Bayer, Amazon, Meta, Oracle — the biggest names in the economy are stripping out management layers and widening spans of control. The savings are obvious. The thing they're cutting is not.
Middle management has become the most fashionable thing to cut. The logic is clean enough to fit on a slide: AI now does much of what middle managers did — synthesizing information, coordinating across teams, drafting the updates, summarizing the decisions — so remove the layer, widen everyone's span of control, and bank a faster, cheaper org. Bayer, Amazon, and Meta have all made public commitments to flattening. Oracle cut roughly 6% of its workforce in a single April day, with more expected. The trend even has a name now: the Great Flattening.
The numbers are not small. Gartner has suggested more than half of middle-management roles could be eliminated by AI-driven restructuring, and surveys already find 41% of employees globally — 44% in the US — saying their organization has removed managerial levels. Average team size per manager is creeping up as companies consolidate. On paper, it reads as overdue de-bureaucratization. Some of it is. But "the manager was overhead" and "the manager was doing invisible work" can both be true, and the flattening wave is being justified almost entirely by the first.
What the layer was actually holding
A middle manager's job description undersells the job. The legible parts — approvals, reports, status meetings — are exactly the parts AI can absorb, which is why they're the parts cited in the business case. The illegible parts are where the value hid: translating strategy into something a team can act on, catching the problem before it became a crisis, developing the junior employee into a capable one, holding the institutional memory of why things are done the way they are.
None of that shows up in a span-of-control spreadsheet, which is precisely why it's vulnerable. You don't miss the coordination layer when you cut it. You miss it three quarters later, when decisions are slower because nobody owns the cross-team seam, when a good performer leaves because no one was developing them, when the same mistake recurs because the person who remembered the last time is gone. Removing managers saves salary immediately and introduces communication gaps, decision delays, and lost knowledge on a lag — the classic shape of a cost that's easy to book and hard to see.
The parts of a manager's job that AI can do are the parts you could see. The parts you couldn't see are the parts you're about to find out you needed.
The span-of-control problem nobody priced
Widening spans is the mechanism of flattening, and there's a human limit it runs into. A span of around six was long considered the upper bound of how many people one manager could genuinely hold in their head — know their work, their development, their state. Spans of 14, 50, or 90 don't extend that capacity; they break it. The manager who technically has ninety reports isn't managing ninety people. They're managing a queue and hoping nothing urgent is buried in it.
The bet underneath the flattening is that AI tools absorb the coordination load that wider spans create. Sometimes they will. But AI can summarize a status update; it can't notice that a normally reliable person has gone quiet, or have the difficult conversation, or decide which of two good options fits the team's actual situation. Stretch spans past the human limit and assume software covers the difference, and you've quietly removed the layer where people were noticed — without replacing the noticing.
Visual 1 — The flattening ledger
What flattening removes | Shows up immediately | Shows up later |
|---|---|---|
Manager salaries | Lower cost (visible win) | — |
Approval/reporting layer | Faster decisions (real gain) | — |
Cross-team coordination | Looks fine at first | Slower decisions, dropped seams |
People development | Unnoticed | Weaker bench, higher attrition |
Institutional memory | Unnoticed | Repeated mistakes, slower onboarding |
How to read it: the business case lives in the top two rows. The risk lives in the bottom three — all on a lag, which is exactly why they're underweighted in the decision.
The contrarian half nobody says out loud
Here's what makes this genuinely hard rather than just cautionary: a lot of middle management was bloat, and pretending otherwise is its own failure. Layers of managers managing managers, meetings that existed to feed other meetings, coordination overhead that existed because the org was badly designed — that deserved cutting, and AI removing the busywork is a real gain. The honest position isn't "don't flatten." It's that flattening is a scalpel being used as a wood-chipper. The same wave that removes genuine bloat is removing genuine capability, because the business case can't tell them apart — it only counts the salary, and a great developer of people and a useless layer of bureaucracy cost roughly the same.
What this means for leaders
Cut roles, not headcount. Decide which specific functions a manager performed are bloat, which AI can truly absorb, and which were load-bearing — and protect the third category by name. A flat percentage cut across a layer guarantees you lose the good with the redundant.
Don't let AI "cover" the human parts on paper. Be honest about what software actually replaces. It handles synthesis and coordination admins; it does not develop people, hold hard conversations, or exercise situational judgment. If wider spans assume it does, you've created a gap, not an efficiency.
Watch the lagging indicators, not the cost line. The savings book instantly; the damage surfaces in attrition, decision speed, and repeated errors a few quarters out. Track those deliberately after a flattening, because they're the only signal that will tell you whether you trimmed fat or cut into muscle — and by the time they show, reversing is expensive.
Flattening will keep accelerating because the upside is immediate and legible and the downside is delayed and diffuse — the exact profile of a decision organizations are worst at making well. The companies that come through it strong won't be the ones that cut the most layers. They'll be the ones that knew which layer was holding something up.
A BusinessInfomatics original. Drawn from 2026 reporting on corporate delayering (Bayer, Amazon, Meta, Oracle), Gartner estimates on middle-management reduction, and workforce data on span-of-control and flattening.
