Three stories for one action.
In April 2025 the chief executive of Shopify told his company that before asking for more headcount, a team had to show why it couldn't get the work done with AI [1]. In June, Amazon's chief executive wrote that agents would mean fewer people doing some of the jobs being done today, and that he expected the corporate workforce to shrink over the next few years [2]. Then the cuts came, about 14,000 corporate roles in October and 16,000 more in January, and the explanation changed. The head of people wrote that the company was reducing layers, increasing ownership and removing bureaucracy. On the earnings call the chief executive said the cuts were not really financially driven, and not even really AI-driven, not right now at least. Really, it was culture [3][4].
I don't say that to score a point. I run a large technology organization and I recognize every one of those sentences. What strikes me is that they're three different stories for one action: an AI story, a cost story, and a design story. When an organization can't say which one it's telling, it is usually telling the first two and calling it the third. The design question hasn't been asked. It's been assumed into the headcount number.
This note tries to ask it. If the last three of these notes are right, that agents are part of the labor supply (Governing Digital Labor), that every delegation to them is a record with a source, a channel and an exit (Authority Provenance), and that the human's job in the loop is the gate (The Gate Is the Work), then the organization chart has a new job. Most of what it used to do moves somewhere else.
Span of control was a measure of attention.
The organization chart answers three questions. Who reports to whom. How many. Through how many layers. All three come from one number, and the span of control was always a measurement of human attention: how many people one person could direct and watch, which in practice landed somewhere between five and ten depending on the work. Layers followed by arithmetic. At spans of eight, ten thousand people need five layers of management, and nobody chose that. When Amazon asked its leaders in September 2024 to raise the ratio of individual contributors to managers by at least 15 percent, the stated aim was fewer layers and a stronger sense of ownership at the front line [5]. Same arithmetic, run in reverse. Still attention math.
Twenty years ago Robert Simons argued in Harvard Business Review that span of control is only one of four spans a job carries. The second is span of accountability: the range of trade-offs a person is held answerable for, which can be wider than the resources they control. His design rule was that the resources a job commands have to match what it's accountable for, and he noted that deliberately widening accountability beyond control creates a kind of entrepreneurial tension [6]. Hold onto that gap. It's about to become the whole problem.
Everything in the classic chart assumes the units on it are people: people who need direction, whose attention has a ceiling, and who can be held to account for what they do.
The labor left the chart.
Agents break the first assumption and the third. They take assignments, hold permissions and produce work you can measure. They don't need watching in the sense a span of control measures, because watching an agent means reading its runtime evidence, and a person can read evidence on a thousand workloads in the time one skip-level conversation takes. And they can't be held accountable, because accountability is a thing that happens to a person. Governing Digital Labor called them labor supply and not workforce members. Here is what that distinction does to the chart.
In July 2024 an HR software company announced it would onboard digital workers into its product as employees, with a place on the org chart and performance management like anyone else. Three days later it withdrew the plan, saying the idea had raised questions with no answers yet [7]. The episode was read as a communications failure. I read it as a finding. The org chart is a map of people who can be held accountable. Putting an agent on it either promotes the agent to something it isn't or demotes the chart to an inventory.
So the labor goes somewhere else, and the last two notes already built the place: the delegation record. What authority was delegated, to what, by whom, under what bounds, on what evidence, until when, from whom instruction may be taken, with whom the agent may coordinate, and what it does when it can't finish. The record becomes the map of who is doing what. The chart keeps the people.
Microsoft got partway there in its 2025 Work Trend Index, which has rigid org charts giving way to outcome-driven "Work Charts," names a new role called the agent boss, and proposes a new metric, the human-agent ratio: how many agents a role needs and how many humans it takes to guide them [8]. The instinct is right and the ratio is wrong. Humans to agents counts the labor. What has to be counted is authority, because authority is what stays with the human, and the question is how much of it a person can hold and still see.
Control narrows. Accountability explodes.
Picture the person left on the chart. They direct fewer people than they did, because most of the tasks and some of the people have moved into the record. They authorize far more work, because every grant in the record traces to a name and theirs is on a growing number of them. In Simons' terms their span of control has narrowed and their span of accountability has widened, and the same thing has happened to everyone around them at once. The gap he described as a lever for entrepreneurial behavior has become, at scale, a control gap.
I'd give it a name, because the thing that gets named gets designed. Span of authority: the grants one person holds and remains in real control of. Real control is the phrase from Governing Digital Labor, and it means being in the loop and not only in the chain. The limit on span of control was attention, which came with a natural ceiling, the hours in a day. The limit on span of authority is evidence and stop, and neither has a ceiling at all. A person can be named on ten thousand grants by Tuesday. Nothing stops it except design.
Three things set the span, and none of them is a headcount.
Governing Digital Labor made instantiation a governed variable, separate from approving the agent pattern. Span of authority is its mirror on the human side. Approving a role and approving how much authority it may hold are different decisions, and the second belongs where capacity planning and risk appetite already meet.
Classify a role by what it decides, not how it spends its time.
If that's the span, what is a role? I've spent part of this year drawing a version of this line through a narrower question: which roles in a technology organization are really project management, and which are domain experts who happen to produce a plan. The test that held was decision rights. A coordinating role sequences other people's decisions; a domain role makes the technical ones. How either spent its day told us almost nothing. Where the decision sat told us everything.
The same test redraws the whole chart. Every role left on it is defined by four things.
Work that is none of these four is labor, and labor now lives in the record whether a person or an agent performs it. What's left on the chart is small and heavy. And a question I spent a whole note arguing settles itself structurally. Job families, career ladders, performance reviews and bills of rights attach to the four things above, and all four require a person. Nothing in the record needs a career.
The Gate Is the Work already drew this chart for one team. The six gates in the AI-native SDLC loop are the four things above with names attached: a product owner owns the intent and spec decisions, an engineer issues the grant that lets the agent touch a file, a code owner and a release manager hold the stops at merge and at production, and a service owner reads the 3 a.m. evidence and decides what it becomes. The review queue that piles up in the playbook's opening is a span of authority exceeded in miniature. Build collapsed to hours, the gates didn't, and the people holding them were named on more than they could see. Play the run at allynshaw.com/sdlc/run and you hold a span of authority for five minutes, which is the quickest way I know to feel the constraint.
They used to aggregate attention. Now they aggregate stop.
Here is where I part company with the flattening instinct, including my own. Layers existed because attention has a ceiling, and a layer is how you stack enough attention to cover an organization. Take the ceiling away and the case for most layers goes with it. Most, not all. Governing Digital Labor argued that a worker who disagrees with an agent needs a door to walk through, a named human reachable without going through the agent. Risk at Runtime argued that every runtime rail has to end in a person who can stop it. Doors and stop authorities have to sit somewhere on the chart, and in every organization I know they sit in the layers.
So flatten by all means, but do it as a control change. If you can't say what a layer decides and what it can stop, it's bureaucracy, and the chief executive who hates it is right. If you can, it's a control, and removing it moves a stop authority and an escalation path to someone else. Name that someone. A flat organization where nobody can say who stops a workload has saved on layers and spent the savings on exposure.
Where the next gatekeepers learn.
The Gate Is the Work ended on the claim that approving well, fast, on evidence, is the scarce skill, and that organizations should hire and train for it. This note owes an answer to the obvious objection: train how? Judgment at the gate was learned by doing the work the agent now does. The product owner who can read a spec in minutes once wrote specs for years. The engineer who can tell a plan is wrong once shipped the wrong plan.
The early data says the ladder is already being pulled up. Researchers at Stanford, working from the payroll records of the largest provider in the country, found a 13 percent relative decline in employment for workers aged 22 to 25 in the occupations most exposed to AI, while experienced workers in the same occupations held steady or grew. The declines concentrated where AI automates work rather than augmenting it [9]. An organization that keeps its senior gates and removes its junior work is borrowing against its own future gatekeepers, and the loan comes due in about a decade, when the people who learned by doing retire and the people behind them learned by approving.
The design answer is that apprenticeship moves into the record too. A junior person holds narrow grants with heavy evidence requirements and widens them by demonstrating judgment, the same earned-autonomy curve Risk at Runtime proposed for agents. One mechanism governs the agents and the people who will one day hold the gate on them. The difference is intent. An agent's curve is as steep as its evidence allows. A person's curve has to be deliberately steep enough to produce a gatekeeper before the organization runs out of them.
Fewer people is what the design produces, not what it is.
I run the business of technology, and part of that business is cost, so I'll be plain about the obvious. An organization designed this way will, in time, have fewer layers and fewer people in them. That's a consequence of the design and a poor substitute for it.
A reorganization is a mass reassignment of grants. Every role that goes held decisions, issued grants, read evidence and could stop something, or it was already labor and belonged in the record. Before the role goes, list what it held and reassign each item to a name. If the list is empty, the cut is overdue. If the list is long, the cut is a control change and should be governed as one, because every grant on it that doesn't land on a new name becomes authority nobody is watching. Risk at Runtime has a word for that: unrated. Decision Integrity has another: silent debt, the shortcut nobody wrote down, discovered later by people who can't understand why their options are so narrow.
The memos I opened with will be judged on their headcount numbers. I'd judge them on whether anyone can produce the list.
Maybe this is management with new nouns.
I'll make the case against this note myself. Simons wrote about spans of accountability in 2005 with no agents in sight. Delegating authority to people has always required evidence, earned trust and someone who can intervene. Good managers have always known the difference between the people they direct and the decisions they own. Nothing here is new; it's the discipline with the dust blown off.
Two answers. First, volume. Span of control never had to contend with instantiation. A manager with eight reports couldn't wake up to eight thousand, and a person named on grants can. Design that assumed a ceiling has to be redone when the ceiling is gone. Second, the social layer. The people inside a span of control were a control themselves. They pushed back, asked questions, complained to a peer, missed a deadline visibly, quit. Much of what made a span of eight safe was that eight humans don't drift in silence. Governing Digital Labor argued that agents do, because our controls key off change events and drift has none. Remove the social layer and the span has to be sized by evidence alone, which is a different problem from the one Simons was solving, even though his vocabulary fits it well.
What this borrows, and what it corrects.
I write these notes as one line of thought, so here is where this one leans and where it pushes back.
From Risk at Runtime: coverage as a first-order control, applied here to the humans on the chart, and stop authority with a clock as what a layer exists to hold. From Decision Integrity: the Ownership test as the definition of a role, and silent debt as the name for an orphaned grant. From Governing Digital Labor and Authority Provenance: labor supply without workforce membership, the nine-field record as the place the labor lives, real control as the standard, and instantiation as a governed variable. From The Gate Is the Work: the gate as the human's job, the six gates of the SDLC loop as the first worked example of a role defined by what it decides, and approving well as a skill, which this note tries to say how to grow.
Two corrections. Governing Digital Labor ended on the seam nobody owns, the delegation model that lives between HR, technology, risk and the business, and left the owner open. I'd now answer it: whoever owns the chart, because the chart and the record are one design and span of authority has to be sized where roles are sized. That note also named the disciplines to extend rather than replace: identity, third-party risk, workforce planning, model risk. It left out organization design, which turns out to be the one with the most to change.
What I read.
- [1] Tobi Lütke, memo to Shopify employees, published April 7, 2025, as reported by TechCrunch. techcrunch.com
- [2] Andy Jassy, "Some thoughts on Generative AI," message to Amazon employees, June 17, 2025. aboutamazon.com
- [3] Amazon third-quarter 2025 earnings call, October 30, 2025, as reported by TechRadar. techradar.com
- [4] Beth Galetti, message to Amazon employees on the January 2026 reduction, January 28, 2026, as reported by TechCrunch. techcrunch.com
- [5] Andy Jassy, "Strengthening our culture and teams," message to Amazon employees, September 16, 2024. aboutamazon.com
- [6] Robert Simons, "Designing High-Performance Jobs," Harvard Business Review, July–August 2005. hbr.org
- [7] "Lattice Scraps Plans to Treat AI Bots as Employees After Backlash," SHRM, July 2024. shrm.org
- [8] Microsoft, "The 2025 Annual Work Trend Index: The Frontier Firm Is Born," April 23, 2025. blogs.microsoft.com
- [9] Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab, August 2025. siepr.stanford.edu