Middle management is back in the crosshairs. Gartner predicts that by 2026 one organisation in five will use AI to flatten its structure, eliminating more than half of current middle management positions (Gartner, 2024). The logic seems linear: if AI routes information, compiles reports and monitors performance, the layer that did those things becomes pure cost. The thesis of this article is that the logic is linear but incomplete — and that organisations cutting managers without refounding the role are removing the organ that produces their executives and holds their context.
What AI actually automates
Break down a middle manager’s work. One part is informational coordination: collecting progress updates, consolidating reports, distributing priorities, chasing deadlines. This part AI absorbs, and rightly so: it was administrative work dressed up as management.
But there is a second part, less visible in process maps and absent from organisation charts: translating strategy into local decisions, filtering upwards the signals that matter, holding together people who would not choose one another, training newcomers on the job. That part is not coordination: it is judgement applied to context. And context is precisely what general-purpose systems do not have.
The evidence on skills demand confirms the distinction: in the occupations most exposed to AI, the capabilities most in demand are exactly the managerial and business ones — managing projects and people, coordinating resources, operational decision-making (OECD, 2024). AI is not making management redundant: it is separating management from the office work that was suffocating it.
Cutting a manager because AI writes his reports is like dismissing a doctor because the machine takes blood pressure.
The hidden cost of flattening
Organisations that flatten aggressively discover three bills no business case had included.
The bottleneck moves up. The decisions the managers absorbed do not disappear: they rise. Executives with twice the direct reports become the organisation’s tail-end queue — and Gartner’s own prediction lists managers overwhelmed by the widened span of control among the risks.
The pipeline goes dark. Executives are not born executives: they are formed in middle management, making mistakes at scales where the error is recoverable. Removing the intermediate layer means the next promotion is from individual contributor to executive, with no training ground in between. It is the same formative debt we describe in our analysis of hybrid skills: visible only when it is too late.
The memory scatters. Managers hold the unwritten knowledge — why that client left, why that process carries that exception, why the shortcut that looks obvious was tried and abandoned twice. AI systems know the documents; managers know what the documents do not say. That knowledge appears in no restructuring business case, because it was never inventoried — and its worth is discovered only when it is needed and gone.
The span of control test
The argument for flattening runs like this: with AI absorbing coordination, each manager can cover more people, so fewer managers are needed. The first half is true. The second contains a logical leap worth illuminating.
A sustainable span of control does not depend only on how much coordination is needed: it depends on how much managerial attention is needed — difficult conversations, developing people, local decisions, handling exceptions. AI reduces the first load, not the second. If the organisation doubles direct reports on the assumption that attention scales like coordination, the result is not efficiency: it is nominal management — people who have a boss on paper and none in practice.
The equilibrium point exists, but it must be computed on the real work: what share of your managers’ time was automatable coordination? In functions where it was eighty per cent, the span can grow considerably. Where it was thirty — because the bulk was people management and decisions — flattening cuts muscle, not fat. The company-wide average is the wrong statistic: the decision must be taken function by function.
Refound, do not defend
Defending the status quo would be the opposite error. An intermediate layer that only passes information along is indefensible, and AI has made that obvious. Serious refounding rewrites the role around what remains scarce.
The post-automation manager has three mandates. First: guarding the quality of the decisions delegated to systems within their perimeter — the point where the boundary of algorithmic delegation stops being policy and becomes daily practice. Second: developing people, now that the administrative time has been returned. Third: acting as a two-way sensor — carrying context upwards and meaning downwards, the one function no dashboard has ever performed.
With mandates like these, the span of control can indeed grow: not because the manager does the same things faster, but because he does different things.
Designing the transition
For those who choose refounding, three concrete workstreams.
Redefine the job in writing. The new mandate — guarding delegated decisions, developing people, two-way sensing — must be written into the appraisal systems, not announced at town halls. If a manager’s performance is still measured on reports delivered, the organisation has changed the conversation and not the work.
Retrain on judgement, not on tools. The useful training for post-automation managers is in the skills AI does not replicate: running critical reviews of automated outputs, managing escalation, developing their teams. Yet another platform course prepares them for the work that is disappearing.
Select honestly. Not every current manager will inhabit the new role well: someone excellent at coordination may not be at judgement, and vice versa. A serious transition assesses people against the future mandate and treats with respect those who do not find themselves in it — the alternative, keeping everyone and changing only the title, produces the old structure with the new rhetoric.
What this means for decision-makers
Before signing a flattening plan, three checks. Where will the decisions this layer currently absorbs go — to an executive already at capacity, or to a system nobody reviews? Who will form the next generation of executives, and in what training ground? How much of the plan is organisational strategy, and how much is a cost cut that AI makes presentable?
A final caution on communication, because Gartner’s prediction circulates in corridors too: your managers have read it. An organisation that stays silent while everyone reads “half of positions eliminated” gets what silence always produces — the best leave first, because they are the ones with alternatives. If your choice is refounding, declare it early and with facts: written mandates, funded training, visible paths. Middle managers’ trust is an asset that is lost exactly once.
If the answers hold, flattening is a legitimate choice. If they do not, today’s saving is tomorrow’s bill — with interest. Organisational structure is a decision for the top, not a by-product of the budget: it is matter for an AI strategy worthy of the name, not for a spreadsheet.
And it is worth remembering where the prejudice comes from: middle management has been the favourite target of every restructuring wave for decades, because it is numerous enough to produce visible savings and quiet enough not to defend itself in public. AI did not invent this temptation — it merely gave it a new vocabulary. Distinguishing real transformation from a disguised cut is, once again, a decision that measures the quality of whoever takes it.