What Org Structure Actually Costs

Key takeaways

• Roughly one American job in eight is supervisory: BLS counts 11.1 million management jobs plus about 7.7 million first-line supervisors out of 155.5 million, or about 7.2 non-supervisory workers per supervisor.
• Widening average span from 5 to 8 deletes a whole layer: in a 10,000-person illustration, about 1,071 manager roles. At the BLS mean management wage of $145,260 that is $156 million a year in wages, or roughly $227 million once loaded 1.46x for benefits at the ECEC ratio.
• Severance was the smaller half of Meta's reorg: of $3.45 billion in 2023 restructuring charges, $2.51 billion was facilities and only $1.17 billion was people, against a $0.22 billion credit on data center assets.
• Structure sets the ceiling, not the outcome: 40% of the variation in management practice sits between plants inside the same firm (Bloom et al., AER 2019).

The bill you can see

Structure has a payroll line, and it is large. In the May 2025 Occupational Employment and Wage Statistics, the U.S. Bureau of Labor Statistics counted 11,132,700 jobs in management occupations out of 155,495,730 jobs overall, at an annual mean wage of $145,260 against $69,770 across all occupations. Multiply the two published figures together and management wages alone come to roughly $1.62 trillion a year, before a dollar of benefits.

That undercounts supervision, because BLS files first-line supervisors inside the group they supervise rather than in the management category. The shift lead is coded with the crew, not with the executives. Summing the twenty first-line-supervisor occupation codes adds another 7,722,740 jobs. Together that is 18,855,440 supervisory jobs, 12.1% of all employment, or about 7.2 non-supervisory workers for every supervisory job in the country. Roughly one job in eight exists to direct other jobs.

The multiplier on all of it comes from the same agency. BLS's Employer Costs for Employee Compensation put civilian compensation at $49.32 per hour worked in the first quarter of 2026, split $33.72 in wages and $15.60 in benefits. Benefits are 31.6% of total compensation, so every dollar of salary carries about $1.46 of employer cost. Any headcount arithmetic that uses salary alone is understating by roughly a third.

None of which means managers are waste. A manager who removes a blocker that would have cost a ten-person team a week has paid for a quarter. The visible bill is misleading in a different way: it invites the wrong question. "How many managers do we have?" is answerable and nearly useless. "What does each layer decide that the layer below cannot?" is hard and is the only version that leads anywhere. A layer that reviews, summarizes and forwards is not managing; it is a relay, and relays are the specific thing that org design is supposed to remove.

Span of control: the arithmetic of layers

Span of control (the number of people reporting to one manager) and total headcount together determine how many layers you have. You do not get to choose all three.

The relationship is logarithmic: layers ≈ log_span(headcount). Illustrative only, with round numbers. Take 10,000 frontline workers. At an average span of 5, you need 6 layers of management above them and about 2,500 manager roles (10,000 ÷ 5 = 2,000 first-line, then 400, 80, 16, and so on up). At a span of 8, you need 5 layers and about 1,430 managers. Widening the average span by three people deletes 1,071 management jobs and one entire level of the hierarchy.

Price it with the BLS figures above: $145,260 mean management wage × 1.46 loading ≈ $212,000 fully loaded per role, so 1,071 roles is roughly $227 million a year. That is the number that gets put in a deck, and it is the least interesting number in this briefing.

The more important saving is latency. A decision that has to climb from the floor to the top and come back crosses 12 handoffs in a 6-layer organization and 10 in a 5-layer one. Assume one business day of queue at each handoff, optimistic in most companies where the wait is for a person's calendar rather than their attention, and that is two and a half working weeks versus two. Over a year of decisions, the compounding of that difference dwarfs the payroll saving, and unlike the payroll saving it shows up in revenue rather than in cost.

Here is the trade that gets left out of the deck. Span is not free to widen, and what breaks first is not the work. It is onboarding, performance management, and the quality of the escalation path. A manager with five reports can hold a real picture of each one: what they are good at, what they are struggling with, whether they are about to quit. At fifteen, they hold a real picture of the loudest three and a status summary of the rest. New joiners take longer to become productive because nobody has time to teach them; weak performers persist because managing someone out is expensive attention nobody has; and the good people who need a sponsor to get promoted do not get one. Those costs are real, they land 6 to 18 months after the reorg that caused them, and nothing in the accounting connects the two.

Which is why span should vary with the work rather than be issued as a policy number. Routine, standardized, independently-executed work supports wide spans: a shift supervisor can genuinely run twenty people doing the same task. Novel, interdependent, high-judgment work does not, because the manager's job there is arbitration and context, and both scale badly. An enterprise that mandates a uniform target span across every function has guaranteed itself the wrong span twice: too tight where the work is routine, too loose where it is not.

Nucor is the live demonstration that the flat end of the range is achievable at scale. Its fiscal 2025 10-K reports approximately 33,000 teammates and states that the organization is highly decentralized, with most day-to-day operating decisions made by division general managers and their teams, and that approximately 200 teammates work in its principal executive offices in Charlotte. Corporate is 0.6% of the company. The 2018 filing tells the same story at a smaller size: slightly more than 100 people in the executive offices against roughly 26,300 employees. The headquarters grew by 100 people while the company grew by nearly 7,000.

Coordination is the cost that never gets itemized

Every person you add to a group adds more relationships than people. That is the whole of the coordination problem, and it is arithmetic rather than culture.

The number of pairwise links in a group of n people is n(n−1)/2. Five people have 10; ten have 45; twenty have 190; fifty have 1,225. Doubling a team from ten to twenty multiplies the relationships that need maintaining by 4.2. Fred Brooks made this the centerpiece of The Mythical Man-Month half a century ago, and the reason the observation has not aged is that the formula has not changed.

Amazon built a structural rule out of it. AWS's own account of two-pizza teams states the principle without hedging: no team should be big enough that it would take more than two pizzas to feed them, and ideally this is a team of fewer than 10 people, because smaller teams minimize lines of communication and decrease the overhead of bureaucracy and decision-making. The document also names the Ringelmann effect (individual effort falls as group size rises) and, crucially, insists that size alone is not the point. Two-pizza teams have single-threaded ownership of one service, end to end, and do not hand what they launch to another team to run. The size cap is what makes the ownership possible; the ownership is what makes the size cap worth having.

Meeting cost is the easiest coordination expense to compute and the least important one. Illustrative only: a weekly 60-minute cross-team meeting with 12 managers. At the BLS mean management wage, $145,260 ÷ 2,080 hours is $69.84 an hour, times the 1.46 loading is about $102 fully loaded. Twelve people for an hour is roughly $1,226; across 52 weeks, about $64,000 a year. Add thirty minutes of preparation and follow-up each and it is closer to $95,000. Real money, and still a rounding error next to what that meeting is actually costing you.

The expensive part is the queue. A weekly forum means that any decision needing that forum waits, on average, half a week and at worst a full week, no matter how long the decision itself takes to make. If eight decisions a quarter route through it, you have bought roughly four weeks of pure waiting for $64,000 of attention. Moving the forum to twice weekly halves the wait and barely changes the cost, which tells you that meeting frequency is a latency lever, not a cost lever, and should be reasoned about that way.

The design conclusion follows. Every recurring cross-team meeting is evidence of a standing dependency. The cheap intervention is almost never to run the meeting better; it is to move the dependency inside a single team so the meeting stops being necessary. When that is impossible, because some dependencies are real and permanent, the next cheapest intervention is to write down who decides when the parties disagree, because most coordination time is not spent exchanging information. It is spent waiting for someone with the authority to end the disagreement.

What a reorg actually costs

A reorganization has three prices, and only one of them is booked.

The booked one is visible in Meta's filings, which are unusually clean because the company disclosed the composition. For the full year 2023 Meta recorded $3.452 billion of restructuring charges: $2,506 million in facilities consolidation, $1,170 million in severance and other personnel costs, and a $224 million credit on data center assets. In 2022 it recorded $4.61 billion. Two years, about $8.06 billion, against $88.15 billion of total costs and expenses in 2023. The 2023 charge alone was 3.9% of the company's cost base. Headcount finished 2023 at 67,317, down 22% year over year.

Read the split again, because it is the counterintuitive part. Severance was 34% of the 2023 charge. Buildings were 73%. When people picture the cost of a reorg they picture payouts to departing staff; in the largest well-documented recent example, the offices cost more than twice what the people did. Any organization carrying leases, fitted-out floors, or committed infrastructure should model the property consequences of a headcount decision before the headcount consequences, because that is where the money is.

The second price is legally fixed. Under the Worker Adjustment and Retraining Notification Act, an employer with 100 or more employees must give 60 days' written notice before a covered plant closing or mass layoff. A plant closing is an employment loss for 50 or more employees at a single site within any 30-day period; a mass layoff is 500 or more, or 50 or more amounting to at least 33% of the employees at the site. That statute sets a floor on the interval between deciding and executing: two months in which everyone affected knows, everyone unaffected suspects, and no one commits to anything new. The notice period is not a cost line, but it is a productivity event, and it is the only part of the schedule you cannot compress.

The third price is never booked at all: the months beforehand, when senior people spend their attention competing over boxes rather than running the business; the months afterwards, when reporting lines, approval paths and informal networks are all being relearned at once; and the institutional memory that walks out with the people who knew why things were built the way they were.

The evidence that this is more than inconvenience is not anecdotal. Guthrie and Datta, publishing in Organization Science in 2008 on a matched sample of U.S. manufacturing firms, found that downsizing is associated with decreases in subsequent firm profitability, and that the negative effects were more pronounced in industries characterized by research-and-development intensity, growth, and low capital intensity. In plain terms: cutting hurts most exactly where the value lives in people rather than in plant, which is to say in most modern businesses.

And then there is the part almost nobody models. Reorganizations reverse. Meta's headcount was 78,865 at 31 December 2025, up 6% year over year: about 17% above the post-cut trough and back to roughly 91% of the pre-cut level implied by the reported 22% decline, on revenue of $200.97 billion. The 2022–23 restructuring was real, it was expensive, and within two years most of the headcount had returned in different shapes. That is not necessarily a failure, since buying two years of cost flexibility may have been worth $8 billion, but it is a different transaction from the one described in the announcement. Model the round trip: the charge, the notice period, the productivity trough, the rehiring, and the re-onboarding. If the case only works assuming the cut is permanent, check that assumption against your own hiring history.

Every structure is a bet about which coordination problem you would rather have

There is no structure without a tax. Choosing one is choosing which handoff becomes easy and which becomes hard, and the honest way to evaluate a proposal is to name the tax out loud before you adopt it.

A functional structure (all the engineers together, all the marketers together, all the finance people together) buys depth of craft, consistent standards, cheap specialist hiring and a clear career ladder inside each discipline. The tax is that nothing reaches a customer without crossing every function, so delivery speed is set by the slowest queue in the chain and no single person owns the outcome. When something ships late, everyone can honestly say their part was done on time.

A divisional or product structure (a team per product, market or customer segment) buys end-to-end speed and unambiguous ownership. This is the two-pizza model. The tax is duplication (every division builds its own version of the same capability), drift (standards diverge because nobody owns them across units), and professional isolation (the only data engineer in the division has nobody to learn from and no obvious next role).

A matrix tries to buy both and pays a third tax instead: two reporting lines, and therefore a permanent supply of disagreements that have no natural resolver. Matrices work in exactly one condition: the tie-break is written down before the tie. Who decides when the functional leader and the product leader disagree, on which classes of decision, and how fast? Answer that and a matrix is a legitimate design. Leave it unanswered and a matrix is a queue with an org chart attached, and the escalation path becomes the real structure regardless of what the diagram says.

The most useful evidence in this whole field is also the most deflating for anyone who believes the chart is the answer. Bloom, Brynjolfsson, Foster, Jarmin, Patnaik, Saporta-Eksten and Van Reenen, publishing in the American Economic Review in 2019 on Census surveys of 35,000 U.S. manufacturing plants in 2010 and 2015, found that management practices account for more than 20 percent of the variation in productivity, a similar or greater share than R&D, ICT, or human capital. That is the case for taking management seriously. The next finding is the case for humility about structure: 40 percent of that variation occurs across plants within the same firm. Same company, same org chart, same policies, same incentives on paper, and enormous differences in how well the place is actually run.

Structure sets the ceiling. It does not deliver the result. A good structure with weak practice underperforms an awkward structure with strong practice, which is why so many reorganizations change the diagram, leave the practice untouched, and produce nothing.

How to buy the benefit without paying for the reorg

Almost every problem diagnosed as structural has a cheaper intervention that tests the same hypothesis first. Run the ladder from the bottom.

In ascending order of cost and irreversibility: change the metric a team is judged on; change who is in the room for a specific recurring decision; move one decision right down a level and write down that you have; name a single-threaded owner for one outcome and give them the authority to match; move one person; create or dissolve one team; and only then move boxes at scale. Each rung tests roughly the same belief, that the current arrangement is producing the wrong behavior, at a fraction of the cost, and each one is reversible in a week.

The diagnostic that beats intuition is to instrument the waiting rather than the work. Take your last ten significant decisions and, for each, account for the elapsed time in four buckets: waiting for a person, waiting for a scheduled meeting, waiting for information, and actually being worked on. Most organizations are shocked by how small the last bucket is. If the time is dominated by waiting for a person or a meeting, you have a decision-rights problem and a reorganization will not touch it, because the same bottleneck will reappear at the new coordinates. If the time is dominated by handoffs between two groups that always work together, you have a genuine structural problem and moving that boundary is the correct fix.

The second diagnostic is a count: how many separate teams must say yes before a typical customer-visible change ships? One or two means the structure matches the work. Five means the structure is taxing every unit of output you produce, and the tax compounds, because each additional approver adds not just their own delay but their own queue, their own priorities, and their own chance of asking for a change that restarts someone else's clock.

If the answer really is a reorganization, four things make the difference between the expensive version and the very expensive version. Set span deliberately per unit rather than as a company-wide target, with the routine work wide and the judgment-heavy work tight. Publish the tie-break rules on day one, in writing, including who resolves disputes between the new units. Spend your communication on why the shape changed and what decisions now sit where, because the chart is the least useful artefact you will produce and the only one anyone reads. And set a date, six months out, to measure the thing you claimed the reorganization would fix, using the wait-time diagnostic above, so that the next proposal has evidence behind it rather than a fresh diagram.

Put it to work

Take your last ten decisions and split the elapsed time four ways: waiting for a person, waiting for a meeting, waiting for information, actually being worked. Waiting on people is a decision-rights problem: name an owner, write the tie-break down, and skip the reorg. Before any headcount case goes forward, load salaries by 1.46, price the buildings before the severance, and add the 60-day WARN clock to the schedule.

Sources & references

Linked entries open the named source directly. Entries without a link say exactly what kind of reference they are — and how to check them yourself.

• U.S. Bureau of Labor Statistics — Occupational Employment and Wage Statistics (national, May 2025) — Employment and wage figures retrieved via the BLS public data API: series OEUN000000000000000000001/004 (all occupations) and OEUN000000000000011000001/004 (management occupations, SOC 11-0000); the first-line-supervisor total is the sum of the twenty first-line-supervisor SOC codes.
• U.S. Bureau of Labor Statistics — Employer Costs for Employee Compensation (civilian workers, Q1 2026) — Total compensation $49.32 per hour worked, wages and salaries $33.72, benefits $15.60; BLS API series CMU1010000000000D, CMU1020000000000D, CMU1030000000000D.
• Meta Platforms — Fourth Quarter and Full Year 2023 Results (restructuring charge composition, headcount)
• Meta Platforms — Fourth Quarter and Full Year 2025 Results, Exhibit 99.1 (SEC EDGAR)
• Nucor Corporation — Form 10-K for fiscal 2025 (decentralization; headquarters headcount)
• Amazon Web Services — Powering Innovation and Speed with Amazon's Two-Pizza Teams
• Bloom, Brynjolfsson, Foster, Jarmin, Patnaik, Saporta-Eksten & Van Reenen, "What Drives Differences in Management Practices?", American Economic Review 109(5), 2019, 1648–83
• Guthrie & Datta, "Dumb and Dumber: The Impact of Downsizing on Firm Performance as Moderated by Industry Conditions," Organization Science 19(1), 2008, 108–123
• 29 U.S.C. §2101 — WARN Act definitions (employer, plant closing, mass layoff thresholds)
• 29 U.S.C. §2102 — Notice required before plant closings and mass layoffs (60 days)
• Frederick P. Brooks Jr., The Mythical Man-Month (1975) — the n(n−1)/2 communication-path argument — Book; no stable public URL. The arithmetic quoted here is the pairwise-link formula, which is checkable directly.