Fewer layers is not more leverage.

Andy Grove, then the president of Intel, opened his 1983 book on management with breakfast. Imagine you are a waiter, as Grove was in college, and you must serve a three-minute soft-boiled egg, buttered toast and coffee, all fresh and hot at the same moment. Then imagine running a factory that does it all morning. Much of what Grove wanted to say about running an organization grows out of that small factory: its limiting step, its inventory, its indicators, and the people who keep it running. A few chapters in, he asks what a manager in such a factory actually produces, and gives an answer that fits on an index card. A manager's output is the output of the organization they run, plus the output of the neighboring organizations they influence.¹

That card is worth taking out this autumn, because many companies are redrawing their org charts with a different measure in mind. They are counting layers removed: how many levels of management came out, and how many people now report to each manager who remains. Grove's card counts something else: the organization's output, the work the whole unit delivers. A company can score well on the first and badly on the second, and the first is far easier to report.

My argument is that AI is taking over the part of management that moves information, and that this makes the rest of the job worth more, not less. Grove would have agreed that information has become cheaper to gather. But his test of a manager was never the ratio of reports to managers. It was output, and the activities he said produce the most of it are decisions, well-timed interventions and knowledge that only one person has. Those do not shrink when the information work goes. Each one now steers more output. The right response to AI is to rebuild management roles around that work, not to count the layers that came out as the result.

The year of the flatter chart

The wave is easy to document. In May, the Guardian reported that the cryptocurrency exchange Coinbase had laid off 14 percent of its workforce and said it would no longer have pure managers: managers would contribute code and other work directly, and would see their direct reports rise to 15 or more. After the payments company Block laid off 40 percent of its workers, some of its engineering managers were assigned as many as 175 direct reports, according to internal organization charts the Guardian reviewed. A former technical lead told the paper that managers there had typically had six to 12. A statement from Block's chief executive, Jack Dorsey, and board member Roelof Botha put the principle plainly: “There is no need for a permanent middle management layer.” The same report noted that Amazon's chief executive had set a goal of raising the ratio of employees to managers by at least 15 percent, and that Meta had moved to flatten its structure too. The four companies declined to comment.²

The trend is not confined to a few technology firms. The Guardian cited research from the workforce data company Revelio Labs finding that openings for middle-manager jobs in the United States at the end of 2025 were 42 percent below their 2022 peak.² Gartner's 2025 survey of chief executives and senior business executives found that more than half of CEOs said they would use AI within the next five years to de-layer middle management roles. Gartner also suggested that AI could let a single manager oversee approximately 20 percent more employees.³ And on October 8, Caixin reported that the chairman of the Chinese recruitment platform Zhaopin sees standardized jobs and middle-management positions already under pressure, as AI takes over routine coordination and information processing.⁴

Each of these is a statement about structure: fewer layers, wider spans, a higher ratio. None of them, on its own, is a statement about output. That is not a criticism of the companies. They may well be right about the work. It is a reminder that the number in the announcement is not the number Grove would have asked for.

What AI really takes: the information work

Grove sorted a manager's day into a handful of activities: gathering information, giving information, making decisions, nudging people, and being a role model, where he thought nothing mattered more than how a manager spends their own time. He kept a log of one of his own days to show the mix, and most of it was gathering and passing on information, with decisions and nudges threaded through.¹ Much of a manager's week, in his account, is spent being a channel.

That is the part AI is taking first. A Fortune commentary on October 8 described a daily operating meeting at CVS Health where agents now prepare the reports and dashboards that once consumed much of a two-hour session. The meeting has fallen to roughly 15 minutes and is now largely about exceptions, problems and decisions.⁵ At Block, according to the Guardian, AI is now primarily responsible for sharing information between managers, their reports and other teams, while people called directly responsible individuals set strategy and priorities and player-coaches look after employees' growth.²

Grove would not have been surprised. He wrote that automation is one way to raise the leverage of every kind of work, administrative work included.¹ If software can assemble the dashboard, the manager who used to assemble it has been given time. The question his book forces is what that time now produces.

What it leaves: the high-leverage work

Grove defined leverage as the output a particular activity generates, and he named three ways a manager's activity earns a lot of it. The first is when many people are affected by one manager, as when a finance manager defines in advance what every participant in the planning process must prepare. The second is when a brief, well-focused set of words or actions changes someone's behavior over a long period. The third is when a large group's work depends on a unique, key piece of knowledge that one person supplies, such as the engineer who sets prices that hundreds of salespeople will have to sell at.¹

None of these is information work, and none of them gets smaller when agents arrive. Each gets larger. When one team can produce much more with AI, a decision that sets its direction steers much more output. A short conversation that corrects how someone frames a problem now affects everything that person and their tools produce afterward. The knowledge of what good work looks like in a particular domain becomes the thing a large volume of generated work is checked against. Grove's three conditions were about how much output depends on one person's act. AI raises that dependence.

The same arithmetic works in reverse. Grove also wrote about negative leverage. A manager who waffles on a decision that others are waiting for stops their work, because no decision has the same effect as a no. A manager who meddles, taking over a subordinate's problem instead of letting them work it through, teaches the subordinate to bring every problem upward.¹ In an organization where each team produces more, an unclear decision idles more work, and meddling at scale has a wider blast radius. The Guardian quoted a consultant making a related point: if one team produces more with AI, the team that must approve its work can be overwhelmed.² Removing a layer does not remove that work. It moves it, often to someone who now has less time for it.

Span of control, rethought

Grove offered a rule of thumb for the size of a team. A manager whose work is mainly supervisory should have six to eight subordinates; three or four are too few and ten are too many. The range comes from his guideline that a manager should spend about half a day a week with each subordinate, enough to monitor without meddling.¹ He offered it as a rule of thumb, not a law, and a fair reader should not treat it as one. It does have one useful property: it states its assumption, so you can test what has to change for a much larger number to work.

Figure 2 does that test with deliberately simple numbers. Give a manager four days a week for supervision, which is roughly what Grove's six to eight reports at half a day each would use. Then suppose AI absorbs part of each report's half day, the part spent gathering and passing on information.

With nothing absorbed, the answer is Grove's: eight. With half the time absorbed, it is 16. To reach 175, AI would have to take over about 95 percent of the time each report needs, leaving roughly 11 minutes a week for everything else. That may be possible for some kinds of work. It is a claim about how little of the manager's half day was ever more than information, and it deserves to be stated and tested rather than implied by an org chart.

The Fortune commentary proposes breaking span of control into several measures: a human span of people managed, an agent span of digital workers directed, a workflow span of the bodies of work someone is accountable for, and what its authors call total productive leverage, the amount of work the combined system can produce.⁵ I think that is right, and the last measure is Grove's card in new words. But Grove adds a refinement that explains why no single number works.

He argued that the right management style depends on a subordinate's task-relevant maturity, an idea he credited to other researchers: a mix of drive, readiness to take responsibility, training and experience, specific to the task at hand. A seasoned sales manager moved to run a factory unit had low maturity in the new job and struggled until he learned it.¹ Low task-relevant maturity calls for closer, more structured supervision; high maturity calls for agreed objectives and lighter monitoring. And Grove was firm that delegating does not end responsibility. In his words, “delegation without follow-through is abdication.”¹

This is where the flattening wave is most exposed. A reorganization of this kind puts many people into new tasks at once: managers who now produce, individual contributors who now coach, teams learning to work alongside agents. Task-relevant maturity falls across the board just as spans widen. The Guardian's reporting on lost mentorship and fewer paths to promotion is what that collision looks like from below.² Grove's framework predicts that these are the months when follow-through matters most, and the structure has just removed the time for it.

The test for every layer

None of this is an argument for keeping layers. Grove spent much of his book attacking low-leverage management work, and a layer that exists mainly to collect and repackage information is exactly what he would have cut. If AI does that work, the layer should change or go.

The test is what replaces it. For each management role in the new design, ask which high-leverage activity it performs: which decisions it owns, whose behavior it is responsible for shaping, and what knowledge it supplies that no one else has. Then ask how the output of the organization it runs is measured. Grove's own advice on measurement applies: pair indicators, so that an effect and its counter-effect are both measured, as when inventory levels are watched alongside shortages.¹ A wider span that produces more work and more rework has not raised output.

A role that cannot answer the first question was probably a channel, and the channel is what AI replaces. A role that can answer it should be sized by what that work requires, including follow-through for people new to their tasks, not by a target ratio.

Objections, considered

Grove wrote for a 1983 semiconductor company. He did, and his span rule came from a time when gathering information took a manager's own hours. That is why Figure 2 lets AI absorb the information work. What does not date is the measure: a manager is judged by the output of what they run. Nothing about AI changes that.

Flatter organizations make decisions faster. Often they do, and fewer handoffs can raise output directly. That benefit is real and should be counted. But it is an output claim, and it should be measured as one, alongside the decisions that now wait longer because the person who makes them has 30 other people waiting too.

The new spans are working. Perhaps. The evidence so far is mostly announcements, internal charts and interviews, and the companies involved have said little in public about results. The Guardian's sources included people who left or were laid off, who may be biased, and one of them said as much.² That is a reason to measure output carefully in the next year, not to assume either answer.

Count what the organization produces, not the layers it no longer has.

Before you redraw the chart

• What does this role decide? Name the decisions it owns and who waits on them.
• Whose work does it shape? Name the people and teams whose behavior it is responsible for, and how often it can reach them.
• What does it know that no one else does? If the answer is nothing, it was a channel.
• How is its organization's output measured? Pair each measure with its counter-effect, such as volume with rework, before and after the change.
• Who follows through for people new to their tasks? Size that time explicitly; a reorganization creates many of them at once.

Back to the index card. It still says what it said in 1983: a manager's output is the output of the organization they run, plus the output of the neighbors they influence. Put it next to the new org chart. Every layer that came out removed some information work, and that is a gain. The card asks a different question of every box that remains: what does it produce, and how would we know?