The Third Great Restructure: How Many Managers You Need

Management is two jobs sharing one title. AI is automating one and making the other harder. Cut the layer as a single budget line and you delete both. Here is how to cut only the half that goes.

Your manager layer is two jobs sharing one title, and only one of them is in danger.

The first job is routing. Collecting status. Translating it upward, translating priorities back down. Sequencing handoffs, chasing the thing that stalled, assembling the summary that tells the floor above what the floor below did. That work exists because human attention could not span the organisation... and it is being automated right now, faster than any other white-collar function.

The second job is judgment under accountability, and the development of people. Deciding with incomplete information and owning what happens next. Telling someone the truth about their work in a way they can act on. That is not being automated. The restructure now underway is making it harder.

Most companies are about to cut both, because they are budgeted as one line item called headcount. This essay is about cutting only the first.

This is the third time we have done this

The manager layer is not a fact of nature. It was invented, on a date, for a reason.

Alfred Chandler's The Visible Hand documents the first restructure: as American railroads and then manufacturers pushed volume and speed past what owner-operators and market signals could coordinate, firms built a salaried administrative hierarchy to do the coordinating. Middle management was the technology that solved a throughput problem. Plumbing, not a moral achievement, and it worked so well that within two generations nobody remembered it had been installed.

The second restructure ran through the late 1980s and 1990s, and we have the panel data. Raghuram Rajan and Julie Wulf, working from Hewitt Associates' confidential compensation survey of large US firms across 1986–1999, found the number of managers reporting directly to the CEO rose from an average of 4.4 in 1986 to 8.2 in 1998, while the levels between division heads and the CEO fell. The layers came out. The spans widened.

Two restructures, same move: push decisions down, take a layer out, widen the span. And neither one removed the routing work. They moved it, which is why the 1990s produced the status meeting, the weekly report, the project management office, and the first software sold specifically to do the coordinating that the deleted layer used to do.

The third restructure is different, and the difference is the whole essay.

What is actually different this time?

For the first time, the routing work itself can be done by something that isn't a person.

Not the decisions. Not the accountability. The routing... reading twelve threads and writing one summary, noticing that ticket 4471 has been open eleven days, turning an engineering update into something a sales lead can act on, assembling the Monday deck. Information transformation with a defined input and a defined output... which is exactly what a well-instrumented agent stack does without complaint at three in the morning.

So this restructure does not relocate the routing work. It deletes the demand for it.

That is the structural claim behind the Hourglass Collapse: the middle of the org chart is narrow because human bandwidth could not carry information directly between a wide top and a wide bottom. Remove the constraint and the layer has no load left to bear. This essay picks up where that one ends. Not is the middle collapsing, but how many managers do I keep, and what do I do about the people.

The honest answer starts by admitting the layer is not one thing.

The half that goes: routing is a technology, not a talent

The test. Take any recurring thing a manager does and ask: does this change the state of the world, or change who knows about the state of the world? State-changing work is creation and decision. Knowledge-moving work is routing.

Routing work, in the specific:

• Status collection: the standup, the weekly roll-up, the "where are we on this" ping.
• Translation between altitudes: an IC's technical update into an executive summary; a strategy memo into three tickets.
• Sequencing and handoff enforcement: noticing A finished and B has not started, then making B start.
• Triage and assignment under known rules.
• Escalation filtering: deciding what the layer above needs to see.
• Reporting, dashboard assembly, meeting logistics and the notes afterwards.

For a sense of volume: Microsoft's 2023 Work Trend Index, measured across Microsoft 365 telemetry, found the average employee spends 57% of their time communicating, meetings, email and chat, and 43% creating. That is everyone. For someone whose entire remit is coordination, the communication share runs higher, and most of it is routing.

Andy Jassy named the pathology precisely in his September 2024 memo to Amazon employees: too many layers had produced "pre-meetings for the pre-meetings for the decision meetings," and "owners of initiatives feeling less like they should make recommendations because the decision will be made elsewhere." His remedy was arithmetic. Every S-team organisation would increase the ratio of individual contributors to managers by at least 15% by the end of Q1 2025. Amazon hit it.

Note what he was optimising for. Not cost. Ownership and speed. The bureaucracy was the product, and the layer was the factory.

The routing/judgment cut is the one the 3-Zone Co-Pilot map makes at workflow level: Replace, Augment, Refuse. Applied to a manager's calendar rather than a department, it produces one uncomfortable, useful number: the fraction of this person's week that is Replace. Run it honestly on a team lead and it comes back higher than they expect.

That fraction is what goes. Not the person. The fraction.

What is genuinely irreplaceable about management?

Now the part where this argument stops agreeing with its own headline. Managers are not obsolete. Three things sit inside the role that survive automation on their merits, and all three get more demanding as the routing disappears.

Judgment under ambiguity

Routing has a defined input and output. Judgment does not. Judgment is what you do when the information is incomplete, the options are all bad, the data supports two readings, and a decision is due anyway. Which of two credible strategies to fund, whether a client relationship is worth losing margin over, calling that a project which is technically on track is actually failing.

Agents are increasingly good at the analysis that precedes such a call. They are not the thing that makes it, because making it means holding the consequences... and consequences do not attach to software.

The common second-order trap: teams assume that because an agent produced a confident recommendation, the judgment has been done. It has not. Recommendation is routing with an opinion attached. The judgment is the acceptance.

Accountability, which is a legal fact and not a feeling

This is where the "AI will manage the team" story runs into something harder than sentiment: statute.

Illinois' HB 3773, amending the Illinois Human Rights Act, took effect 1 January 2026. It makes it a civil rights violation to use AI that discriminates on protected classes, bars ZIP codes as a proxy, and requires employers to tell employees when AI is in use, and the covered decisions include promotion, discipline, discharge and selection for training. Those are management decisions. The law does not care that an agent generated the recommendation; it holds the employer responsible for the outcome.

In the EU, employment and worker-management AI sits in Annex III of the AI Act as high-risk, and Article 14 requires such systems be designed so they can be "effectively overseen by natural persons" in use. The clock moved. The Digital Omnibus, Regulation (EU) 2026/1744, pushed the standalone Annex III deadline from August 2026 to 2 December 2027. But the obligations did not shrink. There, human oversight is the condition of legality.

The counter-example matters too. California's SB 7, the "No Robo Bosses Act," would have barred sole reliance on automated systems to discipline or fire workers. It passed the legislature and was vetoed by Governor Newsom on 13 October 2025 as overly broad. There is no settled consensus that a human must sign every decision. It is a live fight, which is why you should not design your org chart assuming it resolves in your favour.

The practical version: somebody has to be the name on the decision. Not the reviewer of an agent's output, the accountable party. That role does not compress. An agent-heavy organisation needs it defined more sharply, because there are more decisions per human and more places for accountability to go quietly missing.

Development, which is the part nobody budgets

Google ran this experiment already. In 2002, with roughly two thousand employees, it removed engineering managers entirely on the theory that they were overhead. Then reversed it. What its People Analytics team later established through Project Oxygen, reverse-engineering the gap between its best- and worst-rated managers, is that the behaviours predicting team effectiveness were not routing at all. The top one was being a good coach.

Gallup's State of the American Manager found managers account for at least 70% of the variance in employee engagement scores across business units. Seventy percent, sitting in one role. That is not a routing effect. You cannot summarise someone into caring about their work.

Now the uncomfortable connection to your AI programme. Gallup's State of the Global Workplace reports that employees who believe their manager supports their team's use of AI are 8.7 times more likely to strongly agree AI has transformed how much work gets done. Meanwhile manager engagement fell from 31% in 2022 to 22% in 2025.

Read those together and the strategy writes itself in reverse. The manager layer is the strongest determinant of whether your AI adoption sticks, and it is the layer being disengaged and cut. Companies are removing the mechanism of adoption in order to fund the adoption.

So how many managers do you actually need?

You want a ratio. Here is what the evidence supports, and where it stops.

The direction is unambiguous. Gusto's analysis of roughly 8,500 small and mid-sized US employers tracked individual contributors per people manager rising from 3.15 in January 2019 to 5.76 by September 2024. The span nearly doubled in under six years. Gallup, measuring US managers, found average direct reports moving from 10.9 in 2024 to 12.1 in 2025, which it calls nearly a 50% increase in team size since it began measuring in 2013. Amazon's 15% shift is the same move at the top of the market.

Spans are widening everywhere, at every company size, and they were widening before agents were credible. Anyone selling you a target ratio is selling you the trend line with a decimal point added.

What you need is not a ratio. It is an audit, in three steps.

One: measure the coordination share. Per manager, over two real weeks, classify the calendar and the message volume into routing versus judgment-and-development. Not from memory. From the record. Everyone skips this step, because the answer is politically expensive.

Two: automate the routing before you cut the head. In that order, non-negotiably. Cut first and the routing does not vanish. It redistributes onto ICs, who now do coordination instead of the creation work you kept them for. That is how a flattening produces a productivity decline and everyone blames the tooling.

Three: size the remaining layer on judgment and development load, not on headcount. A manager with fourteen reports doing well-instrumented, agent-supported work may be fine. A manager with six reports in a domain that throws three ambiguous calls a week is not overstaffed. Gallup's own finding is that no universal optimal span exists. It depends on team engagement, how much non-managerial work the manager absorbs, manager talent, and how often meaningful feedback actually happens.

One hard brake on the arithmetic, because the counter-evidence is real. METR ran a randomised controlled trial with 16 experienced open-source developers across 246 real issues in their own large repositories. They expected AI tools to speed them up by 24%. With the tools they took 19% longer, and afterwards still believed they had been sped up by 20%.

That is the most important number here for anyone about to cut headcount. Not because AI does not work, but because perceived and actual speedup diverge, in the optimistic direction. Size your org against projected rather than measured productivity and you will cut into muscle. And find out two quarters later, when the people are gone.

Measure first. Cut second. In that order the third restructure works. In the other order it is a layoff with a strategy deck attached.

What happens to the people in that layer?

If you flatten to fund AI, the people who used to route information go to one of three places. Two are good. Say all three out loud. Your team already knows there are three.

Back to IC, with an agent stack underneath them. The strongest destination for the manager with real domain depth beneath the coordination: the ex-engineer, ex-designer, ex-analyst who became a lead. Amazon did exactly this: some managers moved into individual contributor roles rather than out. With agents, an IC of that calibre covers what used to take a small pod. It is a promotion disguised as a title change, and it must be paid like one, or it reads as a demotion and they leave.

Up into judgment and development, at a wider span. The manager whose value was never the routing keeps the role and loses the busywork. Fewer status meetings, more calls that matter, more time on people. Expect to teach this: someone rewarded for six years for smooth coordination has no reps at being rewarded for hard calls.

Out. The manager who was purely a router, no domain depth below, no appetite for accountability above, has no role in the new shape. That is the honest part. Pretending otherwise just produces a slower, crueller version of it: eighteen months in a hollowed-out role, then leaving anyway, having lost eighteen months of re-positioning time in a market getting harder for that profile.

Say it early, fund the transition, let people choose. The alternative, quiet attrition dressed as natural wastage, produces exactly the disengagement Gallup is measuring, in the layer that determines whether your AI programme works.

One structural cost I will not re-argue here, because the apprenticeship crisis is its own essay: the middle was also where people learned to make the calls the top makes. Delete it entirely and you have solved a coordination problem and created a succession problem with a ten-year fuse. The companies handling this well are not refilling the layer. They are building the judgment reps somewhere else, which is a design task nobody has budgeted.

The sequence, and why most companies run it backwards

The default sequence is Tool, then Process, then People. Buy the licences, retro-fit the workflow, tell the humans. It is the sequence that produces licences rotting at "last opened: 47 days ago"... and the manager cut that boomerangs.

ORBIT inverts it: People before Process before Tool. On the manager question that means you do not start with a target span. You find out what the people in that layer actually do all week, redesign the coordination process so the routing has somewhere else to live, and only then buy or build the thing that lives there. The org chart change is the last step. A consequence of the redesign, not the instrument of it.

ORBIT's core is a map of five functions every AI-augmented team has to cover: Orchestrate (direct the agents toward an outcome), Run (manage the workflows the agents are inside), Build (create the systems, tools and prompts they operate within), Influence (drive adoption and culture change) and Translate (bridge agent outputs and human decisions).

Lay the two halves of the manager's job over that map and the restructure gets specific. The routing half doesn't map to a function at all. It's what the agents absorb. The judgment half maps to two: Orchestrate, a CEO function compressed into a single team, and Translate, the person who looks at what the agents produced and says "right, act on it" or "plausible but wrong." And a job appears that nobody budgeted: Run. Operations doesn't disappear when the routing is automated. It gets denser... and someone has to catch what fails silently.

That's the honest answer to "how many." Not a ratio: enough people to cover five functions per pod, by combination in a team of two or three, one name per function at four to six, and a split at ten to twelve.

The full walkthrough is at ORBIT.

The wider cluster on org shape sits under AI org design.

The sequencing failure is not an intellectual mistake. It is a budgeting mistake. The headcount line is available this quarter and the redesign is not, so companies cut what the budget can reach and hope the process catches up. It does not. It lands on the ICs, and six months later the productivity number is worse and the diagnosis is "the AI didn't deliver."

The binary

Management is two jobs. One is routing, and it is going... faster than your planning cycle. One is judgment, accountability and the development of people, and it is not going... it is getting harder.

Two ways to run the third restructure.

Cut the layer as a line item, then discover in two quarters that the routing simply moved onto your ICs, and watch the people who would have made your next set of hard calls leave for somewhere that knew the difference.

Or measure the coordination share first, move that half to the agents, and pay the remaining half more to do the part that was always the actual job.

If you're running a 20–200 person company and you already know which one you're doing, book a Fractional CAIO scoping call and bring one manager's real calendar from the last two weeks. That's the whole diagnostic. It takes thirty minutes and it usually ends the argument.

If you'd rather work it through yourself, the org-design series goes out to the newsletter.

Read this next: The Apprenticeship Crisis... the succession cost this essay deferred: where the next senior people come from once the middle is gone.

FAQ

How many employees do I actually need if AI agents are doing most of the work now?

Fewer than your org chart implies, more than your enthusiasm implies. And you cannot know until you measure. The trap is sizing headcount against projected agent productivity. METR's randomised trial of 16 experienced developers on 246 real issues found they took 19% longer with AI tools while believing they had been 20% faster. So: automate one workflow, measure the throughput change over a real month, then adjust the headcount plan for that workflow. The companies getting hurt cut against a forecast.

What actually prevents AI from managing all my direct reports?

Three things, in descending order of hardness. Law: Illinois HB 3773, effective 1 January 2026, covers promotion, discipline, discharge and training selection when AI is involved, and holds the employer accountable for the outcome; the EU AI Act classes employment and worker-management systems as high-risk and requires under Article 14 that they be effectively overseen by natural persons. Accountability: a decision needs a name on it, and consequences do not attach to software. A recommendation is not a decision. Development: Gallup finds managers account for at least 70% of the variance in team engagement, and Google's Project Oxygen found the behaviour most predictive of team effectiveness was coaching. None of that is routing.

Do companies even still need managers now?

Yes. Fewer of them, doing a narrower and harder job. The routing work that justified most of the layer's headcount is being automated; the judgment, accountability and development work is not, and it is expanding. What disappears is the pure coordinator: the role that exists to move information between two groups who could not otherwise reach each other. Spans widen accordingly: Gusto measured individual contributors per manager going from 3.15 in early 2019 to 5.76 by September 2024 across roughly 8,500 small and mid-sized employers; Gallup measured average direct reports rising from 10.9 in 2024 to 12.1 in 2025. Wider spans, not zero managers.

Why will AI never replace managers?

"Never" is the wrong word and I would not defend it. AI has already replaced a large share of what managers do: status aggregation, translation, sequencing, reporting. What it has not replaced, and has no current path to replacing, is the part requiring someone to be accountable for an outcome and to develop another human being. Accountability is a legal and social fact about persons, not a capability you can benchmark: Illinois and the EU have written that assumption into statute, while California's attempt to go further (SB 7, the "No Robo Bosses Act") was vetoed in October 2025. The boundary is contested, not settled. Google tested the strongest version of the question in 2002, removing its engineering managers entirely, and reversed it. The durable claim is not that managers are safe. It is that management, correctly defined, is a different job from the one most managers spend their week doing.

If my company flattens the manager layer to fund AI, what happens to the people who used to report up through it?

Two things happen and only one is on the plan. The reports get a wider-span manager, Gallup's 2025 average is 12.1 direct reports, so less individual attention at exactly the moment they are asked to change how they work. And the routing the removed layer absorbed lands on them, unless it was genuinely rehomed to an agent stack first. That combination is how a flattening produces a measured productivity decline. The mitigations are unglamorous: automate the coordination before removing the head, name the accountable decision-maker for every workflow explicitly rather than assuming it flows upward, and protect the development conversations first when the surviving manager's calendar tightens. Gallup's 70%-of-engagement-variance finding does not soften because a span doubled.

Sources

• Chandler, Alfred D. Jr. (1977). The Visible Hand: The Managerial Revolution in American Business. Harvard University Press. origin of the salaried administrative hierarchy as a coordination technology.
• Rajan, Raghuram G., and Julie Wulf (2003). The Flattening Firm: Evidence from Panel Data on the Changing Nature of Corporate Hierarchies. NBER Working Paper 9633 (later published in the Review of Economics and Statistics). Data: Hewitt Associates confidential compensation survey, large US firms, 1986–1999. CEO direct reports rose from an average of 4.4 (1986) to 8.2 (1998); depth between division heads and CEO decreased.
• Amazon / Andy Jassy (2024, 16 September). Update on return-to-office plans and manager-to-team ratio, aboutamazon.com. the ≥15% individual contributor–to-manager ratio target by end of Q1 2025; the "pre-meetings for the pre-meetings" description of layer-driven bureaucracy.
• Gusto (2025). The Manager Mass Exodus: How SMBs Are Flattening the Org Chart. individual contributors per people manager rose from 3.15 (January
  2019. to 5.76 (September 2024) across roughly 8,500 small and mid-sized US employers.
• Gallup / Jim Harter (2026, 13 January). Span of Control: What's the Optimal Team Size for Managers? average direct reports rose from 10.9 (2024) to 12.1 (2025), described as nearly a 50% increase in team size since 2013; no universal optimal span.
• Gallup. State of the American Manager: Analytics and Advice for Leaders. managers account for at least 70% of the variance in employee engagement scores across business units.
• Gallup (2026). State of the Global Workplace. manager engagement fell from 31% (2022) to 22% (2025); employees who believe their manager supports their team's AI use are 8.7× more likely to strongly agree AI has transformed how much work gets done.
• Microsoft (2023). Work Trend Index Annual Report: Will AI Fix Work? average employee spends 57% of Microsoft 365 time communicating (meetings, email, chat) and 43% creating; 28-day rolling window ending March 2023.
• METR (2025, 10 July). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089. 16 developers, 246 real issues; 19% longer with AI tools, against an expected 24% speedup and a post-hoc believed 20% speedup.
• Garvin, David A. (2013, December). How Google Sold Its Engineers on Management. Harvard Business Review. Project Oxygen; the 2002 removal of engineering managers and its reversal; coaching as the top predictive behaviour.
• Illinois HB 3773 (effective 1 January 2026), amending the Illinois Human Rights Act. AI use in promotion, discipline, discharge and training selection; notice requirement; ZIP-code proxy prohibition.
• Regulation (EU) 2024/1689 (AI Act), Annex III and Article 14; as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI, in force 27 July 2026). employment and worker management as high-risk; human oversight requirement; standalone Annex III compliance deadline moved from 2 August 2026 to 2 December 2027.
• California SB 7 ("No Robo Bosses Act"), vetoed by Governor Newsom, 13 October 2025. Would have barred sole reliance on automated systems for discipline and termination.