There’s a diagram going round again. Pyramid on the left: CEO, VPs, Directors, Managers, ICs, with a red line struck through the middle two. Tidy stack of rectangles on the right: CEO, Team Leads, small autonomous teams. Underneath it, the exhibits. monday.com cutting a fifth of its workforce on 22 July. A founders’ letter about the organisation built for the last chapter not fitting the AI era. Gartner forecasting that one company in five will use AI to flatten, with over half of middle management as the target.
Good image. Clean, memorable, screenshot-friendly, and it’ll be in four hundred board decks by September. As a description of what happens when you actually do this, it’s nonsense.
McKinsey’s own read is that most companies have added at least one layer between the CEO and the front line over the past decade, and some have added two or three [1]. That’s expensive, it slows decisions down, and if your restructure is stripping out genuine accretion then fine. Do it properly, do it once, and be honest about what it is.
The diagram is making a much bigger claim. It says two whole layers are surplus, that AI soaks up what they did, and that what’s left is a CEO, some team leads and a contented field of autonomous pods. Every part of that is either a renaming exercise or a bill you’ve postponed.
Figure 1. Where most organisations actually are, the flattened structure the decks promise, and what results once leadership work is reassigned rather than removed.
Who leads the pods?
Look at the right-hand box. Team Leads sit between the CEO and a set of small autonomous teams. So who’s running the small autonomous teams?
Two answers, and neither one helps.
Maybe each pod has someone leading it. Setting priorities, holding the standard, doing the one-to-ones, hiring, sitting in calibration. Then you haven’t removed a layer, you’ve renamed one. The Team Leads above are your Directors and VPs with new business cards, and the people running the pods are the Managers you just crossed out in red. Nothing has flattened. The pyramid has been redrawn as rectangles, which is a graphic design decision.
Or the Team Leads really are spanning several teams while carrying delivery work of their own, and the pods self-organise. That’s the version the consultancies mean, and it’s the version that hollows people out. Someone who was managing eight people now has four teams, their own deliverables, and the coordination load that used to sit across two layers. Give it eighteen months. The good ones go. They take an IC role somewhere that pays about the same and asks less, or they leave to build the thing they’ve been turning over for two years. What stays behind is whoever had the fewest options, which isn’t the selection effect anyone had in mind when they called this a high-performance operating model.
Gallup’s span-of-control figure has gone from 8.2 direct reports in 2013 to 10.9 in 2024 to 12.1 in 2025 [2]. The job has grown by half in a decade, before anyone draws the After box. These models assume another big step up from there. I haven’t seen anyone show their working on what happens to quality at twenty.
Amazon already ran this experiment
Almost nothing in the After box is new, which nobody seems keen to mention. Bezos reorganised Amazon around small autonomous teams in 2003, and the idea has been rehearsed ever since through Spotify squads, agile pods, product trios and every operating model refresh in between [3].
But look at what Amazon built, because the deck has it upside down. What made two-pizza teams work wasn’t an absence of leadership. It was the single-threaded leader: one person, fully accountable for one initiative, no competing priorities, no other job. Amazon’s own line is that the surest way to fail at inventing something is to make it somebody’s part-time job [4]. That’s a model that adds dedicated leadership per unit of work. More leaders per initiative, each with a narrower remit.
The flattening diagram borrows the pod and throws away the mechanism. It gives you a Team Lead running four pods plus a delivery remit, which is the part-time-job failure Amazon designed the single-threaded leader to avoid.
Amazon was also honest about the costs. The model assumes teams will duplicate work and compete for resources, and treats that duplication as the price of speed. It needs heavy investment in the interfaces between teams, in written mechanisms, in goal alignment. Amazon has run it alongside a deep and explicit level structure, not instead of one. None of that survives into the After box, because the After box isn’t an operating model. It’s a cost story in an operating model’s clothes.
What happens when firms actually delayer
Here’s what bothers me most about all this. We are not short of research.
Julie Wulf’s work at Harvard Business School, published in California Management Review as “The Flattened Firm: Not as Advertised”, looked at exactly this across large US firms over twenty years [5]. Delayering has always been sold on the promise that removing layers pushes decisions down, nearer the customer, faster. Wulf and colleagues found something closer to the reverse. CEO span of control roughly doubled as firms flattened, from about five direct reports to about ten between the mid-eighties and the mid-noughties. On balance there was more centralisation to the CEO, not less. CEOs of flattened firms spent more of their time on internal interactions.
Now hold that against the After diagram. A CEO with a much wider span, spending more time inside the building, pulled into detail that used to get settled two levels down. That isn’t an autonomous-teams operating model, it’s a bottleneck.
And it bites hardest where you can least afford it. Take performance evaluation. Calibration isn’t admin. It’s how an organisation decides who’s good, who’s developing, who’s in trouble and where the money goes. It needs someone who has watched the work closely enough over a year to make a defensible call, and enough peers in the room with comparable visibility to argue with them. Take out two layers and you have senior leaders calibrating people they’ve seen second-hand, through dashboards and through the summary of someone stretched across four teams. That isn’t more efficient performance management, it’s less accurate performance management, which is a good deal worse than slow. And you’ve dragged your most expensive people into the most transactional version of their job for a quarter of the year.
Same for capacity management. Same for prioritisation across competing demands. Same for the unglamorous business of spotting that someone is struggling six weeks before it turns into a performance case. The diagram has no room for any of it, and prices all of it at zero.
Why the work looks like overhead
I think this is the root of the problem. Management work is largely illegible from above.
What’s visible from the top is the artefact. The status report, the roadmap, the headcount plan, the QBR slide. AI really can produce those faster, and if you think the artefact is the job then the job looks automatable. Zinman’s framing at monday.com pointed at projects that should have taken days taking months, at meetings and friction, and he isn’t wrong that this happens [6]. Most of us have sat in that meeting.
The artefact is the residue though. The work is absorbing ambiguity. Strategy lands at the top of a function as a direction. An engineer needs a task by Tuesday. Somebody converts one into the other, and the conversion involves judgement calls nobody writes down: what to drop, what to defend, whose development this stretch assignment actually serves, which of two plausible plans the team will really execute rather than merely nod at. It’s illegible because it’s being done well. When a manager is on top of this, the visible output is that things are calm, so the layer looks like overhead.
Gallup’s finding that managers account for around seventy per cent of the variance in team engagement has been quoted so often it’s stopped landing, so try it the other way round. If that’s even roughly right, then the management layer isn’t administrative cost sitting between the strategy and the work. It’s your biggest controllable variable in whether the work is any good. You wouldn’t describe your biggest controllable variable as a layer to eliminate. You’d describe it as something to redesign very carefully, with instrumentation, over several years.
The manager layer is already broken
The timing here is the strongest argument against doing any of this now.
Gallup’s 2026 State of the Global Workplace puts global engagement at twenty per cent, the lowest since 2020 and the first back-to-back annual fall they’ve recorded [7]. The attribution is what matters. Gallup pins the decline primarily on managers, whose engagement dropped from thirty-one per cent in 2022 to twenty-two per cent in 2025, five points of that in the last year alone. Individual contributor engagement barely moved. The gap between managers and their teams has narrowed from eleven points to three. Managers have essentially lost the engagement premium they’ve always carried.
That inverts the whole flattening story. The layer the decks write off as coordination overhead is the layer whose deterioration is dragging everyone else’s numbers down with it. The manager layer isn’t idle. It’s maxed out, and the consequences are showing up in everybody’s engagement scores.
Set that next to job satisfaction, which sits far higher, somewhere in the sixties to low seventies depending on whose survey you read. The gap is the whole story. People are reasonably content and psychologically absent at the same time, which is what a workforce looks like when it’s stuck rather than happy. Only about a quarter of US workers now reckon it’s a good time to find a decent job, against roughly seventy per cent in mid-2022.
So the proposal is to take the weakest link, cut half of it, give the survivors much wider spans and a delivery remit, and expect coordination quality to improve. If someone can walk me through the mechanism by which that raises engagement I’ll listen, but I haven’t heard it, and Gallup’s own read is that the problem is structural and about how organisations are led rather than about AI.
Three things the model doesn’t price
All three are talent intelligence problems rather than org design problems, which is presumably why the org design decks skip them.
Attrition, or rather the lack of it. The US quits rate was 1.9 per cent in May 2026, the lowest since August 2020, against a Great Resignation peak of 3.0 [8]. Mobility is frozen across most of the developed world. That cuts both ways. Damage from over-flattening will be invisible, because the people who’d have left in 2022 will stay and quietly check out in 2026, and your retention dashboard will look excellent while the layer rots underneath it. It also makes reversal expensive. These decisions carry an unspoken assumption that if you get it wrong you can hire the layer back. You can’t rehire experienced managers who know your business at any speed, and when mobility returns you’ll be trying to do it in the same window as everyone else who made the same call.
Succession. Manager headcount is already down around six per cent between 2022 and 2025 on Live Data Technologies figures, executive roles down close to five [9]. Korn Ferry has forty-one per cent of employees saying their organisation has cut management levels, and forty-three per cent saying leadership alignment got worse as a result [10]. DDI has around eighty per cent of organisations lacking confidence in their own leadership pipelines [11], and that was before this round started. The Director layer isn’t mainly a coordination mechanism. It’s where VPs come from. For most organisations it’s the only place anyone has ever grown someone capable of running a P&L, because it’s the only job combining scope, ambiguity and consequence at a survivable scale. Take it out and you haven’t saved money, you’ve converted an internal development cost into an external search fee payable in about four years, at a premium, against every competitor who made the same call.
McKinsey put a version of this on the table themselves. Krivkovich calls the junior development question the billion-dollar one: strip out every entry-level role and you end up with an expensive organisation of only senior people, and ten years on you’re missing the generation you needed [1]. The same logic runs one level up and nobody wants to say so. Remove the Manager and Director rungs and you haven’t only lost coordination capacity, you’ve demolished the training ground where judgement gets built.
Then the load transfer. No model I’ve seen shows where the work goes. It doesn’t evaporate. Prioritisation, capacity planning, conflict resolution, career conversations, hiring, onboarding, calibration, stakeholder management. It lands on the remaining leads, who burn out, or on the teams, who now spend a chunk of every week coordinating instead of building. If the answer is that the teams absorb it, say so, and put the productivity loss in the model next to the salary saving. It won’t net out the way the deck implies.
The research is more careful than the summary
Read the source material and there’s a real gap between what it says and what the slide made from it says.
McKinsey’s April 2026 discussion of the agentic organisation is heavily hedged on structure. Krivkovich says org charts are tricky and it’s too early to say what the answer looks like, and that it may differ by domain. On pods that form and reform, she says it’s very hard to do in practice and that most companies are a long way from reorganising around the principle. Her reason is the thing the flattening diagram deletes: “you need a job hierarchy so people can be evaluated” and know who to go to for support [1]. She also notes that more than eighty per cent of companies aren’t yet seeing any bottom-line impact from their AI investment.
So the research position is that nobody knows what the org chart looks like yet, pods are hard, hierarchy is still doing necessary work, and the returns aren’t in. The slide position is that you should delete two layers, and here’s a diagram. Those are different claims, and the second is being sold on the authority of the first.
Headcount is a bad number
Step back and ask what the diagram is optimising for, because the answer is a number that tells you very little.
Every one of these announcements is denominated in headcount. Twenty per cent. Six hundred and twenty people. More than half of middle management. Manager headcount down six per cent. It’s the unit of account for the entire debate and it’s about the crudest measure of organisational design available. Headcount is an input proxy. It tells you how many employment contracts you’re holding. It doesn’t tell you what it costs you to serve a customer, how much value any part of the organisation generates, how long a decision takes to travel from question to answer, or whether the work is any good.
It also moves independently of all of those. Cut headcount and raise cost to serve, which happens routinely when the work reappears as contractors, as an SI statement of work, as an offshore captive, as vendor spend, as rework. Cut headcount and lengthen decision latency, which is what Wulf found when decisions floated up to a busier CEO. Cut headcount and wreck value generation, and don’t find out for three years, because value generation gets measured annually and headcount gets measured weekly.
monday.com illustrates it cleanly. A fifth of the workforce came out, the operating margin outlook went from around thirteen per cent to around fifteen, revenue guidance held. A margin story delivered through a headcount lever. It may well work. But nobody published a cost-to-serve figure, or a revenue-per-employee trajectory, or a decision-latency measure, or any analysis of value per team, because those aren’t the numbers markets react to inside a quarter. The headcount number is the one everybody in the room already has.
Which is why it survives. Not because it’s a good measure but because it’s the only one that’s universally available, immediately comparable across companies, and legible to an investor who knows nothing about your business.
If speed is the goal, measure decision latency and count the approvals between a question and an answer. If cost is the goal, measure cost to serve, including vendors and compute. If value is the goal, measure contribution per team against what that team consumes. Any of those produces a different restructure to the one headcount produces, and usually a smaller and more surgical one. None of them fit on the slide.
The agents nobody is counting
This is where the argument comes apart on its own terms, using the case being made for flattening.
McKinsey’s premise is that AI gives leaders a more superhuman capacity to manage across bigger scopes, and that this is what makes a flatter structure viable [1]. Fine. Follow it through. If the surviving Team Lead is the decision and orchestration layer for a fleet of agents, their span hasn’t narrowed at all. It’s been redefined into something considerably larger than the org chart shows, and the org chart has nowhere to put it.
The scale isn’t hypothetical. IBM’s enterprise survey at Think 2026 has most large enterprises running a digital workforce of more than sixteen hundred agents by the end of this year [14]. Gartner projects a typical Fortune 500 operating over a hundred and fifty thousand agents by 2028, up from fewer than fifteen in 2025 [15]. Nearly every organisation OutSystems surveyed is already using agents somewhere, and ninety-four per cent are worried about sprawl [16].
Put the governance numbers alongside that. Seven in ten executives say their current AI governance isn’t fit for purpose. Around eighteen per cent keep a complete, current inventory of the agents already running inside their own walls. About twelve per cent have a centralised platform to manage them [14]. Gartner’s figure for organisations that believe they have the right governance is thirteen per cent [15].
So the honest After box has ten permanent people and something like two hundred agents attached to them, and none of the two hundred appear in the model. That workload is real: scoping and permissioning agents, monitoring output, catching drift, checking whether the thing still does what it was built to do, sorting out the case where one agent’s unverified output has become another’s input, keeping the audit trail, deciding when to retire something. Supervision, quality control and performance management of a workforce. Which is the exact category of work the flattening model has just declared surplus.
The second workforce is harder to manage than the first in some specific ways. No institutional memory across contexts. No ability to escalate ambiguity, so every ambiguous case bounces up to a human rather than getting resolved sideways by a colleague. It changes underneath you when a model version changes, so the evaluation work never finishes. And it produces confident output when it’s wrong, which makes review more expensive than reviewing a junior’s work, not less.
McKinsey’s own framing sharpens this. Their distinction is between humans in the loop and humans above the loop, where above the loop means agents run the core process end to end and the human contribution is judgement on top [1]. Judgement at volume is the most cognitively expensive work there is. It’s also the least automatable, by their own argument. The proposal is to increase the volume of judgement required per human while removing the layer that supplies most of the judgement capacity.
Figure 2. The same team lead, with the digital workforce added. The org chart counts the left-hand column and prices the right-hand column at nothing.
That’s the diagram I’d like to see a consultancy draw. Not a pyramid turning into rectangles, but two workforces stacked against one supervision layer, with an honest estimate of the hours.
An HR function, in three drawings
Abstractions are easy to argue with, so here are some numbers on a function most readers know from the inside.
A mid-sized HR organisation, sixty-five people, supporting around six thousand employees. A CHRO. Three VPs across talent acquisition, people partnering, and reward and operations. Nine directors under them covering TA, HRBP, reward, payroll, learning, people analytics and HR systems. Fourteen managers. Thirty-eight practitioners doing the recruiting, advising, analysis and payroll. Five layers, average span of about six.
Figure 3a. The starting point. Five layers, sixty-five people.
Apply the flattening model and the director and manager layers come out. The deck says what’s left is a CHRO, three team leads, and thirty practitioners in autonomous pods with AI soaking up the coordination.
I don’t buy the autonomous pod any more. It’s a leftover from the agile era, when the shared thing a team gathered round was a backlog and a standup. That isn’t what agentic working looks like in the functions where I can see it happening. The unit isn’t the pod. It’s the individual, and the individual has their own stack.
What actually forms is this. Everybody, CHRO down to reward analyst, owns an orchestration agent that plans and sequences their work, and beneath it four to six task agents doing the narrow things. Market mapping. Comp benchmarking. Case triage. Attrition signal. Payroll exceptions. Quality of hire. Those agents belong to that person. Not shared with a pod, and emphatically not reports. Nobody appraises them, nobody puts them in a succession plan, and no HR system has a field for them.
Figure 3b. What actually forms. The dashed outline is one person, not a pod. Every individual owns an orchestrator and a stack of task agents, and carries the orchestration and governance for it personally.
Look at what that does to the argument. The flattening case rests on removing management work. This structure spreads it to everybody. Every individual becomes a player-manager. They still do their craft, and on top of it they scope, sequence, monitor, correct, evaluate and eventually retire a small machine estate, and they carry the governance for it themselves. The reward analyst who has never managed anyone now runs five things that need direction, quality control, and a call on when they’ve drifted.
That’s not a lighter organisation. It’s an organisation where management load has been decentralised to people who were never selected for it, never trained in it, and whose job description doesn’t mention it. And the people who were selected and trained for exactly that are the ones the diagram just removed.
Two more things appear that had no equivalent before. Once everybody has a stack, one person’s agents need to talk to another’s, which means a cross-agent orchestration fabric: routing, handoffs, a registry, shared context, aggregation, a standing monitor. And you need common controls across every agent regardless of owner. Data licensing and permitted use. Adverse impact testing. EU AI Act high-risk classification. Required human approval points. Per-agent spend caps. Both are real senior jobs with named owners. Neither existed in the sixty-five-person version. Both land in the column of the ledger the restructure didn’t read.
Figure 3c. The honest version. Thirty-four humans, roughly a hundred and ninety-four agents, and two shared layers the before state never had.
Run the numbers and the saving evaporates. Thirty-four people against sixty-five looks like a forty-eight per cent cut, until you notice the pod leads have reappeared as managers with a new title, that two new shared functions need owning, and that supervision, evaluation and audit for something like a hundred and ninety-four agents has been distributed across those same thirty-four people with no hours attached to it anywhere. Five or six agents per human isn’t a guess; it’s roughly what functions running this properly already report, and it’ll rise.
These exact figures are illustrative. The shape isn’t. Flattening business cases stop at the headline number, and no talent intelligence function I know of is being asked to model figure 3c. That gap is where the argument should be happening.
Why this keeps happening
monday.com is instructive, and not as a criticism of the company, which may be doing something entirely sensible for its own circumstances. Watch the market reaction. The stock rose on the announcement, snapping a six-day losing streak and climbing further over the following days [12]. Operating margin guidance went from around thirteen per cent to around fifteen. Revenue guidance reaffirmed. Six hundred and twenty people left and the market said thank you.
That’s the mechanism, and it’s better said out loud than dressed up as an analytical output. Flattening is legible to investors in a way that almost nothing else in people strategy is. It converts an unmeasurable asset, the quality of your management layer, into a measurable saving on the cost line, on a timescale that fits inside a guidance cycle. The AI framing is what makes it sayable without sounding like a cost cut, which is why every letter insists it isn’t about cost.
If you run a talent intelligence function, this is your job. Not opposing flattening on principle, because some of it is warranted, but making the invisible side of the ledger legible enough to sit in the same conversation. Show what your management layer actually absorbs. Show where the bench really is, and what the internal supply of VP-capable people looks like at current spans and at proposed ones. Show what a comparable flattening did to competitors two years ago and what their leadership churn looks like now. Show the rehire cost and lead time if the answer turns out to be wrong. Ask the consultancy for the counterfactual and the reversal cost. If they can’t produce either, that isn’t an operating model, it’s a slide.
What gets me is the disrespect in these models. Not in a wounded sense, but analytically: an entire category of work has been valued at zero because it doesn’t photograph well. Prioritisation. Capacity management. People development. Team leadership. The judgement to know which of three fires actually matters. None of it appears on the diagram. Nobody weighed it and found it wanting. Nobody weighed it at all.
References
Krivkovich, A. and Rahilly, L. (2026) “AI is everywhere. The agentic organization isn’t yet”, The McKinsey Podcast, McKinsey & Company, 2 April. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/ai-is-everywhere-the-agentic-organization-isnt-yet
Gallup span of control data, reported in Rodriguez Constable, C. (2026) “Management cuts today are shaping a leadership shortage”, Forbes, 30 June. https://www.forbes.com/sites/cindyrodriguezconstable/2026/06/30/management-cuts-today-are-shaping-a-leadership-shortage/
Stone, B., Amazon Unbound, on the 2003 reorganisation around two-pizza teams and the introduction of single-threaded leaders.
Amazon Web Services, “Amazon’s two-pizza teams”, AWS Executive Insights. https://aws.amazon.com/executive-insights/content/amazon-two-pizza-team/
Wulf, J. (2012) “The Flattened Firm: Not as Advertised”, California Management Review, 55(1), pp. 5-23. Working paper: https://www.hbs.edu/ris/Publication%20Files/12-087_bc50bde2-3016-457a-9bee-dc988cb1056b.pdf. See also Rajan, R. and Wulf, J. (2006) “The Flattening Firm”, Review of Economics and Statistics, 88(4).
monday.com restructuring coverage, including the founders’ letter and organisational rationale, CIO, 22 July 2026. https://www.cio.com/article/4200330/monday-com-cuts-20-of-its-workforce-to-restructure-for-the-ai-era.html
Gallup (2026) State of the Global Workplace: 2026 Report. https://www.gallup.com/workplace/708071/global-employee-engagement-continues-decline.aspx. Manager engagement analysis summarised in UNLEASH, 19 May 2026. https://www.unleash.ai/strategy-and-leadership/analysis/gallups-state-of-the-global-workplace-2026-report-three-key-decisions-for-hr-leaders
US Bureau of Labor Statistics, Job Openings and Labor Turnover Survey, June 2026 release. https://www.bls.gov/news.release/pdf/jolts.pdf
Live Data Technologies, reported in The Wall Street Journal and summarised in Forbes, June 2026 (see reference 2).
Korn Ferry, Workforce 2025 survey, on eliminated management levels and leadership alignment.
DDI, Global Leadership Forecast, on organisational confidence in leadership pipelines.
monday.com market reaction, 22-24 July 2026. https://finance.yahoo.com/markets/stocks/articles/mndy-stock-snaps-six-day-155101451.html and https://www.investing.com/news/stock-market-news/mondaycom-stock-rises-after-announcing-workforce-reduction-93CH-4805140
Gartner (2024) “Top predictions for IT organizations and users in 2025 and beyond”, 22 October. https://www.gartner.com/en/newsroom/press-releases/2024-10-22-gartner-unveils-top-predictions-for-it-organizations-and-users-in-2025-and-beyond
IBM enterprise AI survey, presented at Think 2026, on agent fleet size and governance readiness. Summarised at https://beam.ai/agentic-insights/ibm-says-enterprises-will-run-1600-ai-agents-by-year-end-70-cant-govern-the-ones-they-have
Gartner agent sprawl projections, April 2026, reported at https://www.stmicro.net/blog/ai-agent-sprawl-what-growing-businesses-need-to-know/ and https://news.sap.com/2026/08/agent-sprawl-why-ai-governance-is-now-board-level-issue/
OutSystems (2026) State of AI Development 2026, on agent adoption and sprawl concern. https://www.businesswire.com/news/home/20260407749542/en/Agentic-AI-Goes-Mainstream-in-the-Enterprise-but-94-Raise-Concern-About-Sprawl-OutSystems-Res