What happens to work whose purpose was moving information around?

Two different things, and an organisation chart cannot separate them. Some of this work existed because carrying information from where it was to where it was needed cost money, and when the cost falls the work goes with it. The rest of it acquired judgement along the way: the person compiling the monthly pack learned to notice when a number was wrong, and nobody wrote that down because it was never the job description. Removing the first kind is a saving. Removing the second is a saving on the invoice and a loss everywhere else. Detection takes about two quarters. The distinction is worth making before the restructure rather than during it.

Luis Garicano answered this in 2000 and the answer has not been bettered#

The reason organisations have layers at all is not that somebody liked hierarchies. Garicano's model, published in the Journal of Political Economy, treats a hierarchy as an economising device for matching problems to knowledge.

In such a structure, knowledge of solutions to the most common or easiest problems is located in the production floor, whereas knowledge about more exceptional or harder problems is located in higher layers of the hierarchy. Production workers who confront problems they cannot solve refer them to the next layer of the organization, formed by specialist problem solvers. Problems are then passed on until someone can solve them or until the conditional probability of finding the solution is too low to justify continuing the search.

The trade-off he identifies is the one that matters here: "By adding layers of problem solvers, the organization increases the utilization rate of knowledge, thus economizing on knowledge acquisition, at the cost of increasing the communication required." A layer exists because knowledge is expensive to acquire and cheap to consult. Every job whose content is passing a problem upward, packaging it for the person above, and passing an answer back down, is that trade-off made flesh.

Garicano set the question this page is asking, twenty-six years ago, in one line: "will cheaper communication technology make an organization taller or shorter?" The reason his framing survives is that it names the two costs separately. Generative tools reduce the cost of acquiring knowledge, which points one way. They also reduce the cost of communicating, which points the other.

The two technologies pull authority in opposite directions#

Bloom, Garicano, Sadun and Van Reenen tested the distinction on manufacturing plants and found it holds. Their published abstract states the result plainly: "information technology is a decentralizing force, whereas communication technology is a centralizing force". In the data, "better information technologies (enterprise resource planning (ERP) for plant managers and computer-assisted design/computer-assisted manufacturing for production workers) are associated with more autonomy and a wider span of control, whereas technologies that improve communication (like data intranets) decrease autonomy for workers and plant managers".

The mechanism, in the working paper's words: "technologies that reduce information costs enable agents to acquire more knowledge and 'empower' lower level agents. Conversely, technologies reducing communication costs substitute agent's knowledge for directions from their managers, and lead to centralization." Give somebody a tool that lets them answer their own question and their authority grows. Give them a tool that lets a manager answer it for them faster and it shrinks.

Generative AI is both tools in one interface. A model that lets an analyst resolve a question without asking anyone is decentralising. The same model, used to give an executive a same-day answer that used to take a team a fortnight to prepare, is centralising. Which effect dominates is a design decision somebody in the organisation is making, usually without knowing that they are making it. That is drift rather than design, expressed as an organisation chart.

One point of care with this study. The sample description that circulates with it, around a thousand firms across the US and seven European countries, is stated in the 2009 working paper. The published abstract says only "a new data set of American and European manufacturing firms", and the typeset article could not be opened for this page. The finding is quoted from the published abstract; the sample is attributed to the working paper.

Resume data now shows firms flattening after AI adoption, and says so cautiously#

Ewens and Giroud built a measure of corporate hierarchy from "online resumes of 7 million employees" across "over 3,100 U.S. public firms", using a network estimation technique to identify layers. Firms average ten layers and a pyramidal structure. On the question here, their abstract records: "companies flattened their hierarchies following the adoption of artificial intelligence (AI) technologies".

They connect it directly to Garicano, and the connection is the argument of this page compressed into one paragraph: "Artificial intelligence (AI) reduces the cost of knowledge acquisition and information processing. From a theoretical perspective, we would expect hierarchies to flatten following the adoption of AI. For example, in the models of knowledge hierarchy, lowering the cost of knowledge acquisition allows workers to solve a wider range of problems, which reduces the demand for problem solvers and hence the need for hierarchical layers."

Then they publish the caveat that most citations of this paper will drop: "Although the tests are under-powered, we find that the point estimates are significant at the 10% level regardless of the metric of AI adoption." A ten per cent significance level on an under-powered test is a signal worth having and not a fact worth building a restructure on. AI adoption in their design is measured by AI job postings, following Babina and colleagues, which is a proxy for hiring intent rather than for use.

Babina, Fedyk, He and Hodson looked at the composition rather than the shape and report that "AI investments are associated with a flattening of the firms' hierarchical structure, with significant increases in the share of workers at the junior level and decreases in shares of workers in middle-management and senior roles". Their abstract gives direction and significance without magnitude, so no percentage appears here. Note what that composition change means for the estate's central concern: a workforce with proportionally more juniors and fewer of the middle layer that used to develop them is the missing rungs problem arriving through the operating model.

Most of the flattening already under way has nothing to do with AI#

A page arguing that AI removes coordination layers has an obligation to report the evidence that the layers were already going. Gusto, analysing continuous payroll and reporting-relationship data for 8,500 businesses of between two and 500 employees from January 2019 to September 2024, found that "in 2019 people managers were directly responsible for about 3 direct reports, but by 2024 they were directly responsible for about 6 direct reports", that "the share of workers who are in a people manager role decreased by 34%" over the five years, and that "across all small- and medium-sized businesses 14% of managerial roles were cut".

Two things about that source. It is a payroll vendor publishing analysis of its own customer base, which is original data rather than peer-reviewed data, and it covers small and medium businesses only. And Gusto attributes the change to labour costs, inflation and interest rates. The report does not make an AI claim, and it should not be made to carry one.

The nationally representative counterweight is firmer still. The US Census Bureau's 2026 AI supplement found that among firms using AI, "most users (66%) rely on AI solely to augment tasks, while AI-related employment decreases are rare, occurring in only 2% of firms". The same paper found that functional breadth and operational investment are positively associated with employment decreases, "whereas worker-task integration shows no significant link to headcount reduction once functional integration and operational investment are taken into account". Individual workers using AI is not what shrinks headcount. Redesigning a business function around it is.

Autor's inversion: the information was never the valuable part#

The most useful reframing of this question comes from David Autor, who points out that the previous round of this argument was settled and settled against the optimists.

While the utopian vision of the current Information Age was that computerization would flatten economic hierarchies by democratizing information, the opposite has occurred. Information, it turns out, is merely an input into a more consequential economic function, decision-making, which is the province of elite experts.

He names the roles that went: "Away from the factory floor, telephone operators, typists, bookkeepers and inventory clerks, served as information conduits, the information technology of their era." And he states the consequence for the people above them: "the advent of pre-AI computing made the expert judgment of professional decision-makers more consequential and more valuable by speeding the task of acquiring and organizing information. Simultaneously, computerization devalued and displaced the procedural expertise that was the stock-in-trade of many middle-skill workers."

That history is the reason to be careful with the flattening story. The last time information handling got cheap, the organisation did not become flatter. It concentrated decision rights upward and removed the roles that had been carrying information between the layers. Autor's argument is that AI could break that pattern by extending decision-making capability outward, and he is careful about the status of the claim: "My thesis is not a forecast but a claim about what is attainable." His warning is the one that belongs in an operating model review: "The risk is the devaluation of expertise."

What the reporting job was actually doing#

The interpretation this research adds is about what gets lost when a function that looked administrative is removed. Three things travel inside information-moving work and appear on no measure of its output.

It forced a decision to be stated. A monthly pack, a project status report and a planning cycle all have the same underlying property: somebody has to commit a position to a page, on a date, with their name on it. The document was the artefact; the commitment was the function. Generate the pack automatically and the commitment quietly becomes optional, because nobody is now required to have a view before the meeting.

It produced a shared version. The value of a single reporting line is that everyone argues about one set of numbers. Personalised, on-demand synthesis gives each executive a different account of the same quarter, each internally coherent. The disagreements that used to surface in the room now surface as two people who believe they already agree.

It detected the anomaly. Somebody who has compiled the same report forty times can see that a figure is wrong before they can say why. That is tacit knowledge, built by repetition of a task that looked like clerical work. In "Invisible Work" (2025) I called it the checking that keeps an organisation upright while appearing on no measure of output. It is also the first thing to go, because the repetitions that built it were the automatable part.

There is direct evidence that information flow degrades when the channels change, even without any job being removed. Yang and colleagues used telemetry on 61,182 US Microsoft employees over the first six months of 2020, treating firm-wide remote work as a natural experiment against workers who were already remote. They found "firm-wide remote work caused the collaboration network of workers to become more static and siloed, with fewer bridges between disparate parts", with a decrease in synchronous and an increase in asynchronous communication, and concluded that together "these effects may make it harder for employees to acquire and share new information across the network". A change to how information moves reorganised who knew what, without any deliberate redesign at all.

Four questions that separate the conduit from the judgement#

This is a structure for a review rather than a validated instrument. Ask it of a function before removing it.

• If this artefact stopped being produced, who would have to form a view instead? If the answer is nobody, the artefact was a conduit. If the answer is a named person who currently receives the view pre-formed, the function was carrying a decision.
• When this function last found an error, how did it find it? A function that regularly catches things is doing detection work, and detection built on repetition disappears with the repetition. A function that has never caught anything is either not looking or not needed, and the two are worth telling apart.
• Does anyone downstream act differently because of this, or only know more? Reports that change no behaviour are pure information movement and are the cheapest thing an organisation owns to stop producing, whether or not AI exists. The subtractive question is usually more valuable than the automation question and almost nobody asks it.
• Where would the people who currently do this have gone next? Coordination roles are how a great many people learn how an organisation actually works before they run part of it. Removing the layer removes the route, and the cost appears in the succession plan four years later rather than in this year's operating budget. That is the argument of synthetic seniority applied to structure rather than to individuals.

Where this sits in my own argument#

"Knowledge Is No Longer Power" (2025) put the position that Bacon's maxim has broken: knowledge is now instantly and universally available, so advantage comes from judgement about which questions are worth asking rather than from holding the answers. An organisation designed around the scarcity of information is an organisation designed around a condition that no longer holds, and most of them still are. "Rules Before Tools" (2025) made the operational half of the case: fix the processes, the layers and the decision rights first and let the tools slot into them, because chasing model releases substitutes for the redesign.

"This Is Zombie Work" (2025) named the outcome I expect where the redesign does not happen. The layer survives, its title survives, and the judgement is taken out of it, leaving people required to be present and no longer required to decide. That is worse than removing the layer, because it costs the same and produces neither the capability nor the saving.

Attribution note. The knowledge hierarchy is Garicano's. The information and communication technology distinction is Bloom, Garicano, Sadun and Van Reenen's. The devaluation of expertise argument is Autor's. Missing rungs, synthetic seniority, zombie work and drift versus design are mine. Span of control and tacit knowledge are established management and philosophy vocabulary and belong to nobody.

The flattening result is under-powered and its authors say so#

It does not claim AI is causing organisations to flatten. The strongest available finding, from Ewens and Giroud, is a directional result the authors themselves describe as under-powered and significant at the 10 per cent level, using AI job postings as the adoption proxy. The clearest measured flattening, in the Gusto data, is attributed by its own authors to labour costs and interest rates.

It does not claim a magnitude for the middle-management effect. Babina and colleagues report direction and significance in the abstract read for this page; no percentage was read, so none is stated. Two figures that circulate widely on this question, from Live Data Technologies on middle managers as a share of layoffs, could not be traced to any primary publication and are absent for that reason.

It does not claim that any specific function disappears. Nothing in the evidence base names planning, PMO, internal reporting or communications as categories, because no study has measured them as categories. The four questions above are a way of asking about a particular function in a particular organisation, and the answer will differ between two firms with identical charts.

It does not use the vendor figure most often quoted on this subject. The widely repeated claim that knowledge workers spend around 60 per cent of the day on "work about work" is published by a software company that sells against that category, with no methodology stated and with the figure varying between 58 and 66 per cent across its own pages. The measured collaboration quantities above are used instead.

Key sources

• Garicano, L. (2000). Hierarchies and the Organization of Knowledge in Production. Journal of Political Economy, 108(5), 874-904.
• Bloom, N., Garicano, L., Sadun, R. and Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization. Management Science, 60(12), 2859-2885. Sample described in the earlier version, NBER Working Paper 14975, May 2009.
• Ewens, M. and Giroud, X. (2025). Corporate Hierarchy. NBER Working Paper 34162. Abstract.
• Babina, T., Fedyk, A., He, A. and Hodson, J. (2025). Firm Investments in Artificial Intelligence Technologies and Changes in Workforce Composition. In Technology, Productivity, and Economic Growth, NBER Studies in Income and Wealth 83, University of Chicago Press.
• Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs. NBER Working Paper 32140.
• Yang, L., Holtz, D., Jaffe, S., Suri, S., Sinha, S., Weston, J., Joyce, C., Shah, N., Sherman, K., Hecht, B. and Teevan, J. (2022). The effects of remote work on collaboration among information workers. Nature Human Behaviour, 6, 43-54.
• Bonney, K., Breaux, C., Dinlersoz, E., Foster, L., Haltiwanger, J. and Pande, A. (2026). The Microstructure of AI Diffusion. US Census Bureau, CES Working Paper 26-25.
• Terrazas, A. (2025). The Manager Mass Exodus: How SMBs Are Flattening the Org Chart. Gusto Insights, 30 June 2025. Vendor-published analysis of its own payroll customer base.

Related SuperSkills research#

On the shape of the organisation, the shape of the organisation after AI and AI workforce strategy. On what happens to the people in the middle, the mid-career squeeze, missing rungs and synthetic seniority. On the knowledge that leaves with them, institutional memory, keeping expertise in an organisation and tacit knowledge. On decisions and authority, allocating AI decision rights and who manages AI agents. On measuring the result, how long before you know if an AI investment worked.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Garicano article was read as the typeset Journal of Political Economy text, the Autor paper in full at NBER, and the Ewens and Giroud paper in the October 2025 version at the author's own site. The Management Science article itself could not be opened; its finding is quoted from the publisher's abstract and its sample from the working paper, and that split is stated rather than smoothed over. Two widely quoted figures were removed during drafting for want of a traceable source, and both are named in the section on what this page does not claim.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.