How to Restructure for AI

Most advice on restructuring for AI is true and useless. Rearchitect the operating model. Rethink where value comes from. Retrain the people. All correct, all now conventional, and none of it tells you what to build on Monday. The conceptual case was won years ago; what remains is producing the new shape, and that requires knowing which lever actually produces it. Of everything a firm could change around AI, the load-bearing thing is the structure: how work is grouped, who hands off to whom, what gets decided where. The evidence on what happens when you change that is older, richer, and less convenient than the pitch admits.

A general-purpose technology pays off through the shape around it

A general-purpose technology gives back what you put it next to. Bolt a model onto the org you already have and you get the org you already have, slightly faster. Harold Leavitt described the reason in 1965. An organization is an interdependent system of task, structure, people, and technology, and a change pushed into any one of the four produces compensating change in the others.¹³ Push on the technology alone and the rest of the system absorbs the push. The value comes from changing the shape around the technology, and that is where almost everyone underspends.

The clearest number on the underspend comes from BCG, whose guidance for capturing value from AI splits the effort 10-20-70 — roughly 10 percent on the algorithms, 20 percent on technology and data, and 70 percent on people and processes.¹ The same document separates two ways to spend. “Deploy” puts off-the-shelf tools onto existing workflows and lifts productivity 10 to 15 percent, while “Reshape” re-engineers the workflow itself and drives 30 to 50.² Both figures are consultancy estimates with no published methodology behind them, so I take their direction and hold their precision loosely. The direction is enough. Deploy bolts the tool onto the shape you have, Reshape changes the shape, and firms that swap AI into their existing silos and expect transformation are buying the small number and calling it the big one.

The mechanics are older than the technology. In 1968 Melvin Conway observed that any organization designing a system “will inevitably produce a design whose structure is a copy of the organization’s communication structure.”³ If your firm is built as siloed domains, it will build and serve siloed products, because the seams in the org become the seams in the output. A team that owns the whole value stream produces whole work. Conway supplies the mechanical basis for treating structure as the lever. Choose the structure and you have largely chosen the product, before anyone writes a line of code or a word of strategy.

Culture is an output of the design

The usual instinct, once a leader accepts that AI demands change, is to go after culture — run the town halls, change the language, ask people to think differently. It mostly fails, and the reason is structural. Jay Galbraith’s model of organization design treats strategy, structure, processes, rewards, and people as five interlocking levers, and treats culture and performance as outputs of getting those levers aligned.⁴ An output has no handle to push on directly. Try to move the culture while the structure that produces it stays fixed, and the structure wins; change the structure correctly and the behavior follows, because the new shape rewards different behavior and the old habits stop paying.

The catchier version of this claim, “architecture eats culture,” is a riff on a Drucker line with no clean origin, so it stays out of the argument as a maxim.⁵ The defensible form is Galbraith’s, and it is enough. An aligned design produces the culture, which means the leader who wants different behavior gets it by re-cutting the design the behavior grows from.

Provisioned capability turns span and layers into parameters you set

The reason AI reaches all the way to structure is that it changes what fixed the structure in place. Two things have always set the shape of an organization: span of control, how many people one manager can hold, and the number of management layers between the top and the front line. Both were set by human cognitive limits. One person can oversee only so many others whose work they must understand, check, and coordinate, so the hierarchy grew tall to fan out the supervision. Deloitte’s account of the layers adds what the tall structure was carrying. Hierarchies existed because humans managed information scarcity through sequential handoffs, and the layers existed largely to do the “coordination and synthesis work” between them.¹²

Provisioning capability loosens that constraint at its base. When output capacity stops being tied to headcount, part of the coordination and synthesis the layers carried moves into the system itself, which is exactly the work Deloitte finds AI removing. (The labour-economics modeling behind the uncoupling is set out in an earlier paper in this corpus.) Span and layers then turn from inherited constraints into parameters a firm sets, and the setting is a decision with distributional consequences.

The ground is already moving, though the motion is easy to misread. Gallup measured the average number of direct reports per manager at 10.9 in 2024 and 12.1 in 2025, while the median manager held steady at about five to six reports.⁶ The distribution explains the gap between those two facts. Thirty-seven percent of managers oversee fewer than five people, 13 percent oversee 25 or more, and it is the tail of very large teams that drags the average up while the typical manager’s week stays put.⁶

Organize around the stream of value; the shared work moves into the foundation

If structure is the lever, the next question is which structure, and the durable answer predates AI. Organize around the flow of value, and let the flow define the team. Amazon’s “single-threaded leader” is one person fully dedicated and accountable to a single product, on a team that owns its work end to end with no handoffs, governed by guardrails rather than tollgates.⁷ Team Topologies generalizes the same move: the stream-aligned team owns its slice of value and the flow through it, instead of passing work across functional walls.⁸ The seam between teams is where Conway’s law bites, so you put the seams where you want them in the product, around whole streams of value.

AI sharpens this from a good idea into a structural necessity, because a single stream-aligned unit can now carry capability it used to borrow from other departments. It also reverses the direction of design. A siloed firm designs outward from the internal actor — the practitioner’s need, then the features, then the process, then whatever reaches the client at the end. Agents can serve the end-client directly, which makes the honest starting point the other end of the chain. Begin with the client’s need, then the human action that adds judgment or relationship value, then the process, then the systems. Design inward, and hold the human central only where the human is the value.

I run an operating model built on these commitments. A single data foundation sits under the whole group, one foundation for everyone in place of one stack per silo, and infrastructure, governance, and evaluation sit down there with it. An agent layer grows out of that foundation into every function, and I treat it as a layer of talent, a different kind of team member suited to different jobs, with the cognitive load divided across the loop accordingly. The humans sit outside the agent layer and orchestrate, carrying the judgment calls, the audience crossings, and the accountability, which crystallizes at the moment work is delivered to someone who will rely on it. And every spoke carries two lines, work flowing outward and feedback flowing back into the data core, because learning that only walks out in people’s heads is learning the system never keeps. The full model, and why each piece sits where it does, is in a companion essay.

That shape also corrects the absolute version of the stream-aligned story. The tidy telling has every function dissolving into the streams, and a shared data foundation under the whole group is a function by any name. Team Topologies predicts as much. The platform team exists as one of its four team types because pure stream alignment does not survive contact with shared infrastructure.⁸ The streams own the value. What the streams share moves down into the foundation and gets owned there once, for everyone.

Span may contract before it expands, and the shape is a chosen regime

“AI flattens the org” is too glib, and it is too glib in both directions. My own reading of the modeling here forks, and I flag it as my read rather than a settled result. Under one branch, treating agent capability as capital that reduces the friction of managing workers genuinely does widen spans and compress hierarchies. But the same technology produces very different shapes depending on the parameters. Where AI arrives as broad infrastructure available to everyone, the widening spreads; where it selectively amplifies the top, power concentrates into a few superstars. The technology delivers a range of outcomes, the firm lands in one of them by choice, and the choice is a policy lever with winners and losers.

The second branch bears on the timing of the cut. Span of control may contract before it expands, because early, fallible AI needs more skilled oversight, and the widening arrives only as the technology matures. A firm that flattens immediately on the promise of agents is cutting the oversight it most needs precisely in the window when the agents are least reliable. And the slack it would cut into is already thin; Gallup finds the median manager spending 40 percent of their time on individual-contributor work.⁶

Read through that fork, the Gallup distribution looks like widening that is real but uneven and early, a tail of teams that have crossed into the wide-span regime above a large middle that has not and should not yet. I hold that reading loosely. A flat median under a rising average is also consistent with duller stories, post-layoff mechanical widening at large firms among them, and the data cannot pick between them. What the distribution does rule out is the simple pitch, in which everyone’s span widens together because the technology arrived.

Flattening pays only when the layer’s work is re-homed

The case for flattening has a long and inconvenient evidence trail, and it is worth reading before you cut. In 1995 the Institute for Employment Studies looked hard at the previous wave of delayering and refused the headline question outright, judging its own research “too small, and conducted over too short a period of time, to support or deny that flatter is better” and warning employers off “simplistic goals that attribute business success to no more than five layers of management.”⁹ What it would commit to is where the cost lives: “Delayering is unlikely to bring sustainable cost advantages on its own. Changing the way in which the work is done and removing unnecessary tasks that fail to add any value is as important, if not more so, than simply changing the levels of managers doing it.”⁹ The hierarchy you are cutting was doing real work — coordination, feedback, mentorship, synthesis — and that work does not vanish when the layer does. It relocates, usually onto whoever is left.

Thirty years on, the same failure is being rediscovered in real time. After a fresh round of middle-management cuts, Fortune reported nearly half of senior executives doubting their ability to manage the load they had absorbed, with the survey behind the piece finding the same strain running down the organization.¹⁰ The mentorship and career-development function that middle managers quietly performed simply disappeared, and the high performers who depended on it started to leave.¹⁰ The fix the piece lands on is intentional role redesign plus the infrastructure to do what the layer did, and it recommends that over restoring the layer.¹⁰

Gallup gives the discipline an empirical gate. Employees were highly engaged, about seven in ten, at any team size when they strongly agreed they received meaningful feedback; when they did not strongly agree, engagement fell to about one in four.⁶ A wide span works when the feedback survives the widening and fails when it does not. That is the 1995 finding restated as survey data, and it converts “re-home the work” from a caution into a design requirement with a number on it.

So the dissent and the thesis are the same discipline viewed from opposite ends. Span and layers are yours to set, and the setting pays only if the design re-homes the work the layer performed. Cut the layer and draw nothing in its place and you have reproduced the 1995 mistake with better tooling. The widely cited forecast that a fifth of organizations will use AI to eliminate more than half their middle-management roles describes a thing that will happen; whether it works is a separate question, and the evidence says it works only when the coordination, feedback, and accountability the layer carried are explicitly rebuilt somewhere else.¹¹

Map the work first; rebuild the ladder last

The re-homing has a published sequence. Deloitte’s restructuring blueprint runs five steps in order: map the work, instrument the governance, fuse roles selectively, add guardrails, and rebuild the career ladder.¹² The order carries the discipline. Mapping comes first because the unit being redesigned is the work, and Deloitte goes as far as replacing the org chart with a “work chart” on which “tasks and processes — not people — are mapped to show how work gets done.”¹² Governance and guardrails hold the middle because they are the re-homing itself, the built replacement for the checking and coordination a cut layer used to perform. Role fusion sits between them and is explicitly selective, aimed at the “repeatable, artifact-heavy, and coordination-driven” work where handoffs can actually come out.¹² And the ladder is last on the list but on the list, which is the sequence conceding that the restructuring reaches the people who were climbing it.

The costs are concentration and a narrower on-ramp to judgment

Two costs are built into this design. The first is concentration. As agents take on more, the human contribution concentrates into fewer people, each carrying greater breadth. Fewer people. I am not going to call that an upgrade. It is the distributional fact of the model, and a leader owes the organization the plain version of it.

The second is the apprenticeship problem. If agents absorb the entry-level work that used to be how juniors learned the craft, the on-ramp to senior judgment narrows. Deloitte frames the shift as juniors moving from “learning by doing” to “learning by evaluating,” which is real but presumes a base that doing used to build.¹² And where junior roles are disappearing outright, a deployment choice made them disappear. Firms chose automation over augmentation for that work, mostly by default; nothing in the technology forces the choice, since augmenting the junior leaves the rung standing. I have a working answer to the apprenticeship problem and I do not have a proof, and because it is properly a question about the individual’s role, I take it up in the essay on the AI-native contributor. For the restructuring decision the point is narrower and harder. The org chart you draw determines whether that rung survives, so you are making a choice about the next generation’s judgment whether you mean to or not. The flattening and the hollowing are the same decision seen from two heights.

The structure is a decision you keep making

What has stayed durable in my own model is the shape: the foundation under everything, agents as talent, humans orchestrating from outside, feedback drawn as a return line. What has never stayed durable is the split of work between humans and agents. That split is a snapshot, and it moves every time the agents improve, which means the design decision does not stay made.

The organizational literature has a name for the capacity this demands. O’Reilly and Tushman call it ambidexterity, “the ability to simultaneously pursue both incremental and discontinuous innovation … from hosting multiple contradictory structures, processes, and cultures within the same firm,” and after fifteen years of studies they conclude that achieving it “is, at heart, a leadership issue more than a structural one.”¹⁴ The AI version of that leadership test asks for two disciplines held at once. The first is iterating the working model frequently, so the human-agent split tracks what the agents can actually do. The second is questioning, with fresh eyes, whether the assumptions the whole design rests on still hold. Loyalty to a fixed model fails. So does endless iteration that never stops to ask whether the frame still applies. Most of the difficulty is that the two pull against each other, and the leaders who get this right will be the ones who can stand the pull.

Restructuring for AI, then, is the standing exercise of that judgment. The spans and layers of your firm are parameters someone now sets, deliberately or by default; the evidence since 1995 says a new setting pays only when the work the old setting did is rebuilt on purpose; and the technology that turned the shape into a decision will keep moving what the right decision is. The firms that capture the value will be the ones still deciding.

Sources

1. Boston Consulting Group, The CEO's Guide to Maximizing Value Potential from AI (July 2024), p. 2: the 10-20-70 approach — algorithms (10%), technology and data (20%), people and processes (70%). Primary for the 10-20-70 figure; the public asset URL now returns HTTP 403 to automated fetch, quoted here from a locally held copy of the PDF. https://www.bcg.com/
2. BCG, CEO's Guide (July 2024), p. 2: "Deploy" lifts productivity 10–15% with off-the-shelf tools; "Reshape" drives 30–50% efficiency gains through workflow re-engineering.
3. Melvin E. Conway, "How Do Committees Invent?", Datamation, April 1968: "Any organization that designs a system … will inevitably produce a design whose structure is a copy of the organization's communication structure." https://www.melconway.com/Home/Committees_Paper.html
4. Jay R. Galbraith, Star Model (organization design as five interdependent levers — strategy, structure, processes, rewards, people — with culture and performance as outputs of their alignment), documented summary, Strategic Management Insight. Galbraith's primary texts were not available online; the model substance is uncontroversial and widely documented. https://strategicmanagementinsight.com/tools/galbraiths-star-model-explained/
5. "Architecture eats culture" has no single canonical origin; it is a derivative of the line commonly attributed to Peter Drucker, "culture eats strategy." The most traceable public articulation of the escalated phrase is Tim Bouma, "Architecture Eats Culture Eats Strategy," Medium, 31 July 2019, who frames it explicitly as a play on Drucker. The phrase is not used here as a sourced maxim; the underlying claim is backed by Galbraith (note 4). https://trbouma.medium.com/architecture-eats-culture-eats-strategy-925c7f3fb74b
6. Jim Harter et al., "Span of Control: What's the Optimal Team Size for Managers?", Gallup, 14 January 2026: average direct reports rose from 10.9 (2024) to 12.1 (2025), nearly 50% above the 2013 baseline; the median held steady at about five to six. Distribution: 37% of managers oversee fewer than five people, 22% have 10 to 24, and 13% oversee 25 or more. Engagement at wide spans held (about seven in ten) only when employees strongly agreed they received meaningful feedback; without it, about one in four. Managers spend a median of 40% of their time on individual-contributor work. https://www.gallup.com/workplace/700718/span-control-optimal-team-size-managers.aspx
7. Tom Godden (AWS), "Two-Pizza Teams Are Just the Start (Part 2)," AWS Enterprise Strategy Blog, 18 March 2021: the single-threaded leader, end-to-end ownership ("you write the code, you fix the code"), guardrails rather than tollgates. https://aws.amazon.com/blogs/enterprise-strategy/
8. Matthew Skelton & Manuel Pais, Team Topologies, "Key Concepts": organize around the flow of value; the stream-aligned team owns its work with no handoffs; the platform team — "a grouping of other team types that provide a compelling internal product to accelerate delivery" — is one of the four fundamental team types. https://teamtopologies.com/key-concepts
9. Polly Kettley, Is Flatter Better? Delayering the Management Hierarchy, Institute for Employment Studies, Report 290 (1995): "Delayering is unlikely to bring sustainable cost advantages on its own. Changing the way in which the work is done … is as important, if not more so, than simply changing the levels of managers doing it." Also: "This research is too small, and conducted over too short a period of time, to support or deny that flatter is better," and the warning against "simplistic goals that attribute business success to no more than five layers of management etc." A 1995 study of the pre-AI delayering wave, used for its durable structural lesson. https://www.employment-studies.co.uk/system/files/resources/files/290.pdf
10. Lily Mae Lazarus, "Executives are drowning after cutting out their middle managers," Fortune, 21 April 2025, citing the Korn Ferry 2025 Workforce Survey: nearly half of senior executives doubt they can manage the added load; 41% of employees report cut management layers, 37% feel directionless, 43% say leadership is not aligned; the fix is intentional role redesign, not re-adding layers. https://fortune.com/2025/04/21/senior-leaders-burden-cuts-middle-management
11. Gartner 2025 CEO survey prediction (as reported across secondary outlets): through 2026, ~20% of organizations will use AI to flatten structure, eliminating more than half of current middle-management roles in those firms. Secondary and predictive — the Gartner primary is gated and this figure is unverified against it; treated as directional color only. https://www.gartner.com/en/documents/6494771
12. Vivek Kulkarni, Maya Bodan & Jamie Kilgour, "The work chart vs. org chart: AI-driven organizational delayering and cross-functional role fusion," Deloitte US, 16 June 2026: AI reduces the "coordination and synthesis work" that justified management layers; traditional hierarchies managed information scarcity through sequential handoffs; the restructuring sequence runs map work → instrument governance → fuse roles selectively → add guardrails → rebuild the career ladder; role fusion applies to "repeatable, artifact-heavy, and coordination-driven" work; the "work chart" maps "tasks and processes — not people"; entry-level work shifts so that "instead of 'learning by doing,' juniors learn by evaluating." https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/articles/role-fusion-organizational-delayering.html
13. Harold J. Leavitt, "Applied Organizational Change in Industry: Structural, Technological and Humanistic Approaches," in James G. March (ed.), Handbook of Organizations (Rand McNally, 1965), pp. 1144–1170: organizations as interdependent systems of task, structure, people (actors), and technology, in which change pushed into one variable produces compensating change in the others. The 1965 print chapter is not available online; cited from the verified bibliographic record and the documented model summary, not from the original prose. https://www.econbiz.de/Record/applied-organizational-change-in-industry-structural-technological-and-humanistic-approaches-leavitt-harold/10002360285
14. Charles A. O'Reilly III & Michael L. Tushman, "Organizational Ambidexterity: Past, Present, and Future," Academy of Management Perspectives 27(4), 2013, pp. 324–338: ambidexterity defined (from Tushman & O'Reilly, 1996) as "the ability to simultaneously pursue both incremental and discontinuous innovation … from hosting multiple contradictory structures, processes, and cultures within the same firm"; achieving it "is, at heart, a leadership issue more than a structural one." Full text via the open-access HBS-hosted accepted manuscript. https://journals.aom.org/doi/10.5465/amp.2013.0025