What We Lose When the Middle Goes

On Tuesday, we wrote about Amazon’s tokenmaxxing problem, engineers gaming AI usage metrics because the measurement system rewards token volume rather than value creation. We argued that fixing it required harder management, not better dashboards. What we didn’t address is that companies are systematically eliminating the managers who would do that work.

The numbers from Q1 2026 are stark. Amazon has cut roughly 30,000 jobs since October, explicitly citing a move to flatten management layers. Meta has begun layoffs targeting an estimated 20 percent of its workforce. Atlassian eliminated 10 percent. Citigroup expects to reduce headcount by 20,000 in 2026, attributing much of it to AI handling middle-office work. Coinbase CEO Brian Armstrong has spoken publicly about eliminating “coordination tax” and is experimenting with structures that collapse engineer, designer, and product manager roles into single positions. Gartner projects that 20 percent of organizations could eliminate more than half their middle management positions by year-end.

The logic is straightforward. The four largest hyperscalers are committing roughly $700 billion in AI infrastructure capital this year. Human salaries are the only cost line flexible enough to partially offset that buildout on the necessary timeline. Coordination, status reporting, and routine project management—historically middle manager work—are exactly what agentic AI does well. Cutting the layer looks like an obvious win.

It’s a long-term mistake, and the research on what middle managers actually do explains why.

McKinsey’s multi-year analysis of manager behavior found that organizations whose middle managers operate in the top quartile of coaching, talent development, and personal ownership practices generate total shareholder returns 3 to 21 times higher over 5 years than those in the bottom three quartiles. That spread doesn’t show up in any quarterly cost-out exercise. Gallup’s research on employee engagement, replicated across many cohorts, finds that managers account for roughly 70 percent of the variance in team engagement scores. Google’s own Project Oxygen, the internal study that informed how the company built its management practices, ranked coaching as the most important behavior of effective managers, above technical expertise.

None of this is what gets eliminated when companies flatten. What gets eliminated is the cost line. The capabilities behind it are assumed to migrate elsewhere, to AI agents, to senior individual contributors with 15 direct reports, or to peer collaboration. We’ve watched this assumption play out at scale enough times to know what comes next.

The three things that break:

• The first thing to break is coaching. Coaching is a part of management that requires human judgment, sustained relationships, and the willingness to have hard conversations about specific work. AI agents can’t do it. A senior engineer with 15 direct reports can’t do it either, because the math doesn’t work. A real coaching conversation requires roughly 30 minutes of preparation, 45 minutes of conversation, and 15 minutes of documented follow-up per direct report per cycle. At 15 reports, that’s a full work week per cycle. Senior engineers cut to player-coach roles won’t do it. They’ll prioritize the player half because that’s what gets measured.
• Change implementation collapses next. Multiple practitioner surveys and academic reviews, including Hughes’s 2011 analysis in the Journal of Change Management, find that most organizational change initiatives fall short of their stated objectives, with execution at the middle management level routinely identified as a primary failure point. The reverse implication is the part organizations forget. Engaged middle managers are the layer that makes change actually happen. The same layer that translates senior strategy into team-level behavior is the layer being cut to fund AI strategy. The capability needed to implement the change is being eliminated to fund the change.
• Then institutional knowledge fragments. Middle managers carry the unwritten context of how decisions actually get made, which exceptions are real, which deadlines are negotiable, and which cross-team dependencies have historical complications. None of this lives in documentation. Recent research on knowledge transfer in software development organizations finds that, even with formal documentation systems, the knowledge that matters most for daily work tends to live in human heads and walk out the door during layoffs. Trevor and Nyberg’s peer-reviewed study in the Academy of Management Journal found that downsizing of just 1 percent of a workforce is associated with a 31 percent increase in voluntary turnover the following year, and the highest-performing employees leave at disproportionately higher rates — concentrating the loss in exactly the senior individual contributors organizations are now relying on to absorb the work.

Here’s where we expect this to land in 18 to 24 months.

The companies executing the deepest middle management cuts will report short-term productivity gains, mostly driven by AI tool adoption among the remaining workforce. Their cost structures will look strong. Their first-year retention numbers will look acceptable given the soft labor market. Beneath those numbers, three patterns will compound. Coaching frequency will collapse, predicting performance decline. Senior individual contributors will quit at elevated rates, predicting institutional knowledge loss. And the AI tools the cuts were meant to fund will fail to generate the value modeled in capex justifications, because the layer responsible for translating tool capability into team behavior will have been removed.

This brings us back to where Tuesday’s essay ended. Amazon’s tokenmaxxing problem exists because employees calibrate to the measurement system that has consequences. That’s a manager-shaped problem. It requires someone with judgment, context, and a coaching relationship to ask developers what problems they actually solved with AI last week, and to evaluate whether the answer is any good. That work doesn’t scale through a dashboard. It scales through a layer of well-trained, properly empowered middle managers. The same layer that’s being cut to fund the technology that was supposed to replace them.

The capability we’re eliminating is the capability we’d need to make the technology we’re buying actually work. Companies that figure this out will compound their advantage over those that don’t.

This work synthesizes research from peer-reviewed work by Trevor and Nyberg (2008) and Hughes (2011), foundational theory from Argyris and Schön (1974), and institutional studies from McKinsey, Gallup, and Google’s Project Oxygen. Current layoff data is drawn from Q1 2026 reporting in the Financial Times, CNBC, and InformationWeek.

Contact us for a complete list of works cited.