Middle Management Is a Coordination Layer. Agents Eat It.

Pull up the calendar of any middle manager and look at what actually fills it. Standups to collect status. Decks that roll the team’s numbers up a level. Sync meetings to reconcile what one group is doing against what another group thinks it’s doing. Escalations routed up, priorities routed down. Strip away the title and the org-chart box, and the job is mostly a data pipeline made of meetings — pull from the front line, normalize, flag what looks off, push a summary upward.

That pipeline is exactly the shape of work an agent does well. Which is the uncomfortable claim at the center of this post, seeded by Peter Diamandis’s Organizational Singularity conversation (EP #258): middle management in existing companies is “almost completely doing coordination,” and once agents take over the aggregation loop, “that function drops about 90%.” The load-bearing question for anyone building or running an org: how much of that 90% is real, and what does the honest version of the number look like?

The function was already drifting toward coordination #

Start with the part that doesn’t depend on any AI forecast. The manager-as-coordinator wasn’t invented by GPT — it’s where the role had been heading for a decade.

A Harvard Business School linguistic analysis of 34 million job postings from 2007 to 2021 found that supervisory language in manager descriptions fell 23%, while collaborative and coordination language tripled (HBS Working Knowledge). Read that as a quiet redefinition: the modern manager’s value proposition had already shifted from command authority to information brokerage and cross-functional routing. Companies spent fifteen years rewriting the job toward the precise activity that agents are now good at.

That matters for the thesis. You don’t have to argue agents will replace judgment or leadership — only that they can replace routing, and the org chart already told us routing is what the box is for.

The mechanism is specific, not hand-wavy #

The reason this is a sharper claim than generic “AI takes jobs” hype is that the replacement target is a concrete loop, not a fuzzy role. Agents pull status from operational systems, normalize it, flag anomalies, and push dashboard-ready summaries upward. That is the manager’s week, described as a cron job.

There’s early measurement on the underlying shift. A study of more than 187,000 software developers given AI copilots found they cut project-management and administrative tasks by 25% while increasing core output by 12% — and, strikingly, worked with an average of 5 collaborators instead of 22 (HBS Working Knowledge on the GitHub developer study). That last number is the tell: a roughly 79% reduction in collaboration surface per person. Coordination overhead isn’t just getting faster, it’s getting smaller — fewer humans in the loop for a given unit of work. When the loop shrinks, the layer whose job is to manage it shrinks with it.

McKinsey’s macro view points the same way: generative AI could push the share of automatable U.S. work hours to roughly 30% by 2030, up from 21.5% without it, with reporting and coordination tasks among the most exposed categories (McKinsey Global Institute, 2023).

The number is already showing up in headcount #

This isn’t purely prospective. Middle managers were roughly a third of all white-collar layoffs in 2023, and a Korn Ferry survey found 41% of professionals reporting their company had specifically cut manager-level roles (Ramp’s “pure manager layoffs” analysis). The layer is being thinned ahead of the rest of the org — what you’d expect if its work were the most substitutable.

Gartner put a forward number on it: by 2026, 20% of organizations will use AI to eliminate more than half of current middle-management positions, framing AI as a driver of structural flattening — collapsing the coordination layer — rather than task-level speedup (Gartner’s October 2024 top-predictions press release). Note the scope: a fifth of orgs cutting half the layer is a meaningful figure, but it is a long way from a blanket “90% of all middle management evaporates.” Hold that gap; the counter-section lives in it.

Coinbase: a real example, with an asterisk #

The cleanest public attempt to operationalize the thesis is Coinbase. In 2025, CEO Brian Armstrong tied a 14% workforce cut (about 660 people) to an explicit removal of an organizational tier: “pure managers” — people whose job was to coordinate, report, and relay — were replaced by a flat structure capped at five layers below the CEO, with remaining leads expected to be “player-coaches who both lead and build.” Teams were reorganized into AI-native pods, some run by a single person, with AI absorbing compliance monitoring, status reporting, and internal-ops work (The Next Web). On its face: a named company, a named CEO, an explicit deletion of the coordination layer, and a stated mechanism.

The asterisk is large enough that you should treat Coinbase as a real-but-contested data point, not proof. The same reporting argues Coinbase “AI-washed” a cut that was economically forced: revenue down 26%, consumer transaction revenue off 45%, trading volumes at 18-month lows. And 59% of hiring managers in one cited survey admitted emphasizing AI in layoff announcements because, in the article’s framing, it “plays better with stakeholders.” So the restructuring genuinely flattened the management tier — and it happened during a financial contraction that would have demanded cuts regardless. Both are true. Coinbase shows the org-design move is being made in public; it does not isolate AI as the cause.

The honest counter-case #

If you’re building on this thesis, the failure modes matter more than the headline. Four of them are well-supported.

Coordination isn’t purely mechanical. Even the HBS work that documents the drift away from supervision finds that collaborative skill — knowing when to connect disparate groups at an inflection point — grew in value (HBS). Knowing which signal matters and when to escalate is a different thing from relaying the signal. Agents do the relay. The judgment about which relay matters is exactly the residue that doesn’t compress to zero.

Cutting the layer can quietly kill innovation capacity. Alloy Partners argues middle managers historically sheltered cross-functional innovation through budget cycles — “bad innovation capacity is still capacity.” Flatten the layer without replacing its informal sponsorship and you may end up running today’s workflows more efficiently while losing the ability to originate new ones (Alloy Partners).

Klarna ran the aggressive version and walked it back. Klarna cut its workforce from about 5,500 to under 3,500 citing AI automation, including replacing 700 customer-service agents. Within six months, customer satisfaction dropped and the company began rehiring. CEO Sebastian Siemiatkowski’s widely quoted line — that the company “went too far” — comes from his interviews reported in the business press (Bloomberg and others), not from the broader AI-and-work piece in Fast Company that covers the episode. The substance holds regardless of which quote you anchor on: the 90% figure is a theoretical ceiling under ideal conditions, and the real transition carries quality degradation and rollback costs when human judgment is stripped too fast.

Span-of-control limits are real, and dashboards don’t carry legitimacy. After manager-layer cuts, average span of control rose only from 10.9 to 12.1 direct reports between 2024 and 2025 — still short of any AI-augmented ceiling — while 37% of surviving employees reported feeling “directionless” and 43% said leaders “lacked alignment” (Ramp). Even where the agent handles aggregation, a human in the relay loop appears to supply legitimacy and motivational signal that a dashboard does not.

And the firm that issued the bullish forecast has since added the caveat. A May 2026 Gartner study found that autonomous-business and AI layoffs “may create budget room, but do not deliver returns” — separating raw efficiency from the organizational damage of abrupt management removal (Gartner press release; for readers who hit Gartner’s paywall, Fortune’s coverage summarizes it). The same source can be right that the layer is structurally exposed and right that gutting it for cost reasons backfires.

What to actually do with this #

The 90% is best read as a statement about the function, not a payroll target. The coordination work — pull, normalize, flag, summarize, route — really is collapsing toward agents, and the evidence for that is already in the job descriptions, the developer studies, and the layoff mix. The mistake is treating “90% of the coordination function is automatable” as if it meant “fire 90% of your managers next quarter,” which is the move Klarna ran and reversed.

The builder’s version is narrower and more useful. Identify the coordination loops in your org that are genuinely mechanical and hand them to agents now — that capacity is real and compounding. Then protect the residue the number hides: the judgment about which signals matter, the informal sponsorship that funds new bets, the human presence that makes a direction feel legitimate. Cut the relay. Keep the people who know what’s worth relaying. The org that gets the split right runs lean. The one that confuses the function with the headcount pays severance twice.