The Org Chart Was Built for Expensive Information

Your org chart is a price list.

Every layer on it was priced. Every span, every approval gate, every step where one person's work has to be packaged before another person can use it. Each of those was a judgment that information was expensive enough to be worth building a structure around. Somebody, at some point, decided that getting a reconciled answer out of the business required a person whose job was to go get it, and that checking the answer required a second person above them who had seen enough answers to know when one looked wrong.

Those were good decisions at the prices then available. The prices moved. The price list did not.

That is the argument, and here it is in one sentence. Every layer, span, and approval gate in your organization is a priced response to expensive information, and the price changed asymmetrically, which makes the structure mispriced rather than obsolete. Obsolete licenses removal. Mispriced does not. Mispriced structures have to be re-derived, and re-derivation is a different piece of work than cutting, with a different output.

A note on evidence before I start. This series reaches back further than my usual writing, into organizational economics from 1937, 1971, 1974 and 2000, and cites peer-reviewed field experiments rather than benchmarks and vendor reports. That is deliberate. The question is older than the technology.

What I am extending, and what I am not re-arguing

I wrote Superworkers, Not Replacements in late 2025, and its claim was this: "A superworker isn't someone who has been replaced by AI. It's someone whose capabilities have been dramatically expanded by it." I still think that is right, and I am not going to re-argue it here.

That post answered the question everyone was asking. Will AI replace people? No, it amplifies them. It also left a question standing that I did not answer there. Given amplification, what happens to the structure that coordinated those people? Amplified individuals still get routed, still escalate, still have their work checked, and still need someone accountable for the result. This series is about what happens to those functions.

The structure is a price list

Start with why a firm exists at all, because that is where the pricing logic comes from.

Coase's answer, in 1937, was that a firm grows until organizing one more transaction inside it costs the same as buying that transaction on the open market. The boundary is a price comparison. Two things follow, and the second one is the one that gets skipped. First, structure is not a fact about your company. It is a settlement between costs, made at a moment, by people responding to what things cost then. Second, the frame is symmetric. It predicts that the boundary moves when relative prices move, and it says nothing at all about which way. Falling internal coordination cost expands a firm as readily as falling market cost contracts it. Anyone using Coase to argue that cheaper coordination shrinks hierarchy has added a premise and not told you which one.

I have used a transaction-cost argument here before, applied to agents rather than to people, in The Petri Dish. The frame is not new to this site. The subject is.

Simon's point, in 1971, was that when information becomes abundant the thing that becomes scarce is the attention that information consumes. The design problem is therefore not delivering more information, it is filtering it. He wrote that before the organizations reading it had email, which is worth sitting with for a second. It means the constraint every leader now describes as an AI-era problem was correctly specified fifty-five years ago, and the structures we inherited were built by people who understood it.

Galbraith's proposition, in 1974, was that the more uncertain the task, the more information has to be processed among the people executing it, and that an organization has only two families of response. Reduce the need to process information, or increase the capacity to process it. There is no third family. Hold onto the two-ness of that. It does more work later than it appears to do here.

Garicano's model, in 2000, formalizes what hierarchy is for. Acquiring knowledge is costly, so the organization arranges itself so that routine problems are solved by people who cost less, and only the exceptions climb to the people who cost more. Layers are a filter on expensive attention. That is why the pyramid narrows: not because someone likes pyramids, but because expertise is scarce and you ration access to it by making problems earn their way upward.

I want to keep that claim narrow, because it is narrow. It is a formal model of knowledge acquisition and problem-passing, not a theory of everything a manager does. I will come back to that limit, because the strongest objection to this essay is built on it.

What actually got cheaper

Here is my decomposition, labeled before I use it: the four-cost frame is mine. It is inferred, an organizing device rather than a finding, and no source proposes it. Each of the four is separately supported at the task level. The frame that binds them is not.

Cheap machine intelligence lowered four costs.

The cost of knowing. Retrieval, synthesis, summarization, first-draft analysis. What used to require finding the person who knew, or the document that said, now requires a prompt.

The cost of asking. Most questions inside an organization are not asked, because asking is socially and temporally expensive. You have to find the right person, wait for them, and spend a little relational capital to do it.

The cost of routing. Triage, classification, matching work to whoever or whatever should handle it. This is the one that should get a leader's attention, and I will come back to why.

The cost of acting. The mechanical residue of a decision. Generating the document, filing the record, updating the system, writing the code.

The two cost stacks

The costs that fell are the costs hierarchy was built to manage. The costs that rose are not.

The four costs below the line fell, the four above it rose, and the organizational layers drawn on the line are unchanged. Eight labelled elements are arranged around a single horizontal baseline. Below the line, four downward arrows: knowing, asking, routing, acting. Above the line, four upward arrows: verification, legitimacy, judgment, accountability. On the line, three identical neutral rules stand for layers, drawn unmoved. No quantities are shown.

Costs that fell

• Knowingretrieval and synthesis
• Askingthe cost of a question
• Routingtriage and escalation
• Actingthe mechanical residue of a decision

the structure between them

Costs that rose

• Verificationconfirming output is correct
• Legitimacyacceptance by those affected
• Judgmentknowing which side of the frontier you are on
• Accountabilitysomeone owns the outcome

The evidence that knowing and acting fell is good, and it is task-level. Brynjolfsson, Li and Raymond, in the Quarterly Journal of Economics in 2025, studied 5,172 customer-support agents and found about 15 percent more issues resolved per hour with access to a generative assistant, with the largest gains going to the less experienced and lower-skilled workers. Read the measure, not the headline. It is issues resolved per hour, in one firm, in one occupation, in a domain with unusually fast machine-readable outcome signals. That is the friendly case, which is exactly why the number is clean.

Noy and Zhang, in Science in 2023, ran 453 college-educated professionals through mid-level professional writing tasks and reported about 40 percent less time taken and about 18 percent higher quality as scored by blinded graders. Those are short standalone tasks, not organizational work, and the authors themselves caution against reading task-level results into general-equilibrium conclusions. I am honoring that caution, and it costs the argument nothing, because the claim here is only that the two costs fell.

Asking and routing are harder to isolate, and the best available evidence comes at them sideways. Dell'Acqua and colleagues, in Organization Science in 2026, ran a pre-registered field experiment with 791 professionals at one consumer packaged-goods company on new-product-development tasks. Individuals working with AI produced work comparable to two-person teams working without it, and AI access reduced functional siloing, with participants producing more balanced cross-functional solutions regardless of their own background. The silo-breaking half is the durable contribution. AI supplied cross-functional knowledge that previously required assembling people from different functions, which is a coordination substitute.

One pattern I have noticed and cannot prove, offered as a hypothesis rather than evidence and labeled personal observation: much of what gets asked of a model was never asked of a person. It was not asked at all. The friction of asking suppressed low-stakes questions, and removing the friction surfaces demand rather than redirecting it. If that is right, the cost of asking did not fall so much as a market for questions appeared where there had not been one. I would like to see that measured. I have not found it measured.

The pivot

Now put both halves in the same sentence, because this is where the essay turns.

Garicano's hierarchy exists to route problems and ration expensive attention. Simon's design problem is filtering. Galbraith's second family is added information-processing capacity. Coase's boundary is a price comparison on organizing transactions. Those are routing, knowing, asking, and acting, described in the language of organizational economics decades before the technology showed up.

So the four costs that fell are the same four costs the theory says hierarchy exists to manage. That is inferred, and I want it labeled clearly, because it is a join I am making between an established mechanism and my own decomposition. Nobody has measured the mapping. The reason I think it holds is the tightness of the fit: Garicano's model is about routing and expert attention, and routing and expert attention are precisely what got cheap. The reason to state the label out loud is that a tight fit is not a measurement, and I would rather say that than have a reader find it.

If the mapping holds, the structure is mispriced against its own design assumptions. Note what that conclusion is about. It is a conclusion about prices, and only about prices. Nothing above licenses a claim about how many layers you should have, how wide a span should be, or how many of anything belongs anywhere.

What does not follow

The obvious next step is to conclude that the structure should be flattened. It does not follow, and declining it is the real work of this essay.

Cheap intelligence is an enormous increase in information-processing capacity. It does very little to reduce the need to process information. Those are Galbraith's two families, and only one of them is being addressed, which matters because his own argument is that capacity increases hit diminishing returns faster than need reductions do. Adding capacity to a structure whose need is unchanged is the weaker of the two available moves. It is also the one that requires no redesign, and my read is that this is why it is the common one.

The second reason is sharper, and it is empirical. Bloom, Garicano, Sadun and Van Reenen found, across firms in several countries, that information technology and communication technology push decision rights in opposite directions. 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 ones. Both are technology. They do not point the same way.

That distinction is worth keeping, because the rest of this series runs on it. Information cost is the cost of knowing something at the point where the work is done. Communication cost is the cost of moving what is known between people, particularly upward. Generative systems lower both at once. So any sentence that predicts a structural consequence from "AI" or "technology," without naming which of the two costs it means, is carrying a hidden premise about which effect dominates. That premise is a deployment decision, not a property of the model. The three postures I have written about before are a reasonable operational vocabulary for it: what you build determines which cost you actually attacked.

At least three structures are consistent with the same capability. Fewer layers with connections converging upward, which is what you get deploying AI as a reporting and monitoring technology. The same layers with wider spans, coordination automated and decision rights left exactly where they were. Fewer layers with authority pushed toward the work, which is what you get deploying AI as an information technology for the operator. Those are peers, listed in an arbitrary order, and nothing in the technology chooses between them. Part 4 takes up which conditions favor which. Part 1's job is to establish that a choice is being made and that somebody should make it on purpose.

I want to be precise about the refusal, because it is easy to overstate in the other direction. I am not claiming your structure will not change. I am not claiming middle layers are safe. I am claiming that flattening is not compelled by the cost change, that predicting stasis would be the same error pointed the other way, and that the technology underdetermines the answer. Which also means the answer is yours, and you will be accountable for it either way.

Re-deriving a structure means pricing it at today's costs, and that requires knowing which costs went up. Four did, and I will name them here and develop none of them: verification, legitimacy, judgment, and accountability. Each one gets an essay, because each behaves differently from the four that fell and none of them compresses the way coordination does.

There is a related split I will only gesture at here. Part of what a manager does is move information, and part of it commits the organization. Those two halves behave very differently under cheap intelligence, and Part 4 takes them apart properly.

The strongest objection

Here is the best version of the argument against everything above, and I am going to give it in its own words rather than a convenient version.

"The org chart was never really about information cost. It is about control, career ladders, legal accountability, span-of-attention limits, and status. Managers exist to hire, fire, motivate, absorb political risk, and give people somewhere to be promoted to, none of which is an information-processing function. So making information cheap does not touch the load-bearing reason the structure exists."

That objection is substantially correct, and I will not hedge it. The information-processing account of hierarchy is one account among several, not a complete theory, and anyone who has sat through a reorganization knows that career structure and political cover do at least as much work as routing does.

Now notice where the concession lands. If a large share of the org chart is not an information-processing artifact, then cheap information cannot dissolve it. The objection arrives at this series' conclusion from its own premises. The interesting question it opens is how you separate the information-processing half of a role from the rest, and that is Part 4's subject, where it gets a name.

One artifact, and what to do Monday

An illustrative example, and it is illustrative in the strict sense: a recurring pattern in professional practice, not an outcome at any particular organization.

A large enterprise produces a recurring forecast pack. It moves through three levels of aggregation. Each level reformats the level below it into the shape the level above prefers, each adds a reconciliation step, and at the top it takes two signatures before it reaches the executive committee.

That chain was engineered when getting a reconciled number out of the contributing systems was slow, expensive, expert work, and when getting the contributing teams to agree on the number was slower still. Retrieval and reconciliation are now fast and cheap. The sign-off chain takes exactly as long as it always did.

Nothing broke. Nobody is doing anything wrong. The structure simply stopped matching the prices it was built against, and nobody re-derived it, because nobody had ever been asked to say out loud what each handoff was buying down. Hammer's 1990 diagnosis still lands here: accelerating an unredesigned process preserves its waste.

So, the test. One person can run it, with no consultant, no survey, and no budget line.

Pick one recurring artifact your organization produces. A report, a review pack, a standard approval. Count the handoffs.

Then, for each handoff, name which of the four costs it was created to buy down, and what that cost is today.

Any handoff whose original cost you cannot name is the first thing to redesign. Not delete. Redesign, and the difference between those two words is the rest of this series. What you end up with is a list of steps with a price beside each one, and that is something no org chart shows you: not how the organization is arranged, but what it was charging for.

What comes next

Everything measured above was measured at the level of the individual. That is what the productivity evidence covers, and it is what the amplification argument correctly describes. It says nothing about what happens after the amplified individual finishes their work and hands it to somebody else.

I wrote in 2025: "The replacement narrative is a failure of imagination." I still think so. But that sentence was right about the person and silent about the structure, and the structural question is the harder one. An organization full of faster individuals is not a faster organization if the next step in the flow is unchanged, and that next step is very often a human accepting responsibility for work they did not produce.

Part 2 is about that step, and about why it did not get cheaper.

Sources

• Coase, R. H. "The Nature of the Firm." Economica 4, no. 16 (November 1937): 386–405. https://onlinelibrary.wiley.com/doi/10.1111/j.1468-0335.1937.tb00002.x. Paraphrased, not quoted.
• Simon, Herbert A. "Designing Organizations for an Information-Rich World." In Computers, Communications, and the Public Interest, edited by Martin Greenberger. Baltimore: Johns Hopkins Press, 1971. Paraphrased, not quoted, and the chapter is cited as a whole.
• Galbraith, Jay R. "Organization Design: An Information Processing View." Interfaces 4, no. 3 (May 1974): 28–36. https://pubsonline.informs.org/doi/10.1287/inte.4.3.28. Paraphrased, not quoted.
• Garicano, Luis. "Hierarchies and the Organization of Knowledge in Production." Journal of Political Economy 108, no. 5 (October 2000): 874–904. https://www.journals.uchicago.edu/doi/10.1086/317671. Survey treatment: Garicano and Rossi-Hansberg, "Knowledge-Based Hierarchies," Annual Review of Economics 7 (2015): 1–30. https://www.annualreviews.org/doi/10.1146/annurev-economics-080614-115748
• Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." Quarterly Journal of Economics 140, no. 2 (May 2025): 889–942. https://academic.oup.com/qje/article/140/2/889/7990658. Measure: issues resolved per hour, about 15 percent, 5,172 customer-support agents (established).
• Noy, Shakked, and Whitney Zhang. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science 381, no. 6654 (July 13, 2023): 187–192. https://www.science.org/doi/10.1126/science.adh2586. Measures: time taken on occupation-specific writing tasks, about 40 percent lower; blinded-graded output quality, about 18 percent higher; 453 college-educated professionals (established).
• Dell'Acqua, Fabrizio, Charles Ayoubi, Hila Lifshitz, Raffaella Sadun, Ethan Mollick, Lilach Mollick, Yi Han, Jeff Goldman, Hari Nair, Stew Taub, and Karim R. Lakhani. "The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork." Organization Science 37, no. 4 (July–August 2026): 1217–1242. https://doi.org/10.1287/orsc.2025.20702. 791 professionals, one consumer packaged-goods company, new-product-development tasks; published abstract's qualitative findings only (established).
• 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 (December 2014): 2859–2885. https://pubsonline.informs.org/doi/10.1287/mnsc.2014.2013 (established).
• Hammer, Michael. "Reengineering Work: Don't Automate, Obliterate." Harvard Business Review 68, no. 4 (July–August 1990): 104–112. https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate. Cited for the diagnosis only.

Operate. Publish. Teach.