The Real Work Wasn't Measured

On May 7, 2026, Matthew Prince published a memo. Cloudflare had just posted $639.8 million in quarterly revenue, beating consensus by close to $18 million. The stock was fine. The pipeline was fine. He announced 1,100 layoffs, more than 20 percent of the company, and he explained the cuts by citing Peter Drucker.

The memo said internal AI adoption had grown more than 600 percent in three months. Prince drew a distinction between makers and measurers, and said the measurers were the ones going. Middle managers. Finance. Internal audit. Compliance. Legal ops. Revenue recognition. Any role whose main verb was "check on."

Two weeks later, Fortune ran the story under the headline that an entire category of workers was now obsolete. Two weeks after that, Oracle disclosed that its workforce had shrunk by 21,000 people over the fiscal year ending May 31, a thirteen percent reduction, tied explicitly to AI adoption and booked as a $1.8 billion restructuring charge. In July, Amazon confirmed a fresh 14,000-role cut concentrated in middle management, framed as "reducing layers, increasing ownership, and removing bureaucracy." By mid-August, reporting emerged that Oracle was preparing a second round of cuts ahead of its Q2 fiscal start on September 1.

None of these companies is struggling. Cloudflare beat the quarter. Oracle's cloud infrastructure segment grew 77 percent year over year. Amazon's AWS and ads numbers are the strongest in the company's history. The cuts are strategic. They are AI-adjacent. They come out of stated org-design theses. And they are, in every case, aimed at the same layer.

That layer is the one every executive team is now told to compress. The 2026 case for cutting it sounds like a benchmark that finally paid off. It is worth reading carefully, because most of the argument is right, and the part that is wrong is going to be very expensive.

The layer that stopped defending itself

The Gartner projection making the rounds says that by year-end 2026, more than half of middle-management roles in the affected cohort will be gone. An SSRN paper from Guohou Shan and Feng Zhu at Harvard, published in June, shows that firms with high AI exposure are systematically reducing hierarchical layers and widening spans of control. Their measured average span jumped from 8.1 in 2013 to 12.1 in 2025, projected to hit roughly 25 by 2028. That is Prince's math made general.

Prince also cited Drucker, and this is the part everyone quoted. The passage Cloudflare used is from The Practice of Management (1954), the book where Drucker actually invented the profession of the executive as a specific kind of knowledge worker. The line the memo pulled was about how many people a manager can lead without losing sight of them. What the memo skipped is what Drucker said the manager was DOING with that sight.

Drucker said the manager's job was judgment. Measurement was the raw material of the judgment. Cut the measurement and the manager still has a job. Cut the judgment and the measurement has no output.

The 2026 layoffs are cutting the wrong direction. They are cutting the people whose work was the judgment and keeping the dashboards that used to feed them. The dashboards are more accurate than they have ever been. They point at nothing.

What the measurers were actually doing

The word "measurer" is a performance-review abstraction. It describes what appears on the annual form. It does not describe the work.

Ask any of the 1,100 people Prince cut what they did every day and almost none of them will say measurement. They will describe absorbing an executive request that was too vague to run with and translating it into three concrete asks the ICs could act on. They will describe catching a project that was drifting sideways in week three and pulling it back before week six, when the drift becomes irrecoverable. They will describe sitting between a customer success complaint and a product roadmap and deciding whether the complaint was a signal or a fluke. They will describe walking a new senior IC through the political history of a decision that a Confluence page would render unreadable.

Those are all judgment tasks. None of them show up on the org chart as their own line item. All of them get invisibly assigned to the person whose title starts with "senior manager, ops" or "director of program management" or "principal audit lead."

Cloudflare's memo says AI can do the measuring. That is true. The memo does not address who does the discernment.

The discernment does not disappear when the measurer does. It gets reassigned by default. Some of it moves up, to executives who now have to interpret raw data streams the middle used to filter into a weekly briefing. Some of it moves down, to individual contributors who now have to figure out on their own whether their work is aligned with a strategy that used to arrive pre-translated. Some of it lands nowhere, and the drift compounds until a quarter ends badly and a post-mortem discovers something that "somehow nobody noticed."

Nobody noticed because there was no longer a somebody whose job it was to notice.

AI and organizational design at the wrong altitude

The current wave of AI and organizational design consulting is fixated on span of control as if span were the variable to optimize. Span is the OUTPUT of an org design, never the input. What determines a healthy span is how much discernment a single mind can hold across how many parallel streams of ambiguity. AI can change that number, but only if the AI is genuinely holding the ambiguity, not just serving reports about it.

Cloudflare's own numbers say internal AI adoption grew 600 percent in three months. That is impressive as a usage metric. It says nothing about which cognitive tasks the AI is doing. Writing code. Drafting emails. Summarizing meetings. Generating first-pass analyses. All useful. None of them are the discernment that a mid-level ops director does when she reads a customer complaint at 11 pm and decides whether it warrants pulling apart tomorrow's sprint plan.

The right org design question sits elsewhere. It asks what kind of judgment can be safely delegated to a model, what kind cannot, and where the org places the humans who still have to make the calls the model cannot.

The 2026 answer, at most companies, has been to skip that question and cut on the JD. That produces a fast reorg and a clean-looking chart. It also concentrates all the residual discernment in two places: the C-suite, which now handles more raw signal than it can process, and the IC layer, which now runs without translation.

Some of what fills the gap is genuine automation. Most of what fills the gap is the CEO doing what a director used to do, and a senior engineer doing what a manager used to do, and neither of them getting paid extra for it, and both of them updating their LinkedIn profiles.

The Klarna receipt

Klarna is the case study that already ran the experiment.

In February 2024, Sebastian Siemiatkowski announced that Klarna's OpenAI-powered agent was doing the work of 700 human customer service reps. He put a number on the savings: $40 million a year. The agent handled 2.3 million conversations. Resolution times fell from 11 minutes to under 2 minutes. That was the high water mark of the AI-replaces-workers narrative in fintech.

By May 2025, Klarna was hiring human agents back. Publicly. Siemiatkowski told Bloomberg and Forbes that customer satisfaction had degraded on the complex interactions, the cost savings had not fully materialized, and the company was moving to a hybrid where humans handled escalations and emotional complexity. That reversal was reported extensively in spring 2025 and continues to be cited in every fintech operations review this quarter.

The Klarna reversal reads on the surface as a customer service story. That reading is too small. The underlying story is the same category error the current middle-management cuts are repeating. What the human reps were doing was labeled "answering support tickets." What they were actually doing was pattern-matching against a customer's tone, deciding when a policy needed a discretionary exception, and catching escalation signals the ticket text did not carry. That work stayed necessary after the AI arrived. Klarna cut on the label. Then it discovered what the label had been hiding, and it had to buy the work back at higher cost.

The middle-management cuts follow the same shape. Cut on the label. Discover the discernment was doing more than the label said. Buy the discernment back later, either through hires or through more expensive kinds of institutional damage.

The org design that survives

There is a different way to run this reorg, and a small number of companies are quietly doing it.

Start with the actual work. Do a genuine audit of what discernment is currently held by the middle layer. Include the invisible tasks: which manager decides when a customer complaint escalates, which director decides when a project drift is worth a course-correction, which finance lead decides when an unusual invoice pattern warrants a partner call. Write those down as first-class tasks with owners.

Then, and only then, ask what AI can do. In almost every case the answer is that AI can prepare the material for the decision. AI can flag the anomaly. AI can draft the recommendation. AI can queue the escalation. AI cannot make the call that the anomaly is worth acting on this quarter versus next, and it cannot hold the political context that says the partner call has to come from a person, not from a workflow.

The right cut, if there is one, is the layer of paperwork the middle managers were producing to justify the work to the layer above them. That is what should compress. The judgment stays. The reporting shrinks. In some org designs the same person now covers more, because the reporting overhead is gone. In others the same person covers less, because the freed capacity gets pointed at judgment tasks that were previously being neglected under the reporting load.

That is what a serious application of AI and organizational design looks like. The target metric shifts from span of control to discernment budget, and span emerges from how the budget was allocated.

Very few companies in the current cut cycle have done that audit. Most have done a JD sweep. This is why the reorgs look identical across industries, from Cloudflare to Amazon to Oracle to Citigroup, whose CEO told analysts in June that automation would let it run middle-office functions with fewer people. It is also why the second-order costs are going to arrive on similar timelines.

The compounding cost of the wrong cut

The immediate cost of a bad middle-management cut is invisible for a quarter or two, because the residual judgment gets absorbed by overwork at the executive and IC layers. Executives take on more meetings. ICs make more calls they used to escalate. The metrics look flat.

By the third quarter, the cost starts to show. Projects that used to get caught in week three drift into week eight. Customer signal that used to get translated into product changes gets buried in a dashboard nobody has time to read. Financial anomalies that a middle-office lead used to flag show up in the external audit instead of the internal one, which is a very different conversation.

By the fourth quarter, the executives are burning out and the ICs are quitting. The company hires "chiefs of staff" and "business operations managers" and "special projects leads." These are the same jobs the measurers were doing, priced at 30 percent more and stripped of institutional memory. The new hires spend their first six months reconstructing context that a mid-level director had built up over three years.

Amazon has been through this cycle three times in three years. Each round of "removing bureaucracy" has been followed by a quieter round of hiring the bureaucracy back under a new title. Oracle is on its first pass and heading into its second. Cloudflare is one quarter in, and the numbers will not tell the story for another two.

Some of this is unavoidable. Any reorg produces friction. The point is that a lot of the friction being produced right now is structural, and it was avoidable, because the theory the reorgs were based on was reading the JD instead of the work.

The strategic point

The AI-driven flattening is happening because AI genuinely does compress a class of work. It does not compress the work the memos claim. It compresses paperwork, reporting, and low-context summarization. The middle management layer was doing all of that, and also doing something else, and the something else is what the current cuts are getting wrong.

Any executive team about to run this reorg should treat the JD as a starting hypothesis, and only that. The right questions are: which judgment lives in this layer, who holds it, what happens to it if the layer is gone, and where does the org place the human who has to make the call the model cannot.

Cloudflare's memo will be studied for a long time. It is genuinely well argued, drawn from a serious reading of Drucker, backed by real internal adoption data. It will also, in retrospect, be understood as the moment the industry made the same category error at scale that Klarna made in customer service, and had to pay to reverse.

Prince himself has said he expects AI-driven layoffs to become "the new normal." That is probably true. What matters for the executives reading this is whether their version of the new normal comes from a real audit of their own discernment, or from copying a memo that read one paragraph of Drucker and skipped the rest.

Architect the design, or buy it back

The companies that survive this cycle will be the ones that mapped their own discernment before they touched the org chart. Cutting fastest is a different game, and its scoreboard runs on a shorter clock than the one that decides who is still here in 2028.

The audit work is hard. It requires sitting with actual managers and asking what they do that never shows up in a performance review. It requires designing an AI stack around the residual judgment, and around the human who has to hold it, and around the reporting overhead that the AI can safely absorb. It requires resisting the memo that would make you look decisive to the board this quarter and cost you the pipeline in three.

You cannot buy this off a shelf. There is no vendor whose product is "the correct middle-management cut for your business." A consultancy that hands you a benchmark span-of-control number for your industry is selling you the wrong metric, elegantly.

Agor AI Advisory does the audit work. We map the actual discernment inside a company against what AI can and cannot absorb, and we design the residual org around the judgment that stays. We have run this on payroll operations, on customer service escalation paths, on product review, on financial close, on legal ops. The output is a reorg you can defend to the board and to yourself a year from now. The alternative produces a clean chart today and a rehire cycle in eighteen months, dressed up as a new headcount plan.