible complements are being built, and overstating it afterward, because the investment is expensed while the asset it creates is invisible to the accounts.
That evidence does two jobs and this essay needs both. It defends generative AI against premature dismissal, because a null two years in is roughly what the J-curve predicts for a general-purpose technology whose complements have not been built. And it refuses premature headcount enthusiasm, because the return arrives through the reorganization rather than instead of it.
That matters because managerial work is a bundle. If routing and synthesis compress, the recovered capacity may become fewer layers, wider spans, more coaching, more verification, more risk review, or more explicit ownership. The evidence does not pick. Three choices do: what gets deployed and where, who holds decision rights over the redesign, and which controls the new shape runs on. Those sit with leadership, under Part 3's constraint: the people whose work changes hold concurrence over monitoring, over individual-performance use of telemetry, and over material job-content changes, and stronger legal, bargaining, works-council, professional, or policy rights govern.
Three plausible shapes
Three shapes from the same capability
The same capability supports at least three structures, and nothing in the technology chooses between them.
The same AI capability can support three organizational shapes, and the technology does not choose among them. Three equal panels: fewer layers with authority more centralized, the same number of layers with wider spans of control, and fewer layers with authority pushed down to the work. Decision markers show where the decision rights sit in each. The three are presented as peers in arbitrary order, with no ranking and no default among them.
Fewer layers, more centralized
Levels come out, and the decision marker sits on the single node at the top.
Deployment choice: AI used as a communication technology — visibility up, authority up.
Flattening of this kind is a historically observed pattern (Rajan & Wulf 2006; Guadalupe, Li & Wulf 2014 — established).
Same layers, wider spans
The number of levels is unchanged, and each remaining manager carries more of the work directly.
Deployment choice: coordination load absorbed, so spans widen instead of levels coming out.
Fewer layers, authority pushed down
Levels come out, and a decision marker sits under every node at the work, with lateral links across them.
Deployment choice: AI used as an information technology for the operator — verification capability and decision rights at the work.
Decision marker: where the decision rights sit. The three shapes are peers and the order is arbitrary.
Bloom, Garicano, Sadun & Van Reenen (2014) found that information technology and communication technology move decision rights in opposite directions. Generative AI is both. Which effect dominates is a deployment choice (established finding; the application to generative AI is inferred).
The same AI capability can support at least three organizational shapes.
The first is the compressed pyramid. Coordination work is automated, layers reduce, and visibility moves upward, so decisions once filtered through intermediate layers converge toward the top. Efficient, possibly. It can also concentrate decision load and make surveillance feel like operating discipline.
The second is the federated lattice. AI is deployed as information to the operator, so retrieval, synthesis, local decision support, and exception packaging happen closer to the work. Handoffs reduce because people at the edge have more context and better tools for making admissible decisions. That requires real investment in local judgment, verification, and controls, not just empowerment language.
The third is the assurance-heavy structure. Coordination compresses, but the recovered attention is reinvested in verification, accountability, risk ownership, and control evidence. Spans may widen and layers may change, but the distinctive feature is not flatness. It is explicit assurance around consequential decisions. Badly designed, that becomes bureaucracy. It can also be right where the cost of being wrong rose faster than the cost of routing fell.
Those three shapes are peers, not a maturity model and not a forecast. They are a forcing device for a leadership conversation otherwise too easy to hide inside words like efficiency, empowerment, and simplification.
The hard objection
A practical objection carries more force: cost pressure may choose for you.
A board, CEO, or CFO does not need the theory to be deterministic. They need a margin target, visible task automation, and a plausible story that coordination effort has fallen. Once those conditions exist, "the technology does not choose the structure" is formally true and practically irrelevant, and the technology becomes the justification for a structure change already wanted.
Against the naive version of this essay, that lands. Cost pressure, investor expectations, leadership fashion, and vendor narratives can drive a structure change even when the evidence does not compel it. Technology may not be the dominant force in the room.
I am not denying it. I am insisting that if the structure change is a choice, it is owned. Treating delayering as something AI simply caused gives leaders a way to avoid accountability for the design; naming it as a choice does the opposite. It asks who chose the compression, where decision rights moved, what happened to coaching and conflict work, what assurance replaced the removed handoff, and what evidence would prove the choice worked.
There is an opposite failure mode: "flattening is a choice" cannot become permission to do nothing. A company that preserves every layer while automating the coordination load ends up with the same hierarchy, wider spans, quieter central control, and no map of where accountability lives. Stasis is also a design choice after the cost structure moves.
A quantity-free example
Consider a planning role.
The coordination portion gathers inputs from contributing teams, reconciles conflicts in the material, reformats updates for the level above, chases what is missing, and turns scattered status into a coherent view. It matters, and it gets easier when AI can retrieve, summarize, classify, and draft.
The commitment portion decides which conflicting asks receive support, explains the trade-off to the teams that did not get what they wanted, protects the decision when pressure returns, and answers when the plan misses. AI can prepare the facts and draft the explanation. It cannot spend the political capital, repair the relationship, or be answerable.
If the coordination work compresses, the role should change. That does not mean the commitment work disappeared, only that somebody has to decide where judgment, sponsorship, conflict resolution, and accountability now live.
This example is illustrative and sanitized. It names no employer, system, person, proprietary process, metric, interval, volume, proportion, threshold, staffing outcome, or identifiable event.
The Monday test
Run the test before approving a structure change justified by AI.
First, take one management role and sort the work by task, not by meeting title, using the eight-task decomposition: routing, synthesis, monitoring, coaching, conflict, judgment, sponsorship, and accountability. Mark which tasks primarily move information between people and which commit the organization or develop its capacity to make better commitments.
Second, write down which of the three shapes the proposed AI deployment is actually selecting: compressed pyramid, federated lattice, or assurance-heavy structure. If the answer is "all of them," the organization has not made a design decision. If the answer is not written down, the budget is making it.
Third, for every coordination task removed or automated, name where its attached judgment, sponsorship, conflict resolution, and accountability now live. A blank answer is not a savings opportunity; it is the part of the role that moved without an owner.
The output is a task decomposition, a named structural choice, and an accountability map. It is not a headcount plan.
What comes next
The point is not to preserve management as it exists. Some coordination work should be compressed aggressively, some layers should change, and some managerial calendars carry a large amount of work the old information environment forced onto humans and the new one no longer requires.
But coordination structure does not vanish just because coordination gets cheaper. In The Petri Dish, I wrote that "Once norms stabilize, groups create organs: review boards, registries, trusted lists. These reduce transaction costs and increase reliability." That was about agents, not human org charts, but the resonance matters. Coordination structures emerge because reliability has a cost. Cheap coordination changes their form, not the need to decide, sponsor, verify, and answer for consequences.
Once coordination work compresses, the approval becomes the place organizations try to preserve accountability. That is understandable. It is also why the approval has to be examined next, because an approval is not the same thing as a control.
Sources
- Bloom, Nicholas, Luis Garicano, Raffaella Sadun, and John Van Reenen. "The Distinct Effects of Information Technology and Communication Technology on Firm Organization." Management Science 60, no. 12 (2014): 2859–2885. DOI.
- Rajan, Raghuram G., and Julie Wulf. "The Flattening Firm: Evidence from Panel Data on the Changing Nature of Corporate Hierarchies." Review of Economics and Statistics 88, no. 4 (2006): 759–773. Publisher.
- Guadalupe, Maria, Hongyi Li, and Julie Wulf. "Who Lives in the C-Suite? Organizational Structure and the Division of Labor in Top Management." Management Science 60, no. 4 (2014): 824–844. DOI.
- Wulf, Julie. "The Flattened Firm: Not as Advertised." California Management Review 55, no. 1 (2012): 5–23. DOI.
- Dell'Acqua, Fabrizio, et al. "The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork." Organization Science 37, no. 4 (2026): 1217–1242. DOI.
- Brynjolfsson, Erik, and Lorin M. Hitt. "Beyond Computation: Information Technology, Organizational Transformation and Business Performance." Journal of Economic Perspectives 14, no. 4 (Fall 2000): 23–48. Publisher.
- Bresnahan, Timothy F., Erik Brynjolfsson, and Lorin M. Hitt. "Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence." Quarterly Journal of Economics 117, no. 1 (February 2002): 339–376. Publisher.
- Brynjolfsson, Erik, Lorin M. Hitt, and Shinkyu Yang. "Intangible Assets: Computers and Organizational Capital." Brookings Papers on Economic Activity 2002, no. 1 (2002): 137–198. Publisher. Market-value estimates are specification-dependent; cited as order of magnitude only.
- Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." American Economic Journal: Macroeconomics 13, no. 1 (January 2021): 333–372. Publisher.
- Humlum, Anders, and Emilie Vestergaard. "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI." NBER Working Paper 33777, May 2025, revised March 2026. DOI.
- Kellogg, Katherine C., Melissa A. Valentine, and Angèle Christin. "Algorithms at Work: The New Contested Terrain of Control." Academy of Management Annals 14, no. 1 (2020): 366–410. DOI.
- Garicano, Luis. "Hierarchies and the Organization of Knowledge in Production." Journal of Political Economy 108, no. 5 (2000): 874–904. DOI.
Operate. Publish. Teach.