The middle manager is having the worst decade of any job title in the corporate lexicon. US employers were posting 42% fewer middle-management vacancies at the end of 2024 than in spring 2022, according to Deloitte's 2025 Human Capital Trends, which also relays Gartner's prediction that by the end of 2026, 20% of organizations will have used AI to flatten their structures, eliminating more than half of their middle-management positions. The people still in the role can feel the ground moving: Gallup's 2025 State of the Global Workplace found manager engagement fell from 27% to 22% in a single year, the steepest drop of any group it tracks.
The economic logic sounds airtight. If AI can aggregate status reports, draft reviews, and route information, why pay a layer of humans to do it? But the logic contains a hidden assumption: that what middle managers do is what their job descriptions say they do. The evidence says otherwise, and in that gap lives a productivity paradox.
What the middle actually holds
Deloitte's survey of nearly 10,000 leaders found managers spending almost 40% of their time on firefighting and administrative work, and only 13% developing their people. Read cynically, that says the job is mostly waste. Read carefully, it says something else: the middle of the organization is where unplanned problems go to get absorbed. Firefighting is not a distraction from the middle manager's real job. In most firms it is the real job: translating ambiguous directives into workable plans, brokering between functions that do not share incentives, catching the exceptions no process anticipated.
AI genuinely can take over much of the reporting and administrative layer. What it cannot yet absorb is the exception handling and the political translation, and those do not disappear when the layer is cut. They flow somewhere: upward to executives who lack the context, sideways to senior individual contributors who lack the authority, or nowhere, which is how small problems become large ones. The companies now announcing delayering as a productivity measure are, in effect, betting that this invisible work either does not exist or will handle itself.
The evidence for the alternative bet
There is a better-grounded version of flattening, and it predates AI. Aghion, Bloom, Sadun, Van Reenen and colleagues, studying firms across ten OECD countries plus US administrative data, found that companies that had delegated more decision-making power to local plant managers before the Great Recession outperformed centralized rivals in the sectors hit hardest. The mechanism was the rising value of local information under turbulence: when conditions change fast, the people closest to the work know things headquarters cannot. The authors are careful to note decentralization is not universally optimal; it pays specifically in bad times, which is to say, in exactly the volatile conditions AI adoption is creating.
That finding reframes the choice. Cutting the middle while keeping decisions at the top concentrates the firm's judgment at the point farthest from its information. Cutting the middle while pushing decisions down converts the savings into speed. Same headcount reduction, opposite organizational designs.
There is also a version of this from the practitioner's side, and it is older than the current wave. In Brazil's 1990 crisis, as Ricardo Semler has told it, Semco faced the classic layoff moment and did something structurally different: it put the numbers on the wall and let workers decide, and they chose a pay cut in exchange for control over spending and a larger share of profit in the recovery. The recovery came. The point is not the heroics; it is the mechanism. The company did not delete coordination, it relocated coordination to the people with the most information and the most at stake, and gave them the transparency to do it. Families do not open the books, Semler says; companies that respect you do.
The test before the cut
Gallup's data offers one more warning to firms treating flattening as pure subtraction: in organizations it classifies as best-practice, 79% of managers are engaged, nearly four times the global average of 22%. The dispersion suggests managerial capacity is not a fixed tax on the org chart; it is a variable that firms produce, well or badly. Deleting a badly designed layer and deleting the function are different acts.
So the test, before any delayering, is not "what does this layer cost?" but "where will each thing this layer does actually go?" Exceptions need an owner. Cross-functional conflicts need a broker. People still need someone accountable for their development, the 13% of the old job that was arguably the only part AI cannot do at all. If the plan names destinations for those, flattening is a redesign and will likely pay. If it does not, the savings are real, the costs are merely deferred, and the paradox will resolve itself the expensive way: in the second year, not the first.
Sources
- Deloitte (2025). Global Human Capital Trends. Managers spend ~40% of time on firefighting/admin and 13% developing people; US middle-management postings down 42% end-2024 vs spring 2022; relays Gartner flattening prediction. https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2025.html
- Gallup (2025). State of the Global Workplace. Manager engagement fell 27% to 22% in one year; best-practice orgs show 79% manager engagement vs 22% global average. https://www.gallup.com/workplace/659610/state-global-workplace.aspx
- Aghion, Bloom, Lucking, Sadun & Van Reenen (2021). Turbulence, Firm Decentralization, and Growth in Bad Times. AEJ: Applied Economics. Firms that delegated to local managers before the Great Recession outperformed centralized firms in hardest-hit sectors. https://www.aeaweb.org/articles?id=10.1257%2Fapp.20180752