Your AI numbers look good. Your results do not. How planets form explains why.

A young star is surrounded by a flat disk of gas and dust. This is where planets come from.

The disk is almost entirely gas. The solids are a rounding error.

The gas is also useless on its own. It sits there, spread thin across billions of kilometers, going nowhere. Something has to gather it.

That job falls to the solids. Dust and ice clump together, and the clumps pull in their neighbors, growing faster as they grow heavier. In a few hundred thousand years one of them reaches about ten Earth masses. Then it stalls. It has consumed everything within reach.

The models say seven million years of almost nothing. The core sits in an ocean of gas it cannot capture. It pulls in what it can and grows by small amounts.

Then the captured gas grows roughly equal to the core beneath it.

The envelope contracts. Contraction makes room. Gas rushes in. The added mass deepens the gravity. Deeper gravity pulls harder. Every step shortens the next one.

Seven million years of nothing. Then everything.

Not every core gets there.

Earth formed in the same disk, surrounded by the same gas, and never held onto it. One Earth mass of rock. Too small to pull. The air you breathe came later, out of the rock itself.

Uranus got further, and then stopped. Its core drew gas in. Then the disk dispersed, ten million years after it formed, and whatever had not been captured was gone. Uranus today is about fifteen Earth masses.

Jupiter is three hundred and eighteen.

Same disk. Same gas. Three outcomes.

Your organization is sitting in a disk.

The gas is AI capability, and there has never been more of it. Foundation models. Agents. Copilots. Open source weights. Features arriving inside software you already pay for. Employees running tools nobody approved. It surrounds you, expands every quarter, and gets cheaper as it does.

None of that is an achievement.

The gas is not yours. Every competitor you have is sitting in the same disk, breathing the same abundance, reading the same announcements. Access is not a position.

Most organizations have already experienced one steep AI curve and mistaken it for transformation.

The tools arrived. Usage exploded. Employees found the obvious individual applications. Every quarter the numbers rose. Then they flattened, because most of what a person could change alone with a chat window had already been discovered.

That curve has the same shape as core formation. Rapid growth, then depletion, then a plateau. The shape was right. The substance was wrong. The technology spread faster than management changed, and organizations accumulated gas while assuming a planet was forming.

Which has not stopped anyone from counting the gas.

Organizations now measure their transformation in tokens consumed. They benchmark consumption across teams and against competitors. On a podcast in March, Jensen Huang said he would be deeply alarmed if a five hundred thousand dollar engineer had not consumed a quarter million dollars in tokens over a year. Internal leaderboards rank teams by usage. Executives put monthly token growth in board decks as evidence of progress.

Token consumption measures how much gas moved through the organization. Not how much of it stayed.

The gas does nothing on its own. It stays spread thin, passing through a thousand individual sessions that nobody sees and nothing connects. An analyst saves forty minutes. An engineer ships faster. A manager cleans up an email.

The gains are real. They add up. They do not compound.

They do not change how work moves through the organization, how decisions get made, or where resources go. Mass moving through the disk is not mass in a planet.

This is what adoption numbers miss. AI can be present in your organization without being captured by it.

Presence is a survey result. Capture is a production system.

Something has to do the gathering. Something has to be dense enough, and concentrated enough, to pull scattered capability into a coherent body and hold it there against everything pulling it apart.

That is a core.

And in your organization, the core is management.

Not Managers. Management.

The distinction matters. A core is not a population of objects sitting near each other. It is a single dense body. Ten thousand rocks scattered across an orbit can hold the same total mass as one body of equal weight and capture nothing. Dispersed mass pulls. Concentrated mass captures.

Management as a system is the decisions, the structures, the incentives, and the mechanisms that determine what your organization can absorb and what it lets drift past. That system either has density or it does not.

Data can inform. Capital can enable. Models can act. But management is where a company’s agency becomes action it owns, where intent becomes a decision someone is accountable for. That is why it is the core, not one input among many.

Most organizations have been adding the wrong kind of mass.

An all-hands about the AI transformation. A training week. A memo encouraging experimentation. A center of excellence. An executive posting about the prototype they built on Saturday. A new role with transformation in the title. Licenses for everyone.

None of these are wrong. Several are necessary. But not one of them changes what the organization is able to hold.

Managerial enthusiasm has no gravity. Managerial decisions do.

What adds mass is narrower and harder. Changing who decides what, and at what speed. Retiring a process that no longer has a reason to exist. Redesigning a workflow around what the work now is rather than inserting a tool into the step where a person used to be. Changing what gets measured, and therefore what gets rewarded. Moving budget away from something that worked last year. Giving an agent defined authority and accepting the consequences. Changing how you yourself work, in public, where your organization can see it.

Every one of those is a decision somebody has to own. That is why they are rare. That is also why they are dense.

You can tell the difference by asking what would have to be undone. An all-hands can be forgotten by Thursday. A training week leaves no trace in the org chart. But a retired approval step, a changed metric, a reallocated budget line, those cannot be quietly reversed. Something in the organization is now shaped differently.

Mass is what survives the enthusiasm.

This is also why the core forms slowly. Every decision on that list costs something politically. Each one takes authority away from somebody, or exposes a process that a career was built on, or admits that a structure the organization spent years perfecting is now overhead. Managers know this. It is why the easy mass gets added first and the dense mass gets deferred.

And a core that is still deferring the dense mass will sit in the disk, surrounded by everything it needs, and stay exactly the size it is.

The Long Middle

Here is what makes this hard. You can be making real progress and see nothing happen.

Fifty-two percent of American employees now use AI at work. Forty-seven percent say their organization has integrated it, a six point jump in a single quarter. In its 2026 workplace report, Gallup found that twelve percent strongly agree AI has transformed how work gets done.

An NBER survey of nearly six thousand senior executives found that eighty-nine percent had seen no effect from AI on their firm’s labor productivity over the preceding three years.

Adoption is rising. Transformation is rare.

I have watched it happen inside organizations doing everything the playbook prescribes, where the spending was real and the usage was real and you could feel the energy in the teams, and still the value curve would not bend.

But look at what people are actually doing with it. Among employees who use AI, half use it for writing and editing. Half use it for search. Only sixteen percent use it for coding assistance, sixteen percent for automation, and those two applications produce the highest productivity gains of anything Gallup measured.

Breadth matters more than depth of any single use. Employees who use AI for one or two things report a positive effect less than half the time. Employees who use it for seven or more report it nine times in ten.

Gallup is careful about what this shows. It does not establish causation, since people who find value may go looking for more of it. And these are individual productivity reports, not enterprise value.

Which is exactly the point. An employee using AI in seven different ways is not a production system. That is one very capable person inside an unchanged structure. Individual use can keep deepening while the organization holds on to none of it.

The use gets broader. The core does not.

This is what the classical models describe. Phase two takes almost all of the time and contributes almost none of the final mass. At the scale of what follows, the curve looks flat for the entire buildup.

Then it does not.

The mistake is reading a flat curve as a broken process. Sometimes it is. Sometimes the middle is simply long.

But this is exactly where the argument gets dangerous, because more than one thing produces this curve. A core approaching crossover looks like this. A core that stalled permanently looks like this. And so does a disk with nothing forming in it at all, where usage climbed, individual gains accumulated, and the management system was never touched. From the dashboard, all three are identical.

The more uncomfortable possibility, for a lot of organizations, is not that they are in a long middle. It is that the middle has not begun.

What separates them is not how long they have waited. It is whether anything dense is accumulating underneath.

And here the data turns uncomfortable. Gallup found that after technical integration, the strongest predictor of frequent AI use is whether an employee’s manager actively supports it. Employees with that support are 8.7 times more likely to strongly agree that AI has transformed how work gets done. Jon Clifton put it plainly, writing that even the most sophisticated neural network cannot overcome an indifferent team leader.

Read that carefully, though, because Gallup is measuring managers. Individual ones. A manager who encourages, who models use, who explains what is changing. Multiply that by thousands and you still have thousands of separate people doing something reasonable in their own corner. That is not a system, and by the definition we just built, it is not core mass.

But it is not nothing either, and the physics is precise about why.

As the envelope grows, it does something specific. Solid bodies that would have flown straight past the planet now pass through gas instead. The gas slows them. Slowed enough, they are captured. The envelope does not create new material. It makes the core better at catching what was already going by.

AI use does this inside an organization. A manager who runs a team on these tools for six months sees things that were always there and never visible. An approval step that exists only to compensate for a limitation that no longer exists. A meeting that was a workaround for a problem now solved. A reporting cycle measuring activity that no longer indicates anything. Their own job, which would look nothing like this if it were redesigned rather than accelerated.

That learning is solid material. It is what the core is made of. And in most organizations it is never captured.

I have sat with these managers. One knows the approval chain he maintains is theatre. Another knows the report she files every week measures nothing anyone acts on. A third quietly stopped using a system the company still pays for and never told anyone. Each of them is right. None of them can act, because the decision that would act on it sits three levels up, in a system with no mechanism for hearing it and no appetite for what it would cost.

This is what a stalled organization actually looks like. Not indifference. Not resistance. Thousands of people who have individually worked out what should change, orbiting a center that cannot pull them in.

Nobody retires a process they have not watched become unnecessary. But watching is not deciding, and the manager who watched is rarely the one who decides.

So the question that separates organizations is not whether their managers support AI. That gap is real and it is wide. The harder question is what happens to what those managers learn. Does it reach the level where structures, metrics, budgets and decision rights change? Or does it circle indefinitely in the layer that first noticed it?

The surveys do not measure that. It would be the number worth knowing.

The disk is full. The core is thin. And in most organizations, what the company is learning never becomes what the company is made of.

The Crossover

Somewhere in that long middle, if the core keeps growing, something changes.

The threshold is specific. The captured gas becomes roughly equal in mass to the core beneath it. Until then the core has dominated the pull. At crossover the envelope is heavy enough that its own gravity stops being secondary. It contracts. Contraction deepens the gravity. Deeper gravity draws more gas, which adds mass, which drives further contraction.

The thing that was being captured starts doing the capturing.

That is the mechanism. Not more effort. A shift in what is driving the process.

And notice what did not happen. Nothing arrived from outside. The gas at crossover is the same gas that surrounded the core through the entire long middle. What changed was the balance between them. A force that had been secondary became strong enough to run the process.

This is emergence. You change something quantitative far enough and the system changes qualitatively. Not more of what it was doing. Something it could not do before.

Which is why the language people use about AI transformation keeps missing. Faster. More efficient. Higher throughput. Every one of those describes more of the same thing.

Crossover is not more of the same thing. It is where an organization stops being a company that uses AI and becomes a company that runs on it. That is not a difference of degree.

The organizational version follows the same logic. Before crossover, management transforms technology into capability. Every workflow that gets redesigned is redesigned because someone decided to redesign it. Every process that dies is killed by a person who chose to kill it, absorbed the political cost, and defended the decision. The pressure has to be manufactured, decision by decision, and someone has to absorb the cost of each one. It is all push, and it is why the middle is long.

After crossover, the technology starts transforming management.

The agents run enough of the work that an approval layer becomes visibly absurd, and keeping it costs more than removing it. Decisions arrive faster than the governance cycle can process, so the cycle changes, not because a committee designed something better but because the old one stopped working. The reporting structure stops matching how work moves, and the mismatch gets expensive enough that redesign is the cheaper option. Roles get redrawn because the old boundaries have stopped describing where the work actually happens.

Someone still has to make each of those decisions. Nothing changes itself. What changes is who has to make the case. Before crossover, change carries the burden of proof. After, the existing structure does.

And then the loop closes. Each structural change makes the next unit of capability more valuable, because there is now something for it to plug into. Each new capability exposes the next structure that no longer makes sense. Redesign raises the return on capability. Capability raises the pressure for redesign.

Every step shortens the next one.

This is what the ROI conversation is actually about. Not whether the tools work. Whether the organization has crossed the point where they compound.

Two things before this sounds like a promise.

The first is that crossover is not an event you schedule. You will not know the moment you crossed it. What you will notice is that the argument changes. Before crossover, the argument is whether something should change. After, the argument is how fast. That is the tell.

The second is harder. The envelope does not explode at crossover. It contracts, and the organizational word for that is compression.

The distance between information and decision compresses. The distance between decision and execution compresses. Handoffs compress. Approval chains compress. Planning cycles compress. The gap between building something and finding out whether it worked compresses.

Everything built to manage those distances compresses with them.

That is not comfortable. Roles stable for a decade lose their rationale in two quarters. Authority moves. Layers whose entire function was to carry information across a gap discover the gap is gone. The same feedback loop that produces the value produces the disruption, because they are one process seen from different places in the org chart.

So the first evidence of crossover does not look like success. It looks like losing control.

And that is the fork. Slow the technology until it fits the management system you have, or rebuild the management system for the capability you now hold. Most organizations will choose the first, and it will not feel like choosing. It will feel like being responsible.

What Closes

The disk does not last.

Within ten million years, the gas disperses. Whatever a core has captured by then is what it carries forward. The opportunity is over, and no amount of core growth afterward recreates it.

Uranus assembled heavy material in the same range as the giants did. What it never got was a runaway envelope. The ingredients were not the problem. The window closed first.

Here the metaphor needs an honest correction, because AI capability is not going to disperse. There will be more of it every year, cheaper, for as long as anyone can forecast.

What closes is not the supply. It is the feeding zone.

A body that reaches runaway does not just grow. It reshapes its neighborhood. It opens a gap in the disk around it, redirects the material flowing through, and changes what anything else forming nearby can still become. The winner does not simply get larger inside the environment. It changes the environment available to everyone else.

Organizations sit in a shared disk too, and it is not made only of technology. Talent is in it. Customers are in it. Capital, partnerships, proprietary data, the people who know how to do this work, the accounts that have not yet decided who to consolidate with. Most of it is still unattached, which is exactly what makes this moment feel unhurried.

A competitor that crosses over does not simply improve its numbers. It starts pulling that material out of the disk you are standing in, and it pulls harder every quarter, because that is what compounding means.

The material does not evaporate. It relocates, and getting it back costs more than keeping it would have.

So there are three outcomes.

Some organizations never build a core. They will buy licenses, run pilots, report usage, and remain exactly what they were with faster tooling. Gas everywhere. No core. No planet.

Some will build a real core and never cross. They will make hard decisions, retire processes, redesign workflows, and then stop. What remains is the dense mass they deferred, the decisions that are more political and harder to justify, and at some point the organization stops paying. That will look like prudence. They will be substantially better than they are now and permanently smaller than what they could have been. Fifteen Earth masses is not nothing. It is also not three hundred and eighteen.

And some will cross.

Now the objection worth taking seriously. Earth is the good planet. Nothing lives on Jupiter. Not every organization should want to be a gas giant, and an executive whose instinct is that they would rather build something dense and habitable than something enormous and gaseous is not being timid.

But that is a different argument than the one most organizations are actually having.

Mass here is not headcount and it is not size. It is capability held and value produced. The crossed-over organization may well look smaller from the inside than the one that never made it. A dense core holding capability many times its own mass, with the layers that existed to move information across distance compressed away. Work that used to require organizational bulk no longer does.

So the Earth objection is real. Some organizations should choose it.

But choosing it is the thing. Deciding you would rather be Earth is a strategy. Arriving there by default, while telling the board you are pursuing transformation, is not.

Which brings it back to the only question that matters, and it is not how much you are spending or how many people have access.

It is whether your core is still gaining mass.

You will not get a signal when you cross. There is no dashboard for a phase change, and the value curve will look flat right up until it does not. But there is a tell, and it costs nothing to check.

Listen to the arguments in your own organization. If they have become about how fast, something has already begun.

If they are still about whether, that only tells you that you have not crossed. It does not tell you whether you are building toward it or standing still, and those look identical from the outside.

So ask the other question, the one with a factual answer.

What has your organization retired, reallocated, or restructured in the last two quarters?

Not planned. Not piloted. Not announced. Retired, reallocated, restructured.

The gas is everywhere, and it is not yours. What you build in the middle of it is.

It is 2 a.m. and someone is awake, watching.

Not writing code. Not building anything. Watching. On the screen in front of her, a fleet of agents is working through the night. Shipping. Testing. Deploying. Deciding. Her job is to be there for the moment one of them does something a human has to answer for. In four hours she hands off to the next person. There is a rotation now. A schedule. Coverage.

Down the hall there is a standing review. It used to be weekly. Then it moved to daily. Now it runs three times a day, because the work does not wait for the morning. There is an agenda. There is a deck. Someone signs off on what the agents did since the last review and approves what they may do before the next one.

If you squint, it looks like a control room. Squint harder and it looks like something older. A floor that never goes dark. People organized around the clock of a machine that does not sleep. A shift.

We were told this layer was disappearing.

I had been watching management change before the agents arrived. Then they took the execution, and what had been slow turned fast. The sprint emptied out. The standup lost its purpose. I spent much of this series describing it, layer by layer. The process that was no longer yours. The signals beyond your capacity. The delegation that had lost its meaning. The expectation was simple. The management layer would thin out and fall away. Most organizations are only arriving at that point now. One reader told me he had stopped using Jira altogether. No middleman, he said. Not a software one, and soon not a human one.

He was half right. The old middleman is gone. A new one has taken his place, and it answers to a different name. We do not call it management anymore. We call it governance. Oversight. Approval. Audit. Assurance. The org chart did not flatten. It refilled. Different boxes, watching a different kind of worker, at a different speed.

The bureaucracy is back. Nobody decided to bring it back. It moved into a space nobody thought to design, which is also why it does not have to stay. That is the part worth understanding.

Start by giving the core its due. Some of this layer is not optional and never will be.

Somebody has to be accountable. When an agent makes a call that costs money, breaks a system, or harms a customer, a person has to answer for it. That cannot move to the machine. An IBM training manual put it in one line back in 1979. A machine cannot be held accountable, so a machine must never be the one making the decision. Almost fifty years later that sentence is the whole problem and the whole point. The agent can act. It cannot answer. A human still has to.

Authority works the same way. An agent operates inside a boundary, and someone has to draw that boundary and own it. Underneath both there is a deeper pull. On their own, organizations generate administrative structure the way a body generates heat. It is a byproduct of running. Every serious attempt to strip it out has produced more of it in a different costume. I traced that pattern earlier in this series, through Ellul and Graeber, and the governance rising around the agents is the proof they were right. Left alone, the apparatus regrows. The only word that matters in that sentence is alone.

So part of what you are watching at 2 a.m. is real. Accountability has to sit somewhere. Authority has to be bounded. That core is not the problem. And it is not the bureaucracy.

The bureaucracy is everything we built around it.

Look again at the floor that never goes dark. Most of what fills it is not the irreducible core. It is the old org chart, photocopied onto a new kind of worker. The status review. The reporting line. The tier of people whose whole function is to watch other workers, except the workers are now agents. Nobody designed this for agents. We reached for the only shape we had and pressed it down on top of them.

And much of it does not even do the job it claims. Take the approval, the most basic of the new rituals, the human who signs off before the agent acts. Anthropic instrumented its own coding agent and found people approved about 93 percent of the prompts it raised, and read them less carefully the more it asked. The exact number is not the point. The pattern is. When the sign-off comes often enough, it stops being judgment and becomes throughput. We rebuilt the approval to keep a human accountable, and the human is waving through nine of every ten without looking. The loop still has a person in it. The judgment left a long time ago.

Why does that happen? Why does the shape come back, uninvited, every time?

Because nobody planned for it. Not the bureaucracy. The organization.

Here is the thing leaders are very good at, and the thing they almost never do. They are good at plans. Give a strong executive a goal and they will hand you a master plan, a multi-year roadmap, a strategy with milestones and dependencies and a budget that holds. They will plan for the plan to succeed, with room for delay along the way. What they will not do is plan for the organization that exists the morning after the plan succeeds.

The plan has a finish line. The organization does not. The day you hit the goal is not an ending. It is the first day of a company that now works differently, and almost no one has drawn that company. They designed the bridge. They never designed the far bank.

So the agents arrive. The execution compresses. The old roles hollow out. And a new organization is standing there that nobody sketched. The questions land all at once. Who answers for this. Who approves that. Who watches the things that run all night. Into that vacuum walks the oldest reflex in the building. The bureaucracy is not chosen. It is what grows where the far bank is left undrawn.

This is not the first time the pattern has surfaced in this series. It is the same move I described with productivity.

Organizations chased the gain. They planned for output, for speed, for the team of five doing the work of fifty. And I argued, articles ago, that productivity is never the end of the story. It changes roles, then processes, then hierarchies, then the organization itself. That cascade is the far bank. They planned to be faster. They did not plan for what faster turns them into. The hollow roles, the new ones that have to be invented to absorb the change, the coordination that breaks, all of it arrives unplanned, and the organization improvises a response in real time.

You can watch the same thing in any large system that ever went live. The project plan ends the day the system is deployed. Nobody drew the organization that has to run on it. So the workarounds grow in the gap, and a year later the shadow processes are the real processes, and no one designed any of it.

Put it together and the deeper problem comes into view. None of this is being designed. It is being discovered, by stumbling, after the fact.

I read the best current thinking on how to govern agents. It is serious work, and most of what it proposes is right. Identity for every agent. Least privilege. Permission boundaries. A named owner for every consequential action. There are detailed charters for how much authority a single agent may hold, down to its spending limits and its expiry date. There is not a page on what the organization becomes once those agents are everywhere. The controls are engineered to the decimal. The organization they sit inside is left to assemble itself. The guardrails come first. The question of what we are becoming comes later, if it comes at all. The far bank is the afterthought, even among the people thinking hardest about this.

And it is not only the thinkers. Look at what the market is selling. The new products arrive named some version of the same thing. Control plane. Command center. Control tower. They discover the agents, give them identities, trace what they do, enforce the permissions, keep the audit trail. All of it necessary. None of it sufficient. A control tower can tell you where the planes are and whether they are cleared to move. It cannot tell you what airport you are building, which airspace should exist, or what should never be allowed to fly there at all. The market found the metaphor before leaders found the operating model. They built the tower. Nobody drew the airspace.

And here is the part that should stop anyone who has run an operations floor. We already know how to keep humans watching machines that never stop. Security operations centers have run this way for years. Shifts. Coverage. But not only coverage. Severity levels, escalation paths, runbooks, clear authority over when to step in, an after-action review every time something slips through. People watching automated systems around the clock because the threats never rest. That is not oversight as ritual. It is oversight as an operating model, built on purpose by people who had no other option. And the rest of the organization is about to reinvent it the hard way, committee by committee, because almost no one thought to look at the teams that already solved it.

So what is the work now?

It is the part that got skipped. Not governing the agents. Designing the organization the agents create. Deciding, on purpose, what this company becomes once execution is abundant and the old roles are gone. Which parts of the old structure earn their place. Which exist only because nobody stopped to ask. What accountability means when a human cannot read every action. What a career is when the apprenticeship has dissolved. What the human is for, now that the machine does the work.

That thinking cannot be delegated. Not to the agents, which is obvious. Not to a governance committee, which will hand you controls and call it a strategy. Not down the management chain, because the chain is the thing in question. It sits with the people at the top, and it is the one thing this technology does not make easier. More and more of their execution can be handed to a machine. The execution is abundant. This is not. Deciding what your own success should turn you into is the last thing a leader cannot give away.

And it cannot be skipped. That is the whole lesson of the new bureaucracy. Skip it, and the organization does not wait for you. It fills the empty space with the oldest pattern it knows.

I opened this series by telling you the bottleneck had moved. For a century the constraint was execution. Not enough hands, not enough hours. Management existed to wring the most out of scarce human work. Then the agents arrived, execution stopped being scarce, and the bottleneck moved up, onto the management layer itself.

This is where the argument finally lands. The bottleneck moved, and then we put it back. For two years we watched the layer that slowed human work begin to dissolve, and the moment the machines were fast, we built a new one to slow them too. Same instinct. Same shape. Reinstalled for agents faster than we ever managed to remove it for people. The sign-off came back. The watchers came back. We did not decide any of it. We defaulted into all of it. When the night shift comes, it will arrive the same way.

It did not have to go this way, and it does not have to from here. There is a version where you stop, before the vacuum fills, and design the organization on purpose. Decide what the agents are for, what the humans are for, what answers to whom, and what gets left behind for good. That is the harder path, and the only one that ends somewhere you actually meant to be.

The agents were never the question. The question is whether leaders will design the organization on the far bank, or keep waking up inside a factory they never chose to build.

Your engineers are faster than they have ever been. Sprints that took weeks finish in days. Prototypes appear over a weekend. You have seen it, and you have celebrated it. Then you look at your actual products, the things customers pay for, and they look almost exactly like they did two years ago, with an AI assistant panel added to the side. Faster building. Same product.

This is not a coincidence. The methodology behind it was adopted years ago for reasons that were once sound, and never reexamined since. But the methodology is not really what should worry you. What should worry you is what your organization now uses it to protect, which makes this a question of governance, not of release cadence.

Let me show you where it came from.

The constraint that built your company

In 2011, Eric Ries published The Lean Startup, and a generation of builders reorganized around one idea. Execution is expensive, so do not waste it on untested assumptions. Build the minimum viable product. Ship it. Learn. Iterate.

The logic was sound because the constraint was real. Turning an idea into working software took scarce engineering time, long cycles, and real capital. Building the wrong thing at full fidelity could end a company. Ries was not arguing against vision. He was arguing against waste, against pouring scarce execution into an unvalidated guess. That was disciplined. What organizations did with the idea over the next decade was not. The insight was learning under execution scarcity. The mutation was incrementalism without imagination. MVP stopped meaning test the riskiest assumption and came to mean ship the smallest thing that demos.

From that doctrine grew the phrasing your organization uses today. Crawl, walk, run. Introduce a narrow capability to a small audience. Expand scope and automation. Reach full scale. Each phase derisks the next. The framework is now a default across enterprise technology. Consultancies sequence digital labor across crawl, walk, and run phases by value and risk. Vendor roadmaps tell enterprises to start with low-complexity use cases before touching the product itself. It is the reflexive answer the entire field reaches for.

It was the right answer when execution was the bottleneck.

The constraint moved

That bottleneck did not disappear. It moved.

Execution is not free now. Integration, security validation, compliance, data migration, pricing changes, sales enablement, the slow work of earning enterprise trust. All of that is still expensive and still real. What changed is the specific part of execution that MVP was built to ration. A small team, sometimes a single senior product thinker working with agentic tools, can now explore, specify, and prototype at a depth that used to take a full product and engineering cycle. The scarcity that justified incrementalism as the default, the reason you could not afford to think everything through before you built, no longer justifies it.

And yet your organization keeps crawling.

It keeps asking what can ship by the next conference, what will be ready for the next release window. Those questions optimize for a calendar, not a customer. They produce one predictable thing, the shallowest feature set that demonstrates well on a stage. No vision. No architecture. No confrontation with the hard trade-offs.

First, what this is not about

Before I go further, one distinction, because the objection is already forming.

None of this applies to how you adopt AI inside your organization. Rolling out tools to a workforce, building literacy, governing risk, proving value against current operations. Phase that. You cannot hand an entire workforce autonomous agents on a Monday. For internal adoption, crawl, walk, run is responsible change management, and I am not arguing against it.

I am arguing about something else entirely. Not how you adopt AI. What you build with it. Those are two different questions with opposite logics, and the whole problem begins when an organization answers the second using the playbook it learned for the first. Hold that line, because everything else depends on it.

Phased releases are fine. Phased thinking is fatal.

Here is the distinction that matters.

There are two ways to release a product in stages. The first is to release without a destination, hoping the market will reveal where to go. The second is to hold a complete, deeply considered vision and release in deliberate stages toward it, adjusting as customers respond. Your organization has institutionalized the first and convinced itself it is doing the second.

This is not an argument against phased releases. Complex systems require staged delivery. It is an argument against phased thinking. Build the complete vision first. Specify it to the depth the product demands, which you can now actually do. Then release toward something you have fully thought through, not toward a destination you are still discovering one sprint at a time.

The difference sounds subtle. It is not. It decides whether you make your most important choices deliberately or by accident.

What crawling actually costs you

The cost of crawling is not slow progress. It is the decisions you make without knowing you are making them.

Every product has one-way doors. Choices that are irreversible, or so expensive to reverse that you never will. Which data model. Which trust boundaries. Which integration architecture. These shape everything downstream. When you crawl, you walk through them under deadline pressure, picking whatever is fastest for the first release, and then you spend years building around the consequences.

A team that has done the deep thinking first can see every one-way door before it reaches one. It can stop, analyze the choice, test it against the full vision, and decide on purpose. A crawling team discovers its one-way doors only after they have shut.

This is the spine of the whole argument, so let me state it plainly. The path you take while crawling forecloses the destination you claim you are heading toward. You do not arrive at the considered product through a sequence of unconsidered steps.

What your product has to become

So far this is about how you build. The larger question is what you build, and here the stakes are higher than most leaders have let themselves see.

Agentic AI does not only help your teams spec and code faster. It changes what a product can be for a customer. Agents analyze, decide, act, and follow through. They run the workflows that today require a trained professional in front of a console. Which means your product no longer has to be a tool that presents information for a human to act on. It can become the actor.

Take a security product, where the shift is easy to see. A conventional console assumes the analyst is the actor. The analyst triages the alert, inspects the evidence, correlates the signals, decides severity, opens the ticket, escalates, remediates. Every screen and every permission model is built around a human doing those steps. An agentic product assumes the product is the actor. It classifies the incident, gathers the context, verifies whether the threat is exploitable, proposes or executes containment, documents its reasoning, updates the policy, and learns from what the analyst confirms or overturns. The human moves from operating the console to governing the agent. That is a different product built on a different premise, and its surface, data flows, permission model, and pricing all follow from that premise.

Once the product acts, evaluation stops being a QA function and becomes part of the product’s core design.

Most of what your organization ships as AI today is a band-aid. An assistant panel grafted onto an existing console. A summarization layer over existing alerts. A natural language box over an existing schema. These are useful. They are also shallow, because they assume the same human workflows, the same operating model, the same division of labor between product and user that existed before agents could act.

This is the divide I drew earlier in this series. The band-aid injects AI into the product you already have and inherits all of its assumptions. The clean sheet inherits nothing. Mode 3 and Mode 4 were the names for it. Everything in this article is the case for crossing from one to the other, and the reason your organization will fight the crossing.

So be precise about what a clean sheet is. It is not your product with AI added, and it is not a technical refactor. It is the product redesigned around a new premise. Software can now reason, decide, act, and be evaluated as an actor inside the customer’s workflow, rather than a tool the customer operates.

The clean-sheet question reaches past architecture. Your per-seat licensing assumes one human operator per seat. Your support tiers assume escalation paths designed for human analysts. Your packaging maps to human roles. Every one of those assumptions was built for a world where the product could not act on its own. The exercise is to surface all of them, name them, and challenge each one. Some survive. Many will not. None should survive merely because nobody thought to question it.

There is a single question every product team should be made to answer. If you had to build this product from scratch today, with everything agentic AI now makes possible for both the building and the product itself, what would it do, and how would it look? If the honest answer is “not much would change,” the thinking has not gone deep enough.

Why this will never come from below

Crawling is the product mechanism. The funnel is the governance mechanism that keeps it alive. This is the part that makes it your problem and not your teams’.

You might read everything above and conclude the move is obvious. Tell the product teams to do the clean-sheet work. Commission the specs. Wait for the bold ideas to rise through the organization. That instinct is exactly why it will not happen.

The clean-sheet product is, by construction, the worst-looking bet in your portfolio. It serves a customer who may not exist yet. It shows lower margin than the product it would replace. It cannibalizes revenue you are currently booking. And it cannot prove itself in any of the numbers your organization uses to allocate resources, because those numbers measure the product you already have. Ask the clean-sheet bet to justify itself in the metrics of the thing it replaces, and it loses every review. That is not a flaw in your people. It is the design of the funnel.

So when the bet rises from below, it dies. The product manager measured on this quarter’s roadmap rationally declines to champion the thing that threatens it. The reviewer who insists on input metrics that correlate to outputs before committing, which sounds like rigor, is in fact applying the exact filter that passes incremental features and rejects the clean sheet every single time. Everyone behaves sensibly. The disruption dies anyway.

You can watch this happen in any roadmap review. The AI feature that drops neatly into the next release gets a score, an owner, and a date. The proposal that asks whether the console should exist at all gets called interesting, and gets sent back for more validation. One is legible to the process. The other threatens it. The process funds what it can read.

This has a name. Clayton Christensen called it the innovator’s dilemma, and his finding was not that incumbents are stupid or timid. It was the opposite. The incumbents who lost were well run. They listened to their best customers, defended their best margins, and allocated capital responsibly, and that discipline is precisely what killed them, because a disruptive product looks worse on every measure good management trusts, right until the moment it wins completely. Good management was the cause of death.

So drop the idea that your organization lacks imagination. It sees the clean sheet. Its management system is doing exactly what it was built to do, which is protect the product you already sell. The system is not broken. It is working as designed.

Which is why a clean sheet cannot be delegated downward. Your organization’s resource allocation process is structurally incapable of funding it. The only place the decision can be made is above that process. That means you.

What only you can do

The work itself is achievable. The tools exist. The talent exists, somewhere in your organization, though probably not where your incentives have pushed it. The constraint is not capability. It is whether someone with the authority to override the funnel actually does.

That is a short list, and it is yours alone.

Mandate the clean-sheet exercise instead of hoping it emerges. Give teams the explicit charge to specify the product as if no prior version existed, to the full depth the product demands, including how they will evaluate an agent that acts on a customer’s behalf. In a product that decides and executes, you cannot define what “good” means after launch. The evaluation framework becomes a design artifact, built before the code, not a quality gate bolted on after. If a team cannot define what good looks like for an agent making autonomous calls, it does not understand the product well enough to build it.

Protect the work from your own metrics. The clean-sheet bet will lose any review that scores it against the current product. So it cannot sit in that review. It needs its own criteria, its own runway, and air cover from you when the quarterly numbers ask why resources are going somewhere that is not yet earning.

Staff it with the people who can think at this level, and understand that your current system has been quietly driving them out. Deep product thinking is a specific capability. A planning process governed by sprint velocity and release dates reproduces the product you already have. Find the people who can do the other thing, and free them from the machine that rewards the first.

Own the cannibalization decision, because no one beneath you can. The clean-sheet exercise will tell you that the right product for this era undercuts the one you spent years building. It may invalidate assumptions your whole organization is structured around. A product manager cannot choose to cannibalize the company’s revenue. Only you can. That decision is the job now.

The progress that is not progress

One warning before the close, because it is the trap that catches the leaders who are paying attention.

You will watch your AI adoption metrics climb. Usage up, literacy up, real ROI against current operations. The dashboard will be accurate, and the progress will be real. And you will be tempted to read it as transformation. It is not. It is the measure of how well your organization uses AI inside the product it already has. A correct dashboard answering a question you did not ask. The series has kept circling this from different angles, and here is its sharpest form. You can be measurably succeeding at adoption while the clean sheet that decides your future never gets built. The climbing curve is not evidence that you are transforming. It is the comfort that lets you avoid it.

The real challenge is you

Everything in this article is achievable. The tools are here, the depth of thinking is now possible, the people exist. The only missing piece is a leader willing to act on what the clean sheet reveals.

And here is the part that should keep you up. Your metrics will not warn you in time. The dashboard will look healthy right to the end, because it measures the product you have, not the one that is replacing it. By the time the clean-sheet product shows up in your numbers, it will no longer be a strategic option you get to weigh. It will be a market fact you have to survive. Someone will have built it. The only question ever in your control was whether that someone is you.

So the question is not whether your organization can build faster. It can. The question is whether you are willing to authorize the product that makes the one you sell today look obsolete.

That is not a question for your product teams. It was never theirs to answer.

There is a moment in every roadmap review right now that tells you everything. One team shows that agents have cut their delivery cycle in half. Another shows a legacy system rebuilt in weeks by three engineers. A third shows AI features embedded into the flagship product. The room nods. Everyone calls it transformation.

They are not showing the same kind of change. And most organizations have no language for the difference. Without that language, they cannot see what they are doing, what they are missing, or where the real transformation lives.

This article is the map.

Four Modes

Mode 1: Augmented Development

This is where most organizations start. The team adopts AI tools and the process immediately strains. Sprints compress from two weeks to days. The engineer’s role shifts from writing code to specifying and reviewing what agents produce. The manager loses the signals that used to define the job. Velocity, story points, burndown charts. All of them break when agents finish in hours what humans planned for weeks.

This is also where the infatuation lives. The energy is real. Teams are shipping things that used to take weeks, and leadership is excited. The metric everyone watches is consumption. How many tokens, how many agent sessions, how many teams have adopted the tools. Cross-industry leaderboards compare how much AI organizations consume. Internal dashboards track the same. But consumption is not value. Activity is not transformation.

This is real change. Roles and cadence are shifting. The conversations about what the team does and how it does it are happening in every standup, if standups still exist. But it is evolutionary change. Mode 1 inherits everything from the pre-agentic world. The product, the architecture, the assumptions, the organizational structure. It changes how fast the work happens. It does not change what the work is. The organization is asking one question: how do we do what we do, faster?

It never asks: should we be doing this at all?

Mode 2: Rapid Reconstruction

A team of two or three engineers takes a system that was built over two years by twenty people and rebuilds it in weeks with agents. The economics flip. The original build was constrained by execution scarcity. Every feature was a trade-off. Now the constraint is gone, so the team can rebuild the whole thing, not patch it. Mode 2 primarily inherits the product definition. The process, the team structure, the timeline are all new.

But the real value of Mode 2 is not the rebuilt system. It is what the organization learns in the process. You have a known baseline. You know what the system did, how it was built, what it cost. The old system is the control group. The rebuilt system is the experiment. For the first time, the organization can compare two production models against the same functional goal and see clearly what works at agent speed and what does not. Which processes need to change. Whether the infrastructure supports the new way of working or blocks it. What the team actually does versus what you assumed they did.

Mode 2 is the organization’s learning lab. The lessons are concrete and shareable across teams. What processes need to be adapted or redesigned, how roles change when agents handle the execution, what management looks like when three engineers do the work of twenty, and whether the infrastructure supports the new way of working or blocks it. These lessons build the institutional knowledge you will need before anyone can attempt something more fundamental.

Mode 3: AI Capability Injection

The product team adds AI features to an existing product. An AI assistant layer. Natural language interfaces. Intelligent automation. From the outside, the product looks transformed. New capabilities. New user experiences. The press release writes itself.

From the inside, nothing fundamental has changed. Mode 3 inherits the deepest layers. The architecture, the data model, the assumptions about who the user is, what problem they have, and how the product solves it. All inherited from the pre-AI design. The team bolted intelligence onto a structure that was never designed for it.

This is where most organizations focus. It ships visible features. It gets executive attention. It feels like transformation. And it never forces anyone to question whether the product itself, its architecture, its assumptions, its reason for existing, still makes sense in a world where agents can do things that were impossible two years ago.

Mode 4: Clean-Sheet Development

A team asks one question: if we were building this product today, knowing what agents can do, what would we build?

The answer is usually not the product they have. It might not be a product at all in the traditional sense. The user interaction model changes. The architecture changes. The team structure changes. The definition of what the product does changes. Everything is on the table because nothing is inherited.

This is the hardest mode. Not technically. Technically, agents make building from scratch faster than ever. It is hard because it requires leadership to say that the thing we built, the thing that has customers, revenue, a team, a roadmap, may not be the right thing anymore. That takes vision. It takes courage. And it takes a kind of organizational permission that most companies have never granted.

In large enterprises, clean-sheet does not always mean abandoning the existing product. It can mean building a parallel path unconstrained by the current architecture, roadmap, and customer requests, and then deciding what should migrate, merge, or be retired. The point is not destruction. The point is that the new path inherits nothing. It starts from the question, not from the backlog.

Almost nobody is here yet.

The pattern across all four modes is simple. The more an AI effort inherits from the pre-agentic world, the less transformational it is likely to be.

I have seen all three modes firsthand. I have watched teams double their output with agents and celebrate the acceleration. I have seen a legacy system rebuilt in weeks by a handful of engineers. I have sat in reviews where product teams proudly demonstrated AI features layered onto architectures designed a decade ago. Everyone in every one of those rooms called it transformation. I am now pushing leaders and product managers to consider Mode 4. The resistance is real. Not because they disagree with the idea. Because the other three modes are working. They are shipping. They are showing results. Asking someone to question the foundation while the building is producing returns is the hardest conversation in any organization right now.

The Trap

The four modes are not interchangeable options. They form a hierarchy. Mode 1 is where everyone starts. Lowest risk, lowest disruption, real but bounded value. Mode 2 is the proving ground and learning lab. Mode 3 is the comfortable default. Mode 4 is where transformation actually lives.

The trap is Mode 3.

Mode 3 feels like transformation. It ships features. It generates excitement. It gets budget. The product has new AI capabilities. The customers see something different. The demos are impressive. But underneath, the architecture was never designed for this. The data model carries assumptions from a different era. The team is retrofitting intelligence onto a structure built for a world where agents did not exist.

This is the horseless carriage of product development. The carriage has a motor now. It goes faster. It is still a carriage.

And the market reinforces the trap. Your customers are asking for AI features in the products they already use. That is a Mode 3 request. Some are asking you to deliver faster. That is a Mode 1 request. Nobody emails their vendor and says: please rethink your entire product from scratch. Mode 4 has no customer demand signal because customers cannot ask for what they have not imagined. The market will validate Modes 1 and 3 all the way to irrelevance.

Mode 3 does not just disguise the absence of transformation. It competes with it. It absorbs the budget, the roadmap capacity, the executive attention, and the organizational confidence that might otherwise be directed at harder questions. By the time Mode 3 has consumed the oxygen, there is nothing left for Mode 4.

The irony is that some of your most senior leaders have already leapt to Mode 4 on their own. In the previous article, I described the agentic executive, a leader who picks up an agent and builds something themselves, bypassing the organization entirely. They are not doing Mode 1 at higher speed. They are collapsing the chain between vision and execution into one person, starting from a blank page, building what they believe should exist. Mode 4 is already happening. It is just happening outside the operating system of the company. The executive can start from zero because they are not trapped inside the product team’s backlog, architecture, and inherited assumptions. The teams cannot do Mode 4 because they are bound by exactly those things. And the organization learns nothing from the executive’s leap. The weekend prototype does not change how the teams work on Monday. The gap is not closing. It is growing.

Organizations believe they will naturally progress from Mode 1 through Mode 4. That is not guaranteed. Modes 1 through 3 produce signals that something deeper needs to change. The manager who loses her old metrics in Mode 1. The team that discovers its infrastructure cannot support what agents make possible in Mode 2. The architect who realizes the product was designed for a world that no longer exists in Mode 3. The signals are there. But the modes also produce enough visible progress, enough shipped features, enough positive dashboards, that the signals are easy to ignore. The press releases will celebrate the AI transformation.

And you will never reach Mode 4. Because Mode 4 is not the end of a progression. It is a different starting point. It begins with a question that Modes 1 through 3 never ask: what if the thing we built is no longer the thing we should be building?

What Are You Building Toward?

The four modes are not a maturity model. They are a map. Most organizations will recognize themselves in Mode 1 or Mode 3. Some will see Mode 2 as an opportunity they have not yet used. Very few will see Mode 4 as something they are ready to attempt. It requires the boldest leadership. Not just the courage to rethink a product, but the understanding that the theory itself, how you build, how you organize, how you create value, is changing.

The question is not which mode you are in today. Every organization has to start somewhere. The question is whether you have even considered Mode 4. Many organizations have not. They are operating within the modes they know, doing business the way they have always done business, adopting AI within the boundaries of their existing assumptions. Mode 4 was not on the map until now. It is on the map now.

The real danger is not that your organization is behind in adopting AI. It is that your organization is becoming very good at adopting AI into a model that should no longer exist.

Imagine a ship in the seventeenth century, sailing toward a coastline none of the crew has seen before. The captain stands on the deck. A lookout sits in the crow’s nest, high above. The work between them is coordination. The lookout has altitude. The captain has authority. Neither alone can navigate the ship. The lookout sees the horizon and calls down what they see. The captain interprets the call, judges its meaning, and decides the response. The system works because no single person has both the vantage and the command.

Then the telescope arrives.

The instrument does not give the lookout a better view. It gives the captain the lookout’s view directly. The captain can now see the horizon, name the ship on it, judge its course, and decide the response without intermediation. The crow’s nest does not disappear. Navigation still depends on many instruments and many hands. But the telescope changed one thing: for the work the captain chose to do directly, the lookout’s role had become optional.

The technology was given to the captain, not the crew. This is not an accident of who could afford it. It is the structure of the change. Power is distributed when capability has to be distributed. When capability concentrates, power concentrates with it. The telescope did not democratize navigation. It centralized it.

Something analogous is happening now, inside organizations. A new instrument has arrived. The agent. It helps the engineer write code faster. It also gives the executive a capability the executive did not have before: the ability to reach the work directly, without the layers that used to mediate between intent and outcome. The engineer’s role is not diminished. What is new is that the executive, for the first time in a century, can hold both vision and execution in the same hands.

The previous article in this series zoomed out to see the corporation being reshaped by technique, by the moving boundary of the firm, by forces larger than any one organization. This article zooms in. The agentic executive is one of the most visible shapes the reshaping is taking inside the firm. They are not the cause of the change. They are the trial period made visible. The organization that built them is being tested, in real time, against the question of whether it can transform into something the old language does not yet describe. The agentic executive matters not because senior leaders are building again, but because their building reveals that the premise management was built on, achieving outcomes through people, is no longer structurally guaranteed.

The End of the Lean Startup Age

The lean startup was a response to two constraints, not one.

The first was uncertainty. Eric Ries defined his framework precisely: a startup operates under conditions of extreme uncertainty. The builder does not yet know who the customer is or what the product should be. Lean is the disciplined response to that not-knowing. Release a slice. Learn from the market. Iterate. Converge toward a product whose final shape was unknown at the start.

The second constraint was execution. Ries did not need to name it because in his world it was a given. Building anything required people, time, and capital, all of them scarce. Even if a founder held the full vision, they could not build it alone. They needed engineers, designers, testers, managers. The organization existed because the work exceeded what any single person could execute. The concept that best reveals this hidden constraint is the MVP itself. Minimum viable product. The word “minimum” is not just a learning strategy. It is a resource constraint. You build the minimum because you cannot afford to build the maximum. Ries designed the MVP to enable learning with the least amount of effort and the least amount of development time. That design makes sense only in a world where effort and development time are scarce. Lean assumed both constraints simultaneously: you do not know exactly what to build, and you do not have the resources to build everything and see what works.

A different construction philosophy has always existed. Hold the entire vision in advance. Design the architecture, the strategy, the roadmap. Release and refine toward a known endpoint. Aircraft are designed this way. Operating systems are architected this way. Iteration is convergence, not exploration. The endpoint is in sight from the start.

This philosophy required two things at once. A person who could see the full vision. And an execution capacity that could match it. For two centuries of industrial work, no individual had both. Vision lived at the top of organizations. Execution lived distributed across hundreds or thousands of hands. The translation between them was the entire job of the middle.

Both constraints are now lifting for a specific class of builder. Agents close the gap between intent and working artifact. Senior leaders who hold the full product vision, who have spent decades building the judgment to know what the product should be, now have execution capacity to match. They are not operating under Ries’s conditions of extreme uncertainty. And they are no longer bound by execution scarcity.

Two figures emerge in the post-lean age. The solopreneur outside the firm. Pieter Levels, Maor Shlomo, Ben Broca. People building meaningful businesses alone, with agents instead of teams. They have been written about extensively. The figure inside the firm has not received the same attention. That is the subject of this article.

The Phenomenon, Named

I spoke recently with a senior executive who described the experience in a way I have not been able to improve on. He said it was like being a great soccer player who had become a coach, and now, with new capabilities, could step back onto the field and play again. Better than he ever played as a young man.

The metaphor is worth holding onto. This executive spent decades building systems, shipping products, developing the judgment that comes only from years of doing the work. Then he was promoted. Into management. Into leadership. Into the role where outcomes happen through other people. The work he loved, the craft he had mastered, moved to the hands of others. He directed. He reviewed. He approved. He no longer played.

Now the agent gives him the field back. Not in a diminished form. In a better form. He can see the architecture of the entire product because he spent twenty years learning it. He can specify what needs to be built because he holds the full context: the company strategy, the competitive landscape, the technology stack, the customer needs, the conversations with other executives that shaped the direction. No one else in the organization holds all of these at once. And the agent can produce certain classes of working artifacts faster than the team structure around him could absorb, prioritize, and deliver them. Vision and execution, reunited in the same person.

He is not alone. Job van der Voort, CEO of Remote, a global HR platform valued at more than $3 billion, builds features for his own product. He goes through the standard code review process like any engineer on his team. He wanted a feature. He did not file a request. He built it. Moshe Bar, CEO of Codenotary, built a 140,000-line production application using Claude. He is not a programmer. He personally edited roughly ten lines of code. His estimate of the conventional cost: three or four senior developers at $400K to $500K each. Wade Foster, CEO of Zapier, built a personal AI chief of staff in Cursor. Woodson Martin, CEO of OutSystems, described the impulse behind his own experiment in blunt terms: “I was tired of explaining it to somebody who was supposed to build it for me.”

An earlier article in this series called the broader pattern the infatuation phase: the organizational stage where the capability itself becomes the object of attention. The agentic executive is the infatuation phase at its most concentrated. Not the developer in love with running parallel agent threads. The executive in love with playing the game again.

There is a name for this figure. The agentic executive. A senior leader inside an organization large enough that there is structurally a layer between them and the work, who chooses to bypass that layer using agents. Three conditions define the pattern. Authority over the organization. An organization built to do the work. And the choice to do it themselves instead.

The solopreneur fails the second condition. The CEO of a five-person company arguably fails the second. The CVP, the SVP, the CTO, the CEO of a large enterprise meet all three.

These are not personality cases. They are structural ones. The capability has been distributed asymmetrically, and the executives experiencing the return of the field are responding to something real. As one of them told me: I have everything in my head already. Why would I spend months explaining it to people who do not have the context, when I can build it in a weekend?

What Is Collapsing

Start with what management is. Not what it does. What it is.

The discipline of management rests on a premise so foundational that it forgot it was a premise. Management is the practice of achieving outcomes through other people. Not through doing the work yourself. Not through any available means. Through people. This is what made management a discipline distinct from craft. The craftsman achieves outcomes through their own hands. The manager achieves outcomes by directing the hands of others. The entire discipline, its language, its structures, its hierarchies, its training programs, exists because human work required coordination, motivation, and judgment exercised across many people.

The pieces of management that have been dissolving in the agentic age, process ownership, signal collection, delegation, measurement, are not independent failures. They hang from the same through-line. The through-people premise. Remove it and the pieces have nothing to hold them together.

The agentic executive has removed the people from the through-line. Not from the organization. From the through-line. The employees are still there. The org chart is still there. The reporting lines are still there. But for the work the executive chose to do directly, the structural relationship that made people the means to the executive’s ends has changed. The executive now has a path from vision to outcome that does not pass through them.

If management is the practice of achieving outcomes through people, then the executive who achieves outcomes through agents is not managing. They are doing something else. Something we do not have a word for yet. The definition does not stretch to cover it. It breaks.

The break is not just felt. It has a structure.

Between the executive and the work, organizations built a layer of management. That layer exists because direct coordination does not scale. As organizations grow, the person at the top cannot reach the work alone. Managers translate strategy into operations, break vision into tasks, and mediate between intent and execution. This is how large organizations have functioned for over a century. The agentic executive can now reach the work directly, without that mediation. For the work the executive chose to do themselves, the layer between them and the outcome has become optional.

Frederick Taylor formalized the separation that made modern management necessary. Conception belongs to management. Execution belongs to workers. The entire bureaucratic apparatus of the twentieth century was built on this division. The agentic executive un-separates conception and execution. But only at the top. The senior leader recombines them in their own person. For everyone else in the organization, the separation persists. Conception still flows downward. Execution still happens below. Modern organizations softened this separation with agile teams, product ownership, engineering autonomy, and senior IC structures, but they did not erase it. Taylor still governs below. The executive has exempted themselves.

This is an asymmetric reversal of the foundational separation. Not a dissolution. A personal exemption. The executive becomes the person for whom the recombination of conception and execution is backed by the full authority of the organization.

Ronald Coase argued in 1937 that firms exist because coordinating work through markets has transaction costs. Inside the firm, hierarchy replaces the market as the coordination mechanism because it is cheaper. The firm’s boundary sits where internal coordination costs and external market costs meet.

The same logic applies inside the firm. The modern executive always had private context: private judgment, private networks, private intuition. What they never had was private execution capacity. Execution required the organization. Agents change that. They turn executive context into executable output without passing through the organization’s interpretive layers.

The agent does not move the boundary of the firm outward. It folds a new boundary inward. The executive’s transaction cost with their own agent is now lower than their transaction cost with their own organization. The hierarchy competes with a coordination mechanism that is faster, cheaper, and available at any hour. On one side of this new boundary, the executive and the agent. On the other, the rest of the organization. The firm is intact externally. Internally, it has a seam.

What the Organization Receives

The executive ships a feature. Not as a side project. Not as a weekend experiment. As part of how they now work. The building has become integrated into their regular operating rhythm, alongside the strategy meetings, the cross-functional reviews, the one-on-ones. The formal role has not been abandoned. It has been compressed to make room for the work the executive now does directly.

What does their organization receive?

The question is not whether senior leaders should build. They should. The distance between leadership and the artifact has grown too large in many organizations. Executive building can clarify strategy, expose bad process, raise ambition, and reconnect leaders to the reality of what their teams produce. The question is whether what the executive builds returns to the organization as shared context or bypasses the organization as private execution. The first strengthens the firm. The second hollows it out. The line between them is not always visible, and most organizations are not drawing it.

Start with the management layer between the executive and the work. It existed to mediate between strategy and operations. The agentic executive demonstrates, by their behavior, that they no longer need that mediation for some classes of work. The layer did not lose its job. It lost a specific function. Which classes of work the executive will keep mediating, and which they will increasingly do directly, has not been settled by anyone.

There is a measurement problem here. Organizations have spent decades building systems to track executive performance, team velocity, delivery timelines, product quality. None of these systems measure what happens when the executive steps outside the system itself. The organization is winning the metrics and losing the mission. The executive shipping fast registers as success on every dashboard that exists. The cost, the organization’s coherence, its shared context, its ability to function as a unit, is not on any dashboard that exists.

At the leading technology firms, the share of AI-generated code has grown rapidly. At one major firm, it moved from a quarter to more than 75% in under two years. The structural conditions for the agentic executive are not confined to a few exceptional leaders. They are present in every major technology firm. The phenomenon will scale. And the organizational questions it raises will scale with it.

One of those questions concerns the pipeline. Think about how today’s senior leaders became senior leaders. They built products through teams. They learned to translate vision into specifications that others could execute. They developed judgment through years of navigating the gap between what they wanted and what the organization produced. That gap was not inefficiency. It was the training ground. If the agentic executive skips the gap, the training ground disappears. The next generation of leaders does not have a path to the same judgment, because the path ran through the very process the executive no longer needs.

There is a deeper question underneath this one. The infatuation phase, the organizational stage where the capability itself becomes the object of attention, has a natural arc. It is a precursor, not a destination. Whether the agentic executive pattern follows the same arc, a phase that matures into something sustainable, or whether it represents a permanent reshaping of the executive role, is not yet known. The answer depends on whether organizations design for it deliberately or let the pattern run on its own logic.

The Diagnosis

Some teams see it happening. The executive builds something in days that the team had been planning for months. Some react with possibility. If the executive can do this, so can I. The behavior cascades. The agentic executive becomes a proof of concept for the organization.

Others see the same thing and feel the weight of it. The executive did not ask the team to build it. Did not give them the opportunity. The team learned about the result after it existed. The message, intended or not: I did not need you for this. Between peers, that demonstration is inspiration. From the person who decides whether you keep your role, it carries a different weight.

But not all of it is visible. In some cases the executive builds and the team does not know. The team continues working on something the executive has already built or decided to build differently. The standups continue. The planning meetings continue. The status updates continue. The structure looks intact. It is not. The team’s work may be rendered irrelevant without them ever knowing it was being duplicated in parallel.

The pattern does not stop at the individual executive. The solo agentic executive holds vision and execution alone. But executives inside the same division share the same strategic context. They see the same customers, the same competitive landscape, the same gaps in the product. They have the same agent capability. The natural next step is that they team up. The leadership layer of a division forms its own execution unit, building together with agents while the division underneath continues to operate as usual. Call it the super-team. The formal organization persists. Inside it, a parallel organization has begun to form. This is not a redesign. Nobody announced it. Nobody approved it. It is the trial from within. The conventional enterprise, testing whether its leadership can operate as something new while the old structure is still running.

The super-team raises a different risk than the solo executive. A single executive bypassing a team is a morale problem. The leadership layer collectively bypassing the organization it leads is a legitimacy problem. The division was staffed, budgeted, and structured on the assumption that it is the vehicle for the division’s work. When the leadership layer becomes its own vehicle, the division’s reason for existing is no longer self-evident. The people inside it have not been told that the premise has changed. They are operating under a contract that the leadership layer has quietly renegotiated with itself.

Four patterns. The visible and inspiring. The visible and heavy. The hidden. The collective. Each one is a different expression of the same structural shift. The through-people premise, the foundation that made management a discipline, is being sidestepped at the top. Sometimes openly. Sometimes invisibly. Sometimes by one leader. Sometimes by the leadership layer as a whole.

I have been using a phrase in conversations with leaders that tends to land hard. Managerial violence. Not violence by the person. The person may be acting with the best intentions. Violence in the structural effect on the people who are being bypassed, visibly or invisibly, by the leaders who were supposed to achieve outcomes through them. Managerial bypass would describe the mechanism. Managerial violence describes the experience on the receiving end. I am not certain it is the right term. But I have not found a better one, and the reaction it provokes tells me it is touching something real.

The executives drawn into this pattern are not choosing it in any deep sense. Jacques Ellul argued that technique, the logic of efficiency, selects for itself regardless of human preference. The agentic executive may be what technique selects for inside the conventional enterprise. The leaders most committed to their organizations are the ones most likely to find themselves here. They build because they can see the whole, because the agent is faster than the translation, and because the gap between their vision and the organization’s output has always frustrated them. The frustration was always there. The capability to act on it is new.

The Naming Problem

The agentic executive. The super-team. Two figures that did not exist a year ago. Both emerging inside conventional enterprises, not in startups, not in greenfield organizations, but inside firms whose structures, contracts, and expectations were built for the through-people model of management.

When the first automobiles appeared, people called them horseless carriages, defining the new thing by what it lacked rather than what it was. The word automobile came later. Whatever these executives are doing when they bypass their own organizations and build with agents, it is not management. Management is the practice of achieving outcomes through people. This is something else. It is not craft either. The executive is not simply doing the work. They are doing it from a position of authority over an organization that was built to do it for them.

A previous article in this series described the conventional enterprise being reshaped by forces larger than any single organization. The agentic executive and the super-team are how that reshaping becomes visible from the inside. They are not the endpoint. They are the trial. The conventional enterprise is testing, through these figures, whether it can transform into something that might be called the agentic-native company. A firm designed from the start for the world these executives are already operating in.

Whether that transformation succeeds depends on whether organizations design for it deliberately or let the pattern run on its own logic. The answers will not come from theory. They will come from the organizations living through it right now.

Sixth in a series that began with Your Management Model Is the New Bottleneck, continued with Management Is Losing Its Meaning, Beyond Productivity: Who Owns Your Skills?, Management at Machine Speed, and Nobody Measure the Building.

* * *

This article is different from the ones that came before it.

Over five articles, I have deconstructed what is happening to management in the agentic AI era. Each article examined a piece of what is changing. Each one stayed close to the ground: the manager, the team, the sprint, the dashboard. This article steps back.

I want to show you something larger. Not what is happening to management. What is happening to the organization itself.

The five changes I have described are not independent. They are not five separate disruptions arriving at the same time by coincidence. They are connected by a logic that is older than AI, older than the internet, older than the computer. It is the logic that built the modern corporation. And it is now rebuilding it.

To see that logic, we need to start with a sentence that every executive has heard, and nobody has questioned.

The Logic of Efficiency: Technology Versus Technique

No executive argues against “we should adopt AI efficiently.” The sentence sounds like common sense. But the word “efficiently” is doing all the work. It smuggles in an entire logic. And that logic is reshaping your organization in ways the sentence never announces.

To see how, we need a distinction that most organizations have never made. In 1954, the French philosopher Jacques Ellul published The Technological Society. The book made a distinction that most people, seventy years later, still have not absorbed. It is the distinction between technology and technique.

Technology is the machine. The tool. The platform. The thing you buy, deploy, and measure. Technology is visible. It has a name, a vendor, a price tag, and a rollout plan.

Technique is something else entirely. Technique is the logic of efficiency: the processes, rules, structures, metrics, and behaviors that form around a technology after it arrives. Technique is not the tool. It is what the tool does to the organization.

Ellul’s sharpest observation was that the two are constantly confused. People look at the machine and think they are looking at the problem. They are not. “The machine could not integrate itself into nineteenth-century society,” Ellul wrote. “Technique integrated it.” The steam engine sat in a room until someone built the factory system around it. The factory system was the technique. It reorganized labor, time, space, authority, and skill. The engine was just the engine.

The same confusion is happening now.

When an organization says it is “deploying AI copilots,” it is describing a technology. But what actually changes is not the copilot. What changes is everything that reorganizes around the copilot. Review loops shorten or disappear. Approval thresholds shift. Documentation patterns change. Escalation paths are redrawn. What counts as acceptable work changes. What counts as sufficient human involvement changes. Who needs to see what, and when, and why. All of it moves. The manager’s role changes without anyone redefining it.

The technology is the copilot. The technique is everything else. The technology is what you bought. The technique is what you got.

This distinction matters because technique is invisible in a way that technology never is. Technology has a launch date. Technique has no launch date. It seeps. It reorganizes quietly, one decision at a time, one process at a time, one expectation at a time. Nobody announces “we are now reorganizing our authority structures around AI.” It just happens. Each individual change is rational. The cumulative effect is a different organization.

And technique presents itself as unchallengeable. That sentence about adopting AI efficiently? Once efficiency is accepted as the organizing principle, technique takes over. Every subsequent decision follows from it. Not because anyone chose. Because the logic chose.

The corporation did not simply adopt technologies over the past 150 years. It was progressively reorganized by a logic that used those technologies as vehicles. Each technology arrived as a tool. Each departed as a new operating logic. That is the force underneath everything I have described in this series. Not AI. Not the tools. The logic that integrates them. To see where it has been, we need to look at how it entered the corporation, layer by layer.

How the Logic Entered the Corporation

The logic of efficiency did not arrive all at once. It entered the corporation one layer at a time, carried by successive technologies, each one reaching deeper than the last.

Sometimes the translation of technology into organizational method had a recognizable architect. Sometimes it emerged through a cluster of doctrines, institutions, and management fashions. What matters is not who designed the translation. What matters is the pattern: a technology arrives, the logic converts it into an operating method, and that method reorganizes a layer of the corporation that was previously governed by human judgment, habit, or craft. Each time, the logic reaches deeper. Each time, the space for human variability shrinks.

Production. The factory was the technology. Taylor systematized the technique: scientific management. Study work. Break it into processes. Assign each process to the right worker under the right supervision. Measure the output. Optimize the sequence.

Before Taylor, production was personal. Ellul makes this point with precision: the skill of the worker compensated for the crudeness of the tool. The deficiency of the instrument was offset by the expert eye, the practiced hand, the professional know-how. Everything varied from person to person according to their gifts. That variation was not a bug. It was the system.

After Taylor, production was systematic. The logic of efficiency entered the production floor. The human skill that mattered shifted from craft to compliance with process. The worker’s gifts became less important than their adherence to the method. The method was the technique. It sat on top of the technology and told the organization how to use it.

Structure. Scale was the technology. Railroads, the telegraph, and national markets made the large corporation possible. But possible is not operational. A company that spans a continent cannot be run by one person who knows everyone. It needs a structure.

Sloan institutionalized the technique at General Motors: the multidivisional form. Decentralize operations. Centralize financial control. Each division runs its own business, but headquarters governs through standardized reporting, capital allocation, and performance metrics. The logic of efficiency entered organizational design itself.

Something important happened here. The corporation was no longer a collection of people. It was a structure with its own geometry. The boxes on the org chart existed independently of who occupied them. People could leave, be replaced, rotate through positions, and the structure persisted. The entity had begun to develop a form that outlasted any individual inside it.

Cognition. The computer was the technology. Drucker named the shift. He identified the rise of knowledge work and the centrality of information and decision-making as productive functions. He gave middle management its intellectual justification: not as supervisors of labor but as processors of organizational intelligence. Management by objectives systematized this further.

The logic of efficiency entered the cognitive layer. It was no longer enough to organize bodies on a production line or boxes on an org chart. Now the organization of thought, information flow, and decision-making was subject to the same logic. How people think inside the corporation became a management problem. And management problems get optimized.

Middle management expanded dramatically in this era. Not because organizations were inefficient. Because the logic demanded a human information-processing layer between strategy and execution. Drucker saw this clearly. The middle manager existed because the corporation’s nervous system had to run on human cognition. There was no other option.

Interface. The internet was the technology. The technique was less architecturally coherent. Platform economics. Network effects. Agile methodology. Digital transformation. No single thinker. A cluster of doctrines, consultancies, and management fashions, each promising to unlock what the technology made possible.

The logic of efficiency entered the boundary between the organization and its environment. The customer interface. The speed of iteration. The data layer. The organization became porous, interconnected, always on. It could sense its market in real time, respond faster, iterate continuously. The boundary between inside and outside blurred.

But notice what stayed constant through all four layers. Every technology carried the logic deeper into the corporation. The factory reorganized human labor but still required human hands. The divisional structure reorganized human coordination but still required human judgment to run each division. The computer reorganized human cognition but still required human minds to process the information. The internet reorganized the organization’s interfaces but still required humans to build, ship, and decide.

The logic went deeper each time. But it always stopped at the same floor: the human mind.

AI does not stop at that floor. It does not add a new organizational layer alongside the ones that came before. For the first time, it offers the corporation an alternative to the human substrate that every previous layer depended on.

This is the point of discontinuity. Not a new layer. A new medium.

The Paradox: Why It Will Not Simplify

At this point, the reader might expect a familiar conclusion. The logic of efficiency has deepened through five layers. AI reaches the substrate. The corporation should get simpler.

It will not.

In 2015, the anthropologist David Graeber published The Utopia of Rules. His argument was uncomfortable and precise. Bureaucracy did not live only in the state. It hid inside the corporation as well, where the same administrative work acquired a different name and a more respectable suit. When Americans spoke of the Organization Man, the soullessly conformist corporate functionary, they were describing bureaucrats, even if nobody called them that. The word “bureaucrat” attached itself to civil servants. Corporate middle managers doing identical work were called managers.

Graeber’s deeper point was about what rationalization actually produces. Every attempt to make an organization more efficient generates more administration, not less. Scientific management promised to simplify the factory floor. It produced time-study departments, planning offices, and supervisory hierarchies. Business process reengineering promised to flatten organizations. It produced consulting engagements, implementation teams, and change management offices. Lean and Agile promised the same thing in different language. They produced their own coordinators, coaches, frameworks, and certification programs.

Each wave promised reduction. Each wave produced new roles, new processes, new governance, new measurement systems. The total administrative burden never decreased. It transformed. I wrote about this in 2018 when DevOps promised to abstract away IT operations. What it actually produced was a new class of engineers, a new set of tools that needed their own management, and more organizational complexity in every layer. The pattern was visible then. It is visible now.

Efficiency applied to an organization does not produce simplicity. It produces a more complex organization that is more efficient at being complex.

AI is repeating the pattern. Organizations adopt agentic AI to eliminate overhead. They are generating new overhead. AI governance councils. Red-teaming procedures. Human-in-the-loop attestations. Model evaluation pipelines. Responsible AI review boards. Prompt libraries. Policy registries. Agent oversight workflows. Exception escalation paths. Output verification checklists. Some of these functions existed before generative AI. What is new is their proliferation, their visibility, and their centrality inside ordinary corporate workflows.

Graeber put it sharply: “Talking about rational efficiency becomes a way of avoiding talking about what the efficiency is actually for.” The language of efficiency crowds out the question of purpose. Every organization pursuing AI adoption talks about efficiency. Faster cycles. Lower cost per unit. More output per person. Nobody asks what the efficiency is for. Not because the question is unimportant. Because the logic does not permit the question. Efficiency is the answer before anyone asks. In the previous article in this series, I argued that our dashboards have become cave walls, showing us shadows of activity rather than the substance of value. This is why. The metrics measure efficiency. They cannot measure purpose. And the logic ensures nobody notices the difference.

This is the paradox that sits between the historical deepening and what comes next. The administration does not go away when you automate it. It changes form.

AI does not remove the bureaucratic logic. It furnishes it with a better medium.

The Substrate: What the Corporation Runs On

Before the corporation, business was personal. A merchant. A family. A guild. If the owner died, the business often died with them. Liability was personal. Risk was personal. Knowledge lived in a person’s head and hands.

The corporation was invented to solve a specific problem: how do you undertake ventures that exceed a single human life? The answer was a legal fiction. A group of people created an entity that could outlive any of them. It could own property. It could enter contracts. It could sue and be sued. The word comes from corpus. A body. A legal body that was not a human body.

It was a convenient abstraction. And it worked. By the twentieth century, the corporation had become the organizing structure of modern economic life.

But something happened to the abstraction along the way. It developed a trajectory of its own.

A corporation could act against the interests of its own employees. It could pursue goals that no individual within it would choose. It could outlive every person who created it, replace them all, and continue unchanged. Over time, the entity developed something that resembled a will. Not a conscious one. An emergent one. A trajectory that arose from incentive structures, market pressures, legal obligations, institutional momentum, and the logic of efficiency that governed its internal operations.

No single person decided what the corporation wanted. The corporation’s behavior emerged from the system. And the system had its own direction.

But there was a constraint. That emergent trajectory, however autonomous in tendency, always operated through humans. Every decision required a person. Every evaluation required a person. Every act of coordination, translation, escalation, and judgment required a person. The entity needed humans as its operating medium. Not because it valued humans. Because there was no alternative substrate.

For 150 years, humans were the only available technology for the functions the corporation required: processing information, making evaluations, coordinating activity, translating strategy into action. The corporation ran on people. All of its logic, all of its efficiency, all of its autonomous tendency had to pass through human cognition on its way to becoming action.

This is what kept humans at the center. Not indispensability in the philosophical sense. Indispensability in the engineering sense. The system could not run on anything else.

That constraint is breaking.

For the first time, large parts of the corporation’s evaluative, coordinative, and informational work can be executed by nonhuman systems at scale. Agents process information continuously. Agents evaluate outputs. Agents coordinate workflows across systems. Agents translate specifications into execution. The functions that Graeber identified as the core of what bureaucrats do, assessing, auditing, measuring, weighing the relative merits of plans, proposals, and courses of action, no longer require human cognition as their exclusive medium.

This is the direction. In organizations adopting agentic AI, work is beginning to be reviewed, routed, summarized, evaluated, and escalated by systems before a human encounters it. The manager who once performed these functions increasingly receives pre-filtered, pre-evaluated, pre-summarized inputs. Their judgment is not replaced. It is narrowed. The scope within which human variability is permitted shrinks with each iteration. The corporation becomes more purely what the logic of efficiency always demanded.

Ellul saw this seventy years ago. “Man must have nothing decisive to perform in the course of technical operations,” he wrote. “After all, he is the source of error.” The elimination of human variability was always technique’s goal. Not because technique is hostile to humans. Because variability is inefficiency. And the logic does not tolerate inefficiency.

Ellul described what happens when technique encounters an organic obstacle. In one of the most revealing passages in The Technological Society, he traces the history of mechanized bakeries. The machines could not handle bread as a living substance. The dough resisted standardization. More subdivisions, more intervals, more precautions were required in the mechanized bakery than in the non-mechanized one. The machines did not save time. So the industry changed the bread. “Whenever technique collides with a natural obstacle,” Ellul wrote, “it tends to get around it either by replacing the living organism by a machine, or by modifying the organism so that it no longer presents any specifically organic reaction.”

The bread was modified to fit the machine. The machine was not modified to fit the bread.

When agentic AI collides with human judgment, the same logic applies. Technique does not preserve judgment. It either replaces the human or modifies the human until they no longer exercise independent judgment. The scope of decisions narrows. The range of acceptable variation shrinks. The human learns to work within the boundaries the system sets rather than the boundaries their own judgment would set.

This is what deskilling and never-skilling actually are. Not side effects of AI adoption. Not accidental losses. They are technique completing its project. Each skill that erodes is a reduction in human variability. Each skill that never forms in a new hire is one less source of the friction that technique cannot tolerate. The deskilling I described in an earlier article, the never-skilling I named, these are not failures of the AI transition. They are the transition working exactly as the logic demands.

Ellul again: “The ultimate success of mechanization turned on the transformation of human taste.” Not the transformation of the machine. The transformation of the human.

The ultimate success of agentic AI will not turn on better models or faster agents. It will turn on the transformation of human expectations about what work is. What judgment means. What skill consists of. What your role inside the corporation should be.

The corporation is not being transformed by AI. It is being completed by it.

The Three Forms: What the Corporation Is Becoming

In 1937, the British economist Ronald Coase asked a question that nobody had thought to ask: why do corporations exist?

If markets are efficient, why don’t individuals simply contract with each other for every task? Why organize people into a permanent entity with hierarchies, employment contracts, and fixed structures? Why build a corporation at all?

Coase’s answer was transaction costs. Finding the right person, negotiating terms, monitoring performance, enforcing agreements. All of this takes time and money. When the cost of coordinating through the market exceeds the cost of coordinating internally, it makes sense to bring the activity inside the firm. The corporation exists because it is cheaper to organize some activities through hierarchy than through open markets.

This is not abstract theory. This is the reason your organization has the shape it does. Every function that sits inside your company is there because, at some point, someone determined it was more efficient to do it internally than to buy it externally. Every function that sits outside, outsourced, contracted, partnered, is there for the opposite reason. The boundary of the firm is an economic line drawn by coordination costs.

For years, organizations have been pursuing something called digital transformation. The word promises a change of form. Trans-form. I have long advocated for taking that word seriously. True digital transformation should change the depth and breadth of the organization. It should challenge the internal structure, not just the tooling.

And there are real success stories. Banking is the clearest example. Fully digital banks operate today with no branches, no cashiers, no physical infrastructure for customer interaction. What a bank is has changed.

But the digital banks that achieved this are not transformed traditional banks. They are new entities, built from scratch on digital logic. The traditional banks, some of them hundreds of years old, tried to transform and could not. They added digital layers. Mobile apps on top of branch networks. Online portals on top of paper processes. Each layer accumulated alongside the existing structure rather than replacing it. Too much internal friction. Too little leadership willing to challenge the form itself. The result was more administration, not less. A digital layer on top of a physical layer, each requiring its own governance.

The technology could reach the tools. It could not reach the form. And when the form resisted, the logic did not wait. It created new entities alongside the old ones.

Agentic AI reopens this question at a deeper level. Not because it is a better digital tool. Because it changes the coordination economics on both sides of Coase’s boundary.

Inside the firm, agents take over functions whose primary purpose was coordination, translation, and review. Information processing that required layers of middle management can run with fewer people and simpler structures. That part is intuitive. Most executives already see it.

The less obvious part is outside the firm. Agents can search for suppliers, evaluate contractors, monitor performance, negotiate terms, and manage relationships across organizational boundaries at machine speed and minimal cost. The friction that made external coordination expensive, the reason you brought functions inside the firm in the first place, is dropping. When both internal and external coordination costs fall simultaneously, the boundary of the firm comes under pressure from both sides.

Three forms are emerging side by side. These are not stages of one journey. They are different responses to the same economic shift.

The first is the traditional corporation. Adding AI layers on top of existing structure. Getting more complex without changing shape. AI governance on top of existing governance. Agent oversight alongside human oversight. New layers, same form. The traditional bank, replayed across every industry.

The second is the digital native, built from scratch on the new logic. No legacy processes. No internal friction. No inherited form to fight. These organizations are designed around AI from the beginning, the way the digital bank was designed around digital from the beginning.

The third is something else entirely. The agent-native. Not a smaller version of an existing form. A new form. An entity where agents are not tools used by humans but structural participants in the organization itself. We do not yet have a clear example. But the economics point in that direction. If agents can evaluate, coordinate, and execute at scale, the question is not just how many people the corporation needs. It is what kind of entity emerges when the operating logic no longer assumes humans as its primary substrate.

If you are an executive reading this, you are most likely running the first kind. You are competing against the second. You may not yet see the third. And the distance between them is growing. Not because your organization is failing. Because the economics that gave your organization its shape are changing, and the form has not changed with them.

This is what transformation actually means. Not new tools running old structures. A change in the economic logic that determines the corporation’s shape. Which functions stay inside. Which move outside. How many people the entity needs. Whether the boundary of the firm holds, dissolves, or becomes something else entirely.

The question is whether the organizations that already exist will reshape themselves, or whether, as in banking, the new forms will simply grow up alongside them. That question is not theoretical. It is the question behind every AI strategy, every restructuring, every budget decision you are making right now. You are drawing the boundary of your firm. You may not know it.

* * *

Everything I have described in this series is a symptom of this logic.

Management is dissolving because the logic has found a more efficient medium for coordination, evaluation, and control. Skills are being extracted because technique requires the reduction of human variability. The three disciplines I introduced, protecting cognition, governing tempo, governing skills, are not management improvements. They are the deliberate design of the human role inside a system that will, by default, shape that role around its own logic. And measurement is broken because, once efficiency becomes the governing language, the metrics begin to displace the reality they were meant to describe.

What looked like separate observations across five articles is one logic, playing out across every dimension of the organization simultaneously. The parts were always connected. This article is the view from far enough away to see the whole.

None of this was decided. It emerged.

You have spent your career inside this entity. This article was an attempt to show you its shape from the outside. What you do with that view is yours.

You sit down in front of your dashboard. Monday morning. Your velocity is up 40%. AI adoption across your teams is at 96%. Tickets closed this sprint doubled from last quarter. The token consumption number is staggering: 1.2 billion tokens consumed this sprint. The board is green. Everything is moving. You have never had more data. You feel informed. You feel in control.

You are watching shadows.

In Plato’s allegory, prisoners are chained inside a cave, facing a wall. Behind them, a fire casts the shadows of objects onto that wall. The prisoners have never seen the objects themselves. They study the shadows. They name them. They build entire systems of understanding around them. The shadows are not lies. They are real projections of real things. But they are flat, depthless, and stripped of the dimensions that would reveal what is actually happening.

Your dashboard is the wall. Your metrics are the shadows. And the chains are the measurement systems you inherited from a world where humans did the work, humans set the pace, and tracking human activity was a reasonable proxy for tracking value.

That world is gone.

The DORA 2025 report found that AI adoption now correlates with higher throughput and higher instability simultaneously. Two signals on the same dashboard. One green, one red. Speed up, stability down. Most organizations focus on the green one. The shadows are dancing faster. The floor is shaking. And the metric that gets the executive presentation is the one that looks like progress.

Meanwhile, behind the dashboard, Simon Willison, one of the most respected software engineers working with AI today, put it plainly: using coding agents well takes every inch of his 25 years of experience, and by 11am he is wiped out for the day. The dashboard shows unprecedented speed. The human cannot sustain it past mid-morning. That gap is the cave in one image.

The Story So Far

This series has argued that agentic AI has made management the bottleneck (Article 1), that management’s foundations are collapsing across five layers (Article 2), that AI extracts human skills through deskilling and never-skilling (Article 3), and that three disciplines survive the transition: protecting cognition, governing tempo, and governing skills (Article 4). In that last article, I introduced the concept of ghost notes: value that lives in dimensions that current metrics were never designed to capture.

This article is about building the instruments to hear them. Because the dominant system of metrics is structurally blind to the dimensions that matter most.

The Bricklayer and the Robot

Here is the simplest way to understand why your metrics are broken.

For a century, work looked like this: a human performed a task, and a manager measured the human. Take the simplest version of this. A bricklayer lays bricks. How many? How fast? How accurately? Frederick Winslow Taylor built an entire management science around this. Time the worker. Optimize the motion. Measure the output. The primary resistance to Taylorism was about its dehumanizing quality, not its measurement logic. If humans do the work, measuring human activity is a reasonable proxy for measuring value.

Now replace the bricklayer with a robot. The robot lays bricks at machine speed. The human’s job is no longer to lay bricks. It is to decide which wall to build, write the instructions the robot will follow, and verify that the wall is straight. The work still gets done. Bricks still get laid. But the unit of value has fundamentally changed. It is no longer bricks per hour. It is the quality of the instruction that determined which wall to build and whether the wall serves its purpose.

If you keep measuring bricks per hour, you will conclude that productivity has gone through the roof. You will be right. And you will have no idea whether the walls are in the right place.

This is exactly what is happening in every organization that measures AI adoption by velocity, tickets closed, or tokens consumed. The production function moved. The measurement system stayed behind.

Tokens are the most seductive shadow of all. Hundreds of millions of tokens consumed. Billions. The number is big. It is growing. It looks impressive in executive presentations. But tokens are cost, not value. Measuring token consumption and calling it productivity is like measuring the electricity bill and calling it output. An agent burning through millions of tokens on a poorly specified task is the most expensive way to produce nothing. And it looks exactly like progress on the wall.

Jensen Huang said at GTC: “If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.” This is the Cobra Effect. Offer bounties for dead cobras, and people breed cobras. Measure and reward token consumption, and teams optimize for consumption. The metric rewards activity, not value. The more you measure tokens, the more tokens get burned, and the further you get from understanding whether any of it produced anything worth having.

The production function changed. The metrics did not.

The Questions That Matter Now

The fact that it is cheap to build does not make it worthwhile building. When agents can execute anything, the question shifts from “can we build it” to “should we.” Every manager, in every era, moves through four phases of judgment: deciding what to build, governing how it is built, evaluating whether it worked, and steering what comes next. The questions inside each phase have not changed. The ability to answer them has collapsed.

Ask your most senior engineering leader these questions. Ask them right now. Most cannot answer more than one or two with confidence. Not because they are bad leaders. Because the measurement system they rely on was never designed to answer them.

What should be built?

What is my team actually building right now? Not “what tickets are in progress.” Not “what is the sprint velocity.” What is actually being built? What specifications are being written? What are agents executing against? The current measurement system gives you a shadow of this: a board full of moving tickets. It has the shape of an answer without the substance of one.

Is it the right thing for the customer? Visibility is not alignment. You may know what your team is building and still have no idea whether it is what the customer needs. When execution was expensive, the cost of building the wrong thing was the wasted effort. When execution is cheap, the cost is misdirection. You built fast, you built wrong, and you lost the time you could have spent building right.

How should we build it?

Is it being built the way we agreed? Agents execute at machine speed across parallel streams. A team can drift significantly from the plan before anyone notices. The specifications that initiated the work may have been clear. What the agents are actually executing against may be something different entirely. Without a way to measure conformance to intent, you have activity without accountability.

Did we build the right thing?

What was the actual impact? Not “how many story points did we close.” Not “how many lines of code were committed.” What value reached the customer? What did it cost to produce that value?

The ultimate output metric, the customer retention rate, the revenue number, arrives months after the work was done. By the time you see it, it is too late to steer. You have already built the wrong wall. This means you need input metrics that reliably predict those late outcomes. Early signals you can read while the work is still in progress. And you need to validate, empirically, that those early signals actually correlate with the late outcomes. That is the real measurement challenge.

What next?

Can I steer what they build next? Steering requires knowing whether your people have the judgment to decide what to build next and whether they are developing the skills to keep making those decisions. But steering is not only about what comes next. It is about changing direction on what is already underway. A customer’s situation changes. A market shifts. You need to understand that shift and redirect what you are building, while agents are executing across dozens of parallel streams at machine speed. Most organizations today could not answer that question for a single team. You cannot steer what you cannot see.

These are not hypothetical concerns. Faros AI studied over 10,000 developers across 1,255 teams and found that individual engineers were completing 21% more tasks and generating 98% more pull requests. Meanwhile, review time increased by 91%. Individual productivity rose while the organizational capacity to absorb that productivity became the constraint. The productivity dashboard said the teams were faster. The bottleneck had simply moved one level up, to the humans who had to make sense of everything the machines produced.

What to Measure Instead

An F1 race engineer does not measure fuel consumption and call it performance. They measure tire degradation, brake temperatures, sector times, and the gap to the car ahead. Article 4 introduced the F1 metaphor for the agentic era: work has shifted from driving to racing, from linear delivery to continuous cycles at machine speed. The five categories below are the telemetry system for the human-machine interface.

These categories are not theoretical. They are grounded in operational experience. They are not the only way to measure the new production function. But they cover the dimensions that current metrics structurally miss.

The new measurement system must work at three levels: the quality of human input to agents, the quality of human oversight of agent outputs, and the behavior of the human-agent system over time. The five categories span all three.

1. Specification quality. Measure the inputs, not just the outputs. How clear, complete, and actionable are the instructions that humans give to agents? A well-specified task that an agent executes correctly on the first attempt is fundamentally different from a vaguely specified task that requires three rounds of correction. Both produce a completed ticket. Only one represents effective human-machine collaboration. The metric is not whether the work got done. It is whether the specification was good enough to get it done right the first time.

2. Judgment and verification. Measure the human’s ability to evaluate what agents produce. This is the hardest category and the most important. Can your reviewers catch errors in agent-generated outputs? How often do defects escape review? Is the review process rigorous or performative? Research from Princeton found that in over 21,000 agent rollouts, higher reasoning effort actually reduced accuracy in 21 of 36 tested settings. Agents took shortcuts and made catastrophic errors. The question is whether your humans would catch them.

3. Cost of correctness. Measure what it takes to arrive at a right answer, not just whether you arrived. This is not about efficiency in the traditional sense. It is not a utilization metric. It is about the cost structure of the new production function. Stanford research applied production theory to AI coding and found that lightweight models were most cost-effective for basic tasks while reasoning models were necessary for complex ones. Google, UCSB, and NYU researchers found that budget-aware agent configurations cut search calls by 40% and costs by 31% with comparable accuracy, while simply granting larger budgets did not improve performance. An agent that burns through ten times the resources to reach the same answer is not productive. It is expensive. And if engineers are measured on token consumption, the incentive inverts. They will choose configurations that consume more, not configurations that produce better results.

The first three categories measure the human-agent interface directly. The next two measure the broader system those interfaces create.

4. Architectural coherence. Each bricklaying robot builds a perfect wall. But who designed the building? Who ensured the walls connect, the load-bearing structure works, the plumbing runs through the right places? When every individual human-agent loop optimizes locally at machine speed, the architecture is where failures surface. Code that passes every unit test but fails at integration. Features that are individually correct but structurally incoherent. Documents that are locally polished but globally contradictory. The productivity dashboard shows every component as green. The system does not work. This is both a management problem and an architecture problem. Someone must hold the integrative view across all the parallel streams of agent-produced work. Someone must be accountable not for the walls but for the building.

5. Sustainability of human capability. Measure whether the humans in the system are getting stronger or weaker over time. This is where cognitive load, burnout, and skill erosion converge. Is your team’s judgment improving through practice, or degrading through delegation? Can your junior engineers debug without agents, or have those foundational skills never formed? Can your team still do independently what they could do a year ago? Research shows the average focused work session has dropped to just 13 minutes. Among AI users experiencing cognitive fatigue, intent to leave rises by nearly 10%. The sustainability of the human is not a wellness metric. It is an operational one. When the human degrades, the quality of specification, verification, and judgment all degrade with it.

These five categories form a system. If specification quality drops, the agents produce worse outputs, which increases the burden on judgment and verification, which accelerates cognitive fatigue, which further degrades specification quality. The loop tightens until something breaks. You do not manage quality by inspecting outputs anymore. You manage it by managing the humans who direct and evaluate the agents.

What This Demands

Measuring different things is not the hard part. The hard part is accepting what the new measurements require.

If you measure specification quality, someone must be accountable for improving it. That is a management function. If you measure judgment and verification, someone must govern which reviews are rigorous and which are performative. That is a management function. If you measure sustainability, someone must decide when to slow down even when the machine can go faster. That is a management function.

Every one of the five categories above requires a manager who understands what they are measuring and why. Not a dashboard owner. Not a metrics analyst. A manager who governs the human-machine interface.

This is why the ROI is not showing up. Organizations invested in the machine side of the equation. Nobody invested in the management transformation required to capture the value. That is like buying an F1 car and skipping the driver program, the race engineers, and the telemetry systems. The car is fast. The results are terrible.

Here is the line I want you to sit with.

This is not change management. This is change of management.

The management model itself is the thing that must change. Not the tools the managers use. Not the cadence of their meetings. Not the dashboards they review. The fundamental definition of what they manage, how they measure it, and what they are accountable for.

The organizations that figure this out will not just find the AI ROI everyone is looking for. They will build something more valuable: the capability to keep finding it as the technology evolves.

The Ship of Theseus asks: if you replace every plank of a ship, is it still the same ship? If you replace every metric, every process, every definition of what management means, is it still management? The answer matters. Most organizations have not asked the question yet. They are still watching shadows.

Across the industry, from boardrooms to engineering all-hands, the same question is surfacing: where is the AI ROI?

Companies are spending millions of dollars to equip their engineers with the best and most advanced AI tools. Many billions of tokens are being consumed every month. The cost of running software is increasing fast. This time, the investment is not a technology expense. It is a workforce expense. Unlike every previous technology cycle, AI is not competing for the IT budget. It is competing with the labor budget, replacing and augmenting what people do. But the returns are not showing up where leadership expects them: in faster delivery, higher quality, competitive advantage, or increased revenue. The costs are rising. The value is not rising with them. The dashboards show activity. They do not show value. The value is there. But it lives in the ghost notes, the dimensions that current metrics were never designed to capture, and that leaders do not yet know to ask about.

Here is what the dashboards also do not show. On the ground, something very different is happening. Teams are energized. Developers are shipping features that used to take weeks in a matter of hours. Engineers are running multiple agent threads in parallel, reviving long-deferred projects, tackling work they never had the capacity for. Senior technology leaders are personally building prototypes with tools like Claude Code and posting about the experience with visible excitement. The energy is real. I see it in every organization I work with.

These two signals seem contradictory. Missing ROI at the top. Genuine excitement at the bottom. They are not contradictory. They are two symptoms of the same root cause.

Management has not evolved to match the technology.

The ROI will come. But not while organizations are still in the infatuation phase. Today, companies measure their engineers’ token consumption internally. Cross-industry leaderboards compare how much AI organizations consume. Jensen Huang said at GTC: “If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.” But consumption is not value. The ROI will show when organizations move from infatuation to industrialization: when the processes, methods, structures, and management are in place to leverage these new capabilities, not just deploy them. This article is about that evolution.

The Adrenaline Phase

A recent study from Berkeley Haas, published in HBR, found that AI tools do not reduce work. They consistently intensify it. Employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day, often without being asked to do so. They did this because the tools created a sense of momentum, a feeling of having a “partner” that could help them power through their workload.

That momentum is real. It is also unsustainable. But not immediately. The cognitive cost accumulates over weeks and months, invisible until it surfaces as errors, disengagement, or resignation. This is precisely why management matters. By the time the symptoms are visible, the damage is already done. Someone has to be watching before the crash, not after it.

Separate research from BCG and UC Riverside identified what they call “AI brain fry”: cognitive fatigue that occurs when AI use exceeds a person’s ability to process the volume of information and decisions it generates. It produces mental fog, difficulty focusing, slower decision-making, and information overload. 14% of AI-using employees reported experiencing it. Among those who did, intent to leave rose by nearly 10%. And a 2026 workplace study found that the average focused work session has dropped to just 13 minutes, down 9% in two years.

There is a positive signal here too. For years, employee engagement has been in continuous decline. AI tools have the potential to reverse this by absorbing the repetitive work that drains motivation and freeing people to focus on work that energizes them. That is a real and valuable shift. But without deliberate management, the same tools that re-engage people by removing drudge work will overload them by increasing the pace and volume of everything else. The upside and the downside come from the same source. The question is whether anyone is governing the balance.

The executives communicating their AI excitement in all-hands meetings and industry events are not wrong about the technology. They are wrong about what their personal experience predicts for the organization. A senior leader spending a Saturday afternoon building a prototype with an AI coding assistant is having a personal experience. It tells them nothing about what happens when 500 people operate at that intensity for twelve months. One person’s exhilaration on a Saturday is not an organizational strategy. But it becomes one when that leader walks into Monday’s all-hands and broadcasts the excitement as the new standard. Every manager down the line reads the signal: speed is the expectation. Anyone raising concerns about sustainability is swimming against the current.

The adrenaline phase looks like transformation. It is the precursor to a crash. And the organizations that mistake one for the other will lose their best people first.

Why Management Was Never Ready For This

Before AI arrived, management was already running on fumes.

Gallup has spent decades documenting the failure. Most companies promote workers into management because of tenure or performance in a non-management role, not because they have management talent. Only one in ten people possess the inherent ability to manage. The result: organizations fail twice. They lose their best individual contributor. They gain a bad manager. This costs the U.S. economy between $319 and $398 billion annually.

But the failure goes deeper than selection. Most managers never formally learn what management is. There is no structured curriculum, no defined competency model, no progression framework. They learn on the job, which means they imitate. They imitate their own manager, who was also promoted without training. They draw from articles, from cultural mythology about what leadership looks like, from MBA programs still teaching frameworks built for a world that no longer exists. The entire management training infrastructure, from universities to corporate leadership programs, remains ill-equipped for this age. The signal degrades with every copy. By the time it reaches today’s manager, the operating model is a patchwork of inherited habits, half-remembered advice, and instincts that may or may not apply.

Management today consists of more than a dozen distinct disciplines. Most managers could not name them. Most organizations have never made them explicit. You cannot transform what you have not defined.

Some organizations sense that something is missing. Across the technology industry, companies are creating new roles like AI Transformation Lead, charged with driving adoption and accelerating speed. But these roles are aimed at the tool side of the equation, not the management side. They are not equipped to handle the organizational and management transformation that the technology demands. It is another instance of solving for the machine while ignoring the human.

And now the speed has changed by orders of magnitude.

From Road Speed to Race Speed

Here is the metaphor that clarifies what is happening.

An F1 car and a regular car both put a human behind the wheel of a machine. But the similarity ends there. A regular car takes you from one place to another. An F1 race is circular. The driver ends exactly where they started. The point is not the destination. It is sustained performance under extreme conditions, lap after lap, at intensities where the human body becomes the limiting factor, not the machine. The car can corner faster than the driver’s neck can sustain. It can brake harder than the driver’s cardiovascular system can handle. The entire engineering effort around an F1 team is not about making the car faster. It is about managing the human-machine interface at extreme intensity.

The agentic era is the moment when work shifted from driving to racing. From linear delivery (start a project, execute, ship) to continuous cycles of production at machine speed, where the work never stops and the human must sustain performance across every lap. The machine can go faster. The question is whether the human can last.

In F1, this is solved by design. The race engineer filters information so the driver’s cognitive bandwidth is protected. The team governs pace, telling the driver when to push and when to conserve, because maximum intensity for the full race distance destroys the tires. The car is designed around the driver’s constraints, not the other way around. And the driver must maintain fundamental skills even as the car’s systems take over more functions, because a driver who loses feel for the tires or the braking points becomes dangerous.

Most organizations are doing the opposite. They are asking humans to match the machine’s pace. That is the equivalent of running qualifying laps for 58 laps. It works for the first few. Then the tires degrade, the brakes overheat, and either the car fails or the driver makes a catastrophic error. Even with the best race management, F1 drivers sometimes finish so physically depleted that they cannot get out of the car without assistance. These are the fittest athletes in motorsport, trained specifically for this intensity. Your engineers are not. And unlike the driver, they do not get a two-week break between races. They do it again Monday morning.

The fans, by the way, recently pushed back against F1 going electric. They want the roaring sound of the engines. Organizations have their own version of this: the attachment to standup meetings, sprint reviews, performance ratings, and org charts with reporting lines. These are the roaring engines. They feel like management. They are not where the performance comes from anymore.

The Three Disciplines That Matter Now

Most of the traditional management disciplines are being absorbed by agents. Task management, resource allocation, budget tracking, schedule management, routine communications, process execution. Agents do these well and will do them better. This is not a prediction. It is already happening. As one reader commented on my second article: “Three months ago I stopped using Jira. I just work with an agent and the code directly. No need for a middleman. Not a SaaS middleman, and soon, no human middlemen.”

What remains is what agents cannot do. Three disciplines that are not just surviving the transition but are becoming the entire point of management.

Protecting cognition. The old manager protected budgets, timelines, and deliverables. The new manager protects the human brain. When your team is running multiple agent threads, producing output at machine speed, and experiencing the cognitive load that research now documents as brain fry, the manager’s primary job becomes sustaining human performance. This operates at two levels. At the individual level, it means monitoring cognitive load, governing the number of parallel workstreams, and creating space for deep focus in a world where the average focused session lasts thirteen minutes. At the team level, it means preserving coherence. Each person is in their own human-agent loop. Shared context evaporates when everyone is producing at a pace that makes it impossible to track what teammates are doing. The manager creates the synchronization points that keep the team functioning as a unit rather than fragmenting into parallel streams.

Governing tempo. Work now moves at machine speed. The old constraint was scarcity: not enough people, not enough time, so you prioritize ruthlessly. The new constraint is abundance. Agents can do everything. The question becomes what should we not do. The manager becomes a constraint by design, not by default. They are the person who says: we could pursue twenty initiatives this cycle because the agents can handle the execution, but we are doing six because that is what the humans can meaningfully specify, review, and learn from. This is the F1 race engineer telling the driver to conserve on lap twelve. Not because the car cannot go faster. Because the race is won over fifty-eight laps, not five. Organizations that push human hours to match machine speed, the 996 culture now spreading in Silicon Valley, are running qualifying laps for the entire race distance. The research is clear on where this ends: burnout, turnover, degraded decision-making, and the very brain fry that erodes the quality of human judgment that AI depends on.

Governing skills. In my previous article, I argued that AI extracts skills through deskilling and never-skilling. That extraction happens in real time, inside the daily work. Every time an employee delegates a task to an agent without understanding what the agent did, a skill degrades. Every junior hire who never learns the fundamentals because agents handle them is a never-skilling casualty. This is not something HR can manage with quarterly training programs. It happens in every task, every day. The manager’s role is not to approve each delegation decision, but to set the framework that governs them: which capabilities the team must maintain, which skills are strategic assets, and where human practice is non-negotiable. Without that framework, the default is to delegate everything, and the skills quietly disappear. At the team level, this means governing the collective skill portfolio. If one person’s judgment in a critical domain erodes because they have delegated too much to agents, the entire team is exposed. Skill governance is an operational management responsibility now, not an HR function.

These three disciplines are not independent. They form a reinforcing system. If you fail to protect cognition, the quality of specification and review degrades. If you fail to govern tempo, cognitive overload accelerates. If you fail to govern skills, the humans lose the ability to direct agents effectively, which means more errors, more rework, and more cognitive load to manage the consequences. The loop tightens until something breaks.

And quality, the outcome that every organization cares about, lives inside this system. Specification quality, verification rigor, and human judgment are all products of cognitive health, sustainable pace, and maintained skills. You do not manage quality by inspecting outputs anymore. You manage quality by managing the humans who direct and evaluate the agents.

The ROI organizations are looking for will not come from the tools. It will come from transforming the management, the processes, and the roles that govern how those tools are used.

Beyond the Team

Everything I have described so far operates at the team level: one manager, one team, managing the human-machine interface. But the challenge does not stop there. Organizations will need to manage synchronization at four nested levels: the individual orchestrating multiple agents, the team integrating output across multiple human-agent loops, the organization governing all of this through its own multi-agent architecture, and ultimately the customer interface where your agents interact with your customers’ agents. Most organizations have barely addressed the first level. The third and fourth are coming faster than anyone is prepared for.

This is why the CEO cannot see the ROI. The investment went into the machine side of the equation. Nobody invested in the human-machine interface. That is like buying an F1 car and skipping the driver program, the race engineers, and the telemetry systems. The car is fast. The results are terrible. And the board is asking why the car is not winning races.

The Fork in the Road

Organizations will go one of two directions from here.

Some will push humans to keep pace with agents. More hours, more intensity, more parallel workstreams. They will celebrate the adrenaline phase and optimize for the metrics it produces. Their best people will burn out first because the best people are the ones pushing hardest. Then the juniors will hollow out because nobody invested in their skill development. Within eighteen months, these organizations will have fast machines and degraded humans, and they will wonder what happened to the AI returns they were promised.

Others will recognize that the agentic era requires a new kind of management. Not management with AI bolted on. Not AI-driven management with humans holding on. A deliberately designed system where the human and machine elements are integrated, each doing what they do best, with the manager governing the interface. These organizations will invest in their managers the way F1 teams invest in their drivers and race engineers. They will define what management actually consists of, probably for the first time. They will build structured development paths that equip managers for the three disciplines that matter: cognition, tempo, and skills.

The technology is extraordinary. The machine side of the equation is solved and getting better every quarter. The human-machine interface is where the value is trapped. Management is the key that unlocks it.

The question is not whether your organization will adopt AI. That is already happening. The question is whether your management model will evolve fast enough to turn that adoption into value.

The race has started. The car is on the track. Who is engineering the driver’s seat?

In October 2025, the Nobel Prize in Economics went to Philippe Aghion, Peter Howitt, and Joel Mokyr for their work on innovation-driven economic growth. At the center of the award was creative destruction: the process by which new innovations displace old ones, old markets collapse, and new ones emerge. Schumpeter described it almost a century ago. The Nobel committee confirmed it as the engine of modern prosperity.

Creative destruction has a simple promise. Innovation creates new value while destroying old structures. The car replaced the horse-drawn carriage. E-commerce displaced the shopping mall. Streaming killed the video rental store. Each time, the destruction was the price of progress. We accepted it because the creation that followed was worth it.

Organizations love creative destruction. They love the creative part, anyway.

And right now, the creative part has a name: AI. Generative AI and agentic AI arrived with a promise that organizations have not heard since the early internet: a step function in productivity. Not incremental improvement. A fundamental leap in what a person, a team, an organization can produce. Every leadership offsite, every board presentation, every quarterly review is now dominated by one question: how fast can we capture these productivity gains? The race is on. Organizations are pushing hard, investing heavily, and measuring obsessively.

And the success stories are everywhere. Sprints that took two weeks finish in days. Teams of five produce what used to require fifty. Code, analysis, strategy, content, all flowing faster than anyone thought possible a year ago. Organizations see these results and want to replicate them. They want to be the next case study, the next conference keynote, the next proof point.

But productivity is a stepping stone, not a destination.

Productivity changes roles. Then it changes processes. Then hierarchies. Then entire organizations. Then complete sectors. This is not a five-year forecast. This is the sequence that creative destruction follows every time, and it has never failed to complete itself. The question is not whether the sequence will play out. It is how far along it already is. Most organizations are at step one, heads down, chasing the productivity gains. They are not looking at steps two through five. They do not see the cascade coming.

And while they race to capture what AI is creating, they are not paying attention to what it is destroying. Not products. Not companies. Not markets. Something more fundamental.

Their skills. Your skills.

Let’s look at what is actually happening to skills when productivity accelerates.

Skills Have Always Moved

Throughout history, every major technological shift has dislocated skills. The first industrial revolution moved production skills from artisan workshops to factories. A weaver who worked from home became a machine operator in a mill. The second industrial revolution moved craft skills into mass production. A blacksmith became a steel worker on an assembly line. The digital revolution moved physical skills into software. A typesetter became a graphic designer. A clerk became a data analyst.

In every cycle, entire professions disappeared. In every cycle, new ones emerged. The pattern was painful but productive.

And in every cycle, skills also moved from humans to machines. The loom took over weaving. The calculator took over arithmetic. The robot took over welding. This was not a failure of the system. This was creative destruction working exactly as designed. A skill moves to the machine. A new human skill emerges in response. The loom operator needs different skills than the hand weaver. The CNC programmer needs different skills than the machinist. The data analyst needs different skills than the filing clerk.

For two centuries, this interaction between human and machine skills stayed balanced. Every time a machine absorbed a skill, humans moved up to the next level. Machines took execution, humans kept judgment. Machines took the repetitive, humans took the cognitive. Machines took the routine, humans took the exceptional. The escalator kept working. Humans always had somewhere to go.

But research from MIT economists has shown that even before AI arrived, the creation of new human tasks was already failing to keep pace with automation’s displacement. The escalator was already slowing down.

The question now is whether agentic AI stops it entirely.

Agentic AI is not absorbing the repetitive and leaving the cognitive. It is absorbing the cognitive too. Judgment. Reasoning. Analysis. Strategy. Code architecture. Legal interpretation. Financial modeling. When the machine takes over the thing that humans moved up to last time, where do humans move up to next?

I am not arguing against creative destruction. I accept Schumpeter’s framework. I accept that skills move to machines. I accept that destruction is necessary for growth. But the framework assumes that creation follows destruction. It assumes the escalator keeps moving upward. It assumes there is always a next floor.

What if this time there is not? Or more precisely: what if there is, but nobody is building the staircase to reach it?

Nobody is checking.

The Visible Response: Upskilling and Reskilling

Organizations have a default response to technological change: training programs. Upskilling (learn to use the new tools) and reskilling (learn to do a different job) are the visible, budgeted, comfortable answers. They show up in quarterly reviews. They have dashboards. They make leadership feel like the problem is being managed.

Upskilling is necessary. But it addresses only half of the equation. It tracks what is being added. It says nothing about what is being lost.

Reskilling is harder than anyone admits. Research consistently shows that meaningful reskilling across industries and professions has significant barriers beyond a certain career stage. A 45-year-old project manager does not become a machine learning engineer because the organization offers a training budget. The promise of reskilling at scale is comforting. It is not a strategy.

Both responses share the same blind spot: they focus entirely on the creative side of creative destruction. They assume the destruction will take care of itself. It will not.

Let’s look at what organizations are not seeing.

Deskilling: The Unmanaged Crisis

Every time an agent takes over a task, the human skill behind that task begins to atrophy. Not because anyone decided it should. Because nobody is paying attention. The skill quietly migrates to the agent, and one day you realize nobody knows how to do it anymore.

This is deskilling. It is happening in every organization that has adopted AI. And it is happening without governance, without tracking, without any strategic framework for deciding which skills are acceptable to lose and which are critical to preserve.

The process is invisible because each individual instance is rational. Why would an engineer manually review logs when an agent does it faster and more thoroughly? Why would a product manager write a competitive analysis from scratch when an agent produces a first draft in minutes? Why would a security analyst manually triage alerts when an agent handles it at machine speed?

But it is not just happening to your teams. It is happening to you.

When was the last time you wrote a document from a blank page instead of asking an AI to draft it first? When did you last summarize a meeting from your own notes instead of reading the AI-generated summary? When did you last write an email entirely from scratch instead of editing what the AI suggested? When did you last form an opinion about a problem before asking an AI what it thinks?

These feel like small things. They are not. Each time you delegate a cognitive task to an AI, you gain speed. And you lose a small piece of the skill that made you capable of doing it yourself. Writing clearly. Synthesizing information. Forming independent judgment. These are not minor skills. They are the foundation of what makes you effective.

Each decision makes sense. The cumulative effect is that the organization’s human capability base is eroding, from the bottom of the org chart to the top. And nobody is measuring the rate of erosion.

Some of these skills are core. They are not peripheral conveniences. They are the capabilities that define what the organization knows how to do. Its institutional knowledge. Its competitive differentiation. Its ability to operate independently when the tools are unavailable, unreliable, or wrong.

The United States offers a cautionary parallel. Over several decades, American companies made a series of rational, individual decisions to outsource semiconductor manufacturing to Asia. Each decision made sense in isolation. Lower costs. Higher margins. Focus on core competencies. Nobody governed the cumulative effect. And then one day the country woke up and realized it could not build chips anymore. The skills were gone. The know-how was gone. The manufacturing infrastructure was gone. The CHIPS Act was a $52 billion admission that unmanaged skill dislocation had become a national security problem.

Organizations are making the same mistake now. Each decision to let agents handle a function is rational. Faster, cheaper, better. But nobody is tracking the cumulative loss. Nobody has a ledger that says: here is what we could do independently last year, here is what we can do independently this year, and here is the gap. Nobody is asking: which skills can we afford to lose, and which ones are we going to desperately need when the agent is unavailable, when the model hallucinates, when the system fails at 2 AM and a human needs to step in?

And here is the difference that makes this worse than chips. When semiconductor manufacturing moved to Asia, it moved to other humans. The knowledge still existed somewhere. You could, in theory, bring it back. It costs $52 billion and takes a decade, but the path exists. When skills move to AI, the return path is far less clear. The humans who held those skills have atrophied, retired, or never learned them in the first place.

Never-Skilling: The Void

Deskilling is about loss. Something you had is eroding. Never-skilling is different. It is the creation of absence. Skills that will never exist in the first place.

Consider navigation. A generation of drivers has grown up with Waze and Google Maps. They do not have a degraded sense of direction. They never developed one. Nobody decided this should happen. Nobody governed the transition. It just did. One day you realize you cannot drive to a place you have been to fifty times without an app telling you where to turn. For younger drivers, there was never a “before.” The skill did not atrophy. It was never formed.

Waze replaced one skill: spatial navigation. Agentic AI is replacing entire skill families simultaneously. Code architecture. Legal reasoning. Financial modeling. Strategic analysis. Debugging. Systems thinking. The breadth is incomparable. And the speed matters. The deskilling of semiconductor manufacturing took decades. The deskilling of navigation took about ten years. The never-skilling of the agentic AI era is happening in months.

A junior developer who joined a team six months ago and has used agents from day one has never written a function from scratch. Never debugged by reading a stack trace line by line. Never built the mental model of how systems connect and why they fail. Never sat with a failing build for three hours and emerged understanding something fundamental about the architecture. This is not a degraded developer. This is a developer whose foundational skills never formed. And there are thousands of them entering organizations right now.

The same is true one level up. A new manager who has always had AI available has never built a project plan from pure judgment. Never estimated effort by gut feel and experience alone. Never navigated a team conflict without an agent drafting talking points. Never made a hiring decision without an AI-generated candidate comparison. The management skills that the previous article in this series deconstructed are not just dissolving for experienced managers. They are never forming in the next generation.

The implications for organizations are different from deskilling. For deskilling, you can theoretically recover. The knowledge existed somewhere. Maybe it is documented. Maybe someone still remembers. Maybe you can hire someone who learned it the old way. The path back is expensive and slow, but it exists.

For never-skilling, there is no path back because there was never a path forward. Recovery is not retraining. It is training from zero, for something the organization may no longer have anyone capable of teaching. How do you teach junior engineers to debug manually when no senior engineer has done it in two years? How do you develop managers who can make judgment calls without AI when no current leader practices that skill anymore?

Never-skilling is not inherently bad. Nobody mourns the loss of hand-calculating logarithm tables. Some skills genuinely do not need to exist anymore. But that is only true when the decision is deliberate. When the organization has consciously decided: we will never build this skill internally, we understand the dependency this creates, and we accept the risk. When it is unmanaged, ungoverned, invisible, it is not a decision. It is negligence.

The difference is between a controlled demolition and a building collapsing on its own. Same result from the outside. Entirely different in terms of safety, intent, and what you can build next.

The Extraction Economy

We have spent years debating the extractive nature of AI. The conversation is well established. Training models requires energy. Data centers require water and land. Hardware requires rare earth minerals. Training data is scraped, licensed, and fought over in courtrooms. Data labeling requires cheap human labor in the Global South. These are acknowledged costs, sometimes measured, occasionally regulated.

It is time to add human skills to that list.

We have grown accustomed to thinking of data as the valuable extractable asset of the AI era. Data is the new oil. Data moats. Data flywheels. But there is a second extractable resource that nobody is discussing with the same urgency: human skills. Not data about what people do. The actual capabilities: the judgment, the reasoning patterns, the professional intuition that people have built over careers.

AI is not just consuming resources to function. It is extracting human capabilities to improve. And there is now an industry built around doing this at scale.

Even Sam Altman, CEO of OpenAI, framed it in these terms when he noted in February 2026 that it takes “20 years of life and all of the food you eat during that time before you get smart.” He meant it as a defense of AI’s energy costs. But he understated the investment. Twenty years gets you to adulthood, not expertise. A doctor takes 30 to 35 years of education, training, and clinical practice. A senior engineer takes decades of building systems, debugging failures, and developing architectural judgment. That is the real cost of producing human skill. And if it takes decades to build a skilled professional but months to extract that skill into a model through reinforcement learning, the economics favor extraction every time. Unless someone decides the decades-long investment is worth protecting.

Every technology in history has absorbed human skills. That is not the problem. The point is not to slow AI adoption. The point is to govern what happens to human capabilities while adoption accelerates.

Mercor, a San Francisco startup valued at $10 billion, has built its entire business on harvesting human skills. They recruit engineers, lawyers, doctors, bankers, and journalists. These experts evaluate AI outputs, grade model performance, and teach agents the judgment, nuance, and reasoning that only comes from years of professional experience. They manage over 30,000 contractors and pay them over $1.5 million per day for the privilege of encoding their expertise into AI models. The agents learn. The experts become less necessary. One contractor told the Wall Street Journal she joked with friends that she was training AI to take her job someday. She was not wrong.

Mercor understands something that most organizations do not. Human skills are a monetizable, extractable asset. They built an infrastructure to identify, evaluate, and harvest those skills at scale. Their customers include the top AI labs in the world.

And they solved a problem that economists believed was unsolvable. In 1966, philosopher Michael Polanyi observed that “we know more than we can tell.” We cannot articulate the rules behind our own judgment, our pattern recognition, our intuition. For decades, this was treated as a permanent firewall protecting human labor from automation. If we cannot codify tacit knowledge, machines cannot acquire it. Economists built entire labor theories on this assumption. Polanyi’s Paradox was the reason humans would always be needed.

Mercor broke the paradox. Not by asking experts to articulate their knowledge (that was always the bottleneck), but by having them evaluate, grade, and compare AI outputs. The expert does not need to explain how they reason. They just need to say: this is right, this is wrong, this is better than that. The tacit knowledge gets encoded through reinforcement learning from human feedback without ever being made explicit. The paradox is not solved. It is bypassed.

This means the extraction is not limited to explicit, teachable skills. It reaches into judgment, intuition, the “I cannot explain why but this feels wrong” that senior professionals carry. The deepest skills, the ones we thought were permanently human, are now extractable too.

Here is the uncomfortable question. A 22-year-old CEO in San Francisco built a $10 billion company because he understood the value of human expertise. Do you, as a manager, as an executive, understand the value of the skills inside your organization? Do you understand the value of your own skills? Do you have a framework for cataloging them? For protecting them? For deciding which ones are strategic assets and which ones you can afford to let migrate?

Organizations spent the last decade learning, painfully, that data was a strategic asset that needed governance and protection. GDPR. Data sovereignty. Classification frameworks. Access controls. All of that emerged because organizations realized their data was walking out the door and being monetized by others.

Skills are now in the same position data was ten years ago. Being extracted, encoded, and monetized without governance. No skills sovereignty. No classification of which capabilities are strategic. No controls on how institutional knowledge flows out. No awareness that the asset is even being depleted.

The first era of AI was about data extraction. This era is about skill extraction. And organizations are making the same mistake twice: treating a strategic asset as though it is infinite and uncontested, until one day it is gone.

The Scarcity Inversion

In the previous articles of this series, I argued that the iron triangle of project management (time, cost, scope: pick two) is breaking. Execution resources are becoming effectively unlimited. A team of five with agentic AI can produce the output of fifty or five hundred. The constraint has shifted.

But if execution is abundant, what is scarce?

Human skills. The judgment that tells the system what to build. The experience that recognizes when an output is wrong. The institutional knowledge that understands why something was done a certain way. The taste that distinguishes adequate from excellent. These skills live in people, not in models. And they are finite, depletable, and currently unprotected.

Organizations are treating the scarce thing as infinite. They assume skills will always be there, that humans will always know what they know, that institutional knowledge is permanent. Meanwhile, they obsess over the abundant thing: execution speed, tool adoption, productivity metrics.

The scarce resource is the one being extracted. And you, as a leader, are not exempt. Your judgment, your experience, your institutional knowledge, these are part of the scarce resource. They are also being quietly outsourced, one agent interaction at a time.

What Comes Next

I have laid out a problem. Let’s talk about what to do with it.

In a previous article in this series, I introduced a framework for how managers can operate in the agentic age: Sense, Make Sense, Decide, Act. That framework applies directly here.

Sense. Map the skills in your organization. About twenty years ago, I led the development of an internal skills system at Amdocs. The goal was talent management: mapping who could do what, identifying gaps, planning development paths. The system worked. And like most of its kind across the enterprise world, it was eventually treated as HR overhead and deprioritized. Today, organizations need that map back. Not for talent development. For survival. Begin with a simple question: what could your teams do independently last year that they cannot do without AI today? What skills are your new hires arriving without? Most organizations have never done this inventory. Start.

Make Sense. Classify what you find. Which skills are strategic assets that define your competitive advantage? Which ones are acceptable to let migrate to AI? Which ones are you not building in the next generation, and is that deliberate? Not all deskilling is bad. Not all never-skilling is negligence. But you cannot tell the difference without a map.

Decide. For each category, make an explicit choice. Preserve, release, or accept the dependency. The key word is explicit. The problem is not that skills are moving to AI. Skills have always moved. The problem is that nobody is deciding which ones should move and which ones must stay. An unmanaged transition is not a strategy. It is a slow-motion version of the CHIPS Act waiting to happen inside your organization.

Act. Build the governance. Track skill erosion the way you track technical debt. Create the organizational equivalent of what data governance did a decade ago: classification, protection, and accountability for a strategic asset. Make skill governance part of leadership responsibility, not an HR side project.

Creative destruction won the Nobel Prize because the theory assumes that creation follows destruction. For two centuries, that assumption held. The escalator kept moving. But the data shows it was already slowing before AI arrived. Agentic AI may slow it further. The organizations that thrive will not be the ones that adopt AI the fastest. They will be the ones that govern the transition deliberately: tracking what they are gaining, what they are losing, and what they are never creating.

The deeper question is who in the organization owns this. Who is responsible for the skill architecture of the enterprise? That is a question about the future of management itself. It is the subject of the next article in this series.

But you do not need to wait for it. You have a framework. Start sensing.

The crisis is not that management needs to evolve. It is that management, as a concept, is losing its meaning, and we have not yet built the new one.

If you are a manager today, you have probably done everything right. You learned the AI tools. You encouraged your teams to adopt them. You watched development cycles compress from weeks to days. You nodded along at the executive briefings about transformation. And yet, something still feels off. There is a discomfort you cannot quite name. A sense that the ground beneath your professional identity is shifting, even though you have checked every box on the adoption checklist.

I want to name that discomfort for you. It is not about the tools. It is not about speed. It is about the fact that the very definition of what you do, what management is, is dissolving. Not evolving. Dissolving.

In a previous article, Your Management Model Is the New Bottleneck, I argued that agentic AI has shifted the organizational constraint from execution to management. That article identified the bottleneck. This one asks why it exists. Because the bottleneck is not simply that management is too slow. The bottleneck is that management, as we have understood and practiced it for over a century, is built on foundations that are collapsing. All of them. Simultaneously.

Let me show you what I mean through something you know well.

Monday Morning

Picture a typical Monday morning for an engineering manager. You sit down to plan a two-week sprint. You open the backlog. You estimate effort with your team. You break features into stories, assign them, sequence dependencies. You set up the daily standups. You know that by Friday of the second week, you will run a retrospective, review velocity, and feed the insights into the next sprint. This is the rhythm you have practiced for years. It is how you manage.

Now picture the same Monday morning after your team adopts agentic AI. The sprint that used to take two weeks finishes in two or three days. The backlog you planned for is empty by Wednesday. The standup you scheduled for Thursday morning has nothing to stand up about. The retrospective covers a cycle so short that there is barely anything to reflect on.

You might think this is simply a speed problem. Plan shorter sprints. Adapt the cadence. Move faster. But it is not a speed problem. What happened to your Monday morning is the surface. Underneath it, five layers of what made you a manager are coming apart. Let’s take them one at a time.

Layer 1: The Process Is No Longer Yours

Start with the sprint itself. You designed it. You set the cadence, defined the ceremonies, decided what goes in and what stays out. The sprint was your process. You owned it.

But when agents execute the work in days instead of weeks, the sprint as a management process loses its function. It is not that you need a faster sprint. It is that the container you built around your team’s work no longer contains anything. The planning meeting, the daily standup, the backlog grooming, the retrospective. Each of these was a process you owned. Each of them gave your role structure and meaning.

The management practices we use today were born in the factories of the second industrial revolution. Frederick Winslow Taylor formalized them in 1911. Henry Ford operationalized them at scale. Study work scientifically. Break it into processes. Assign each process to the right worker under the right supervision. We are now in the fourth industrial revolution, and we are still managing with the second one’s operating system.

Every industrial revolution dislocates skills. The first moved production skills from artisans’ homes and workshops to factories. The skills did not disappear. They relocated to a new context with new rules. The fourth industrial revolution is doing the same thing, not just to execution skills, but to management skills. Process design, orchestration, coordination. These functions are migrating from humans to agents. The skills do not vanish. They dislocate.

Management, in the old model, meant owning processes. Designing them. Running them. Governing them. Being accountable for their execution. This model shaped everything that followed, from manufacturing floors to Agile transformations. The fundamental unit of management was the process.

Agentic AI does not improve your processes. It takes them over. When autonomous agents handle code generation, testing, deployment, and monitoring, the manager is no longer the process owner. The sprint was yours. Now it is not. And it is not that the process evolved and you need to learn the new version. The process moved outside human management entirely.

This is the first thing managers need to unlearn: the belief that their value comes from process ownership.

Layer 2: The Signals Are Beyond You

Go back to your Monday morning. Part of what you did during the sprint was collect information. You sat in standups not just to hear status but to read the room. You noticed that one developer was stuck. You saw that two teams were building toward a conflict. You caught the early signal that a dependency was going to slip. This was not busywork. This was sense-making. You were the nervous system of your team, collecting signals, filtering noise from meaning, and relaying decisions.

The signals you collected were designed for human-speed work. Velocity measured how fast your team moved through stories. Story points estimated human effort. Sprint burndown tracked human progress over human time. These metrics are meaningless when agents complete in hours what humans planned for weeks. Velocity is infinite. Story points measure nothing. The burndown chart is flat by Wednesday. The signals that gave your Monday morning its structure and purpose are gone.

And what replaces them is not a better dashboard. It is a torrent of agent-generated data, logs, events, decisions, and outputs, running continuously, that no human can synthesize by reading standups or walking the floor.

Peter Drucker, perhaps the most influential management thinker of the twentieth century, described this sense-making function as the core role of middle management. Managers were information processors. They collected signals from processes and employees, synthesized them, and turned raw data into organizational intelligence. Without this function, executives operated blind and frontline workers operated disconnected.

Now your agents are running 24 hours a day, 365 days a year. A single agentic system managing a software pipeline produces more data points in a day than you could process in a month. The signals are there. They are richer, faster, and more detailed than anything your standups ever produced. But no human can collect them at the speed and scale required.

Drucker’s model assumed human-speed organizations producing human-volume signals. Neither assumption holds. The nervous system metaphor breaks, because the organism now runs at machine speed. A human nervous system cannot keep up with a machine body.

This is the second thing managers need to unlearn: the belief that their value comes from being the organization’s information processors. The information has outgrown you.

Layer 3: Delegation Has No Meaning

Back to Monday morning. After you planned the sprint and assigned the stories, what did you actually do? You delegated. You took a body of work, broke it into pieces, and transferred each piece to an employee along with the authority and responsibility to complete it. Then you monitored execution. This is not one approach to management. This is management. The entire discipline rests on this transfer chain.

But now one of your engineers receives the story you assigned, opens an agentic coding tool, writes a specification, and the agent delivers working code in hours. What did you delegate? Not the work. The agent did the work. Not the judgment on how to execute. The engineer made those decisions in their conversation with the agent. Your delegation was a formality. The real working relationship is between the engineer and the agent, not between you and the engineer.

Authority, accountability, responsibility. These are not just management terms. They are the operating system of management. And in the agentic AI age, they need to be completely redefined. Not adapted. Not modernized. Redefined.

If the process is not yours to manage, what is the meaning of delegation? If the signals are too fast and too many for you to process, what is the meaning of oversight? If the execution happens without human hands, what is the meaning of accountability?

These are not rhetorical questions. They are the questions every manager will face.

Layer 4: When Language Fails, Identity Breaks

Most conversations about AI and management stop here. They talk about new tools, faster processes, different metrics. They do not go where we are about to go.

Go back to your Monday morning one more time. Think about how you described your job to someone at a dinner party last month. “I manage a team of twelve engineers. I run two-week sprints. I own the release process. I make sure the right people are working on the right things. I remove blockers.” Every word in that description (manage, run, own, assign, remove) comes from a vocabulary built over a century of management theory. These words are how you think about your role. They are how you understand your professional identity.

Now consider: if the processes are not yours to own, if the signals are beyond your capacity, if delegation has lost its meaning, then every word in that dinner party description is wrong. Not outdated. Wrong. Your job title no longer describes your job.

Ludwig Wittgenstein wrote in his Tractatus Logico-Philosophicus: “The limits of my language mean the limits of my world.” We can only think about what we have words for. The boundaries of our language set the boundaries of what we can conceive. Managers today are trying to describe an agentic world using pre-agentic vocabulary. They say “delegation” and think they are adapting. They say “oversight” and think they are leading. But the words are pulling them back into the old model. The language itself prevents them from seeing the new reality.

This is not a new problem. Language always lags behind fundamental shifts. When the first automobiles appeared on streets in the late 1800s, people had no word for what they were seeing. They knew horses. They knew carriages. So they called this strange new machine a “horseless carriage,” defining it by what it lacked rather than what it was. The word “automobile” came later. A century after that, when self-driving vehicles appeared, the same pattern repeated. People saw a car without a driver. “Driverless car.” Only later did the term “autonomous vehicle” emerge, describing what the thing actually is rather than what it is missing.

Management is in the horseless carriage phase right now. We talk about “AI-augmented management” or “managing with agents” or “AI-era leadership.” Every one of these terms defines the new reality by referencing the old one. We do not yet have the word for what management is becoming. We only know what it is no longer.

This is why the discomfort I described at the start of this article is so hard to name. It is literally hard to name. The words for it do not exist yet. Managers sense that something fundamental has changed, but the language available to them keeps pointing back to a world that no longer exists.

The unlearning required is not behavioral. It is not about learning new tools or adopting new workflows. It is linguistic and conceptual. Before managers can act differently, they need to think differently. And before they can think differently, they need new language.

When the language that defines your professional identity breaks, your professional reality breaks with it.

Layer 5: The Pipeline Is Emptying

There is one more layer, and it concerns the future.

There is a growing conversation about the junior developer crisis. If agentic AI handles code execution, how do junior developers learn the craft? The apprenticeship model, where juniors grow through doing, breaks when there is nothing left to do in the traditional sense.

I believe a similar crisis is coming for management. And it may be worse.

Think about how you became a manager. You performed well as an individual contributor. You got promoted to lead a small team. You learned by doing. You practiced delegation. You learned process ownership. You developed intuition for signal collection through years of experience. MBA programs formalized this with frameworks rooted in Taylor and Drucker. Leadership development programs reinforced it.

But if the management functions I have just deconstructed continue to dissolve, then the path to becoming a manager disappears too. There will be no process ownership to practice on. No delegation chain to learn. No signal collection to develop intuition around. The apprenticeship model that created you as a manager will not be available to the next generation.

The institutions that train managers are not yet aware of this trajectory. MBA programs still teach Taylorist efficiency and Druckerian information management. Today, that education still has value. But if the direction I have described in this article continues, and I believe it will, these institutions will find themselves training people for a profession that no longer exists in the form they are teaching it.

The crisis ahead is not just for current managers who need to unlearn. It is that the path for creating the next generation of managers is eroding. Because we have not yet defined what management is in this new world.

The Blank Slate

I have taken apart five layers of what management means today. Process ownership. Signal collection. Delegation. Language and identity. The pipeline that creates new managers. Each one is collapsing under the weight of a world that no longer matches the assumptions they were built on.

Perhaps the first signal that something deeper was breaking came in late 2024, when Paul Graham’s essay on “Founder Mode” versus “Manager Mode” ignited debate across Silicon Valley. The argument was simple: professional managers, trained to hire well, delegate effectively, and stay out of execution, were underperforming. Founders who stayed close to the work, who went deep rather than orchestrating from a distance, were outperforming them.

That debate was easy to read as a personality argument. It was not. It was a structural one.

Founder mode is not about charisma. It is about proximity. When the abstractions of management stop working, when delegation chains distort signal and process ownership slows adaptation, leaders instinctively collapse the layers. They move closer to the work because the managerial interface no longer transmits reality clearly enough.

Look at what is happening at the highest levels. Satya Nadella is not delegating Microsoft’s AI transformation down a management cascade. He is driving it from the center. Sergey Brin returned to Google to work directly on AI systems rather than remaining at a strategic distance. These are not stylistic quirks. They are structural responses to a management model that is no longer reliably mediating between strategy and execution.

Founder mode, then, is a compression response. It reduces the managerial abstraction layer to restore signal clarity.

But founder mode still assumes human execution. It assumes that if a sufficiently capable human goes deep enough, the system will realign. In the agentic AI age, even that assumption fails. When execution itself migrates to autonomous agents operating at machine speed, collapsing layers is not enough. There is no stable layer left to collapse into.

The shift ahead is not from manager mode to founder mode. It is from both to something else entirely, something we do not yet have language for.

I am not going to pretend I have all the answers for what replaces them. That would be premature and, frankly, dishonest. What I do know is this: every industrial revolution dislocated skills, and every time, the human role was redefined around what remained uniquely human. The first industrial revolution moved manual skills to factories, and humans became machine operators. The skills of management are now migrating to agents. The question is not whether humans still matter. They do. The question is what human skills remain when the management skills have relocated to machines.

The organizations that will thrive in the agentic AI age are the ones willing to stand in front of this blank slate and start building. Not adapting the old model. Not optimizing it. Not putting AI wrappers around Taylorist processes and calling it transformation.

Go back to your Monday morning one last time. The backlog is empty. The standup has no purpose. The sprint is over before it started. And the words you used to describe what you do at that dinner party no longer mean what you thought they meant. You are sitting in front of a blank calendar. That blank calendar is the blank slate.

Building from nothing.

The unlearning comes first. You cannot fill a space that is already occupied by obsolete assumptions. Managers must let go of the identity that made them managers. Organizations must let go of the structures that made them manageable. Institutions must let go of the curricula that made management teachable.

Only then can we ask the right question. Not “how do we manage in the age of AI?” but “what does management mean now?”

That question deserves its own answer. And it will get one. But first, we need to be honest about the fact that the old answer is gone.

For the last 50 years, project management has evolved alongside software development. New methodologies promised better outcomes: waterfall gave way to Agile, Scrum replaced traditional planning, velocity became the measure of progress. Yet despite these advances, the fundamental constraint remained the same. Projects were limited by human capacity to execute work. That constraint no longer exists.

Software development has achieved what seemed impossible: a three-order-of-magnitude increase in productivity. Projects that took months now complete in days. The bottleneck that defined half a century of project management has been removed. This changes everything about how we run and manage projects.

The central argument of this article is simple: management models designed for human execution speeds now throttle AI-enabled organizations. Everything that follows, the collapse of estimation frameworks, the inversion of roles from executors to orchestrators, the obsolescence of the iron triangle, flows from that single shift. The technology has moved. The question is whether your management structures can move with it.

The Evolution: From Black Holes to Compressed Timelines

Waterfall Era: The Black Hole Problem

For decades, waterfall methodology dominated software projects. Teams moved through sequential phases: design, then development, then testing. Customers saw nothing until testing began. This created what I call a “black hole”: a period where the system remained invisible to users, sometimes lasting months or years.

The risk was severe. During the black hole, customer needs changed, processes evolved, requirements shifted. When users finally saw the system, misalignment was common. Rework was expensive. Organizations paid the price of building the wrong thing, and only discovered it late in the process.

If we plot this on a graph where the X-axis represents time and the Y-axis represents value delivered, waterfall produced a hockey stick. The line stayed flat for most of the project timeline, then jumped sharply near the end. No value for months, then everything at once.

Agile Era: Distributing Value, Not Accelerating It

Agile methodologies emerged in the early 2000s with the Agile Manifesto, but gained mainstream adoption in organizations throughout the 2010s. The goal was to solve the black hole problem. Through iterative development and frequent releases, customers could interact with the system much earlier. Changes were easier and cheaper to integrate. Feedback loops shortened from months to weeks. The risk of building the wrong thing decreased.

I witnessed this transformation firsthand. As Chief Architect of the Israeli Ministry of Justice, I led the adoption of Agile methodology, making the ministry the first public sector entity in Israel to implement it. The change was significant: we moved from long, risky projects to iterative delivery with continuous feedback.

This was a significant improvement. But here is what Agile did not change: the overall time the project took. A 12-month project remained a 12-month project. What changed was how value was distributed across that timeline.

Instead of zero value until month 11, value appeared at month 2, 4, 6, 8, and 10. The hockey stick became a steady climb. Agile was a risk management innovation, not a speed innovation. It made projects more adaptable and visible, but not fundamentally faster.

Agentic AI Era: Compression and the Return of the Hockey Stick

Agentic AI changes the time dimension of projects. For the first time, we can actually compress overall project duration by orders of magnitude.

In August 2025, I wrote about how generative AI had eliminated the “man-month” constraint that Frederick Brooks identified in 1975. Brooks observed that developers wrote approximately 9 to 12 lines of debugged, production-quality code per day. This number held constant for 50 years. When I was CTO at a previous company in the mid-2010s, I conducted an internal analysis and found that this number was still 9 lines per day. It was a disappointing but revealing realization.

GitHub Copilot increased this to approximately 18 lines per day, a 56% improvement. But agentic coding tools like Cline and GitHub Copilot Agent Mode have moved us from 11 lines of code per day to 11,000. This is not incremental improvement. This is a three-order-of-magnitude leap.

In recent projects I have observed, tasks that historically took two to three weeks were completed in minutes using agentic workflows. Two-person founding teams are building entire software products without hiring engineers. The constraint is no longer human coding capacity.

But the value curve has changed in an unexpected way. We are returning to the hockey stick pattern, but now it is compressed on the X-axis.

Here is why. Agentic AI requires significant investment in specifications upfront. You must define requirements with precision, design interfaces clearly, establish quality criteria explicitly, and document expected behavior in detail before the agent writes code. This front-loading creates a design phase similar to waterfall. Then agentic systems execute development in rapid bursts, often measured in hours instead of weeks.

The result is a compressed hockey stick. The black hole returns, but it is tolerably short. A project that took 12 months in waterfall and 12 months in Agile might now take 6 weeks: 4 weeks of intensive specification work, then 2 weeks of agent-driven execution and integration.

The trade-off is clear. You sacrifice Agile’s steady visibility for waterfall’s concentrated delivery. But you gain dramatic speed.

The Man-Month is Dead: What This Means

If human execution is no longer the constraint, the frameworks built around that constraint begin to fail. The implications extend far beyond faster coding. Where do the next bottlenecks emerge?

The answer: management structures, deployment infrastructure, organizational culture, and project management approaches. All of these were designed for constraints that no longer exist.

Breaking the Iron Triangle

Traditional project management operated within the iron triangle: scope, time, and resources. All three were limited. All three constrained what could be accomplished. You could optimize two, but the third would suffer. Project management meant navigating these trade-offs with skill and precision.

In software-centric knowledge work, agentic AI breaks the triangle.

Resources are no longer the binding constraint in most software development contexts. The economics have shifted. Traditional projects were constrained by headcount: you needed N developers, each costing $120K-200K annually, and you could only hire and onboard them so fast. Now organizations purchase tokens. A company can provide unlimited token access to every employee for a fraction of the cost of a single developer’s salary. Rate limits and compute constraints exist, but they are manageable operational details, not strategic bottlenecks. The question is no longer “how many people can we afford?” but “how effectively can our people direct AI capacity?”

Time compresses by orders of magnitude. Development that took weeks happens in minutes. The velocity of execution has increased by orders of magnitude.

Budget shifts from people to tokens. GitHub Copilot costs $20-40 per user per month. Claude API access scales with usage but remains marginal compared to human salaries. The cost profile of software development has changed. Organizations spend money on tokens, infrastructure, and the smaller number of humans who orchestrate the AI, not on large teams of code writers.

When the foundational constraints disappear, the entire optimization game changes.

New constraints emerge:

Traditional Constraint

New Constraint:

• Number of developers

• Specification quality

• Developer productivity

• Integration complexity

• Training and onboarding time

• Human review capacity

• Salary budget

• Infrastructure and tooling costs

• Communication overhead in large teams

• Organizational readiness for change

The project management practices that were optimized for the old constraints are now misaligned with the new reality. We are managing projects as if human coding capacity is still the bottleneck. It is not.

The Taylorism Problem

Current management practices trace back to Frederick Taylor and the second industrial revolution. Taylorism was born in Henry Ford’s factories over 100 years ago. Its principles persist today: break work into granular tasks, managers plan while workers execute, measure and control output, optimize the assembly line for efficiency.

Agile challenged some Taylorism assumptions. It introduced cross-functional teams, self-organization, and iterative planning. But Agile preserved the fundamental division: managers set direction, teams execute work. One group thinks, another group does. Agentic AI makes this division obsolete.

Everyone is a Manager Now

When employees operate agentic AI systems, they are no longer executors. They are managers. They define goals for agents, allocate work across autonomous systems, review outputs for quality, approve results, and refine specifications based on feedback. The actual execution has been delegated to machines.

Andrej Karpathy captured this shift in December 2025 when he wrote: “I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last year.”

This sentiment reflects a broader transformation. The work is no longer writing code. The work is orchestrating systems that write code. This applies beyond software development to project management, coordination, and organizational operations.

The implications run through every level of the organization.

For employees, if you manage agents that produce the output of 10 or 100 people, your role is different in kind. You are no longer measured by the work you perform with your own hands. You are responsible for the quality of direction you provide, the clarity of specifications you write, and the judgment you apply when reviewing agent outputs.

For managers, if your team members are now manager-operators of agentic systems, you are managing managers, not executors. Your role shifts from task assignment and progress tracking to something else entirely. What is that “something else”? Strategic direction, system design, quality oversight, capability building, and organizational alignment. Manager upskilling is not optional. It is mandatory. Managers who cannot operate effectively in the agentic paradigm will become bottlenecks to their own teams’ productivity.

For organizations, team capacity multiplies. A team of five employee-managers operating agentic systems might produce the output that previously required 50 or 500 people. How do you structure organizations when capacity scales this way? What do hierarchies look like? What does career progression mean?

These are not theoretical questions. They are immediate and pressing.

Let me give you a concrete example of what this looks like in practice.

A product manager at a large enterprise software company used to spend 60% of her time coordinating: writing status updates, chasing down developers for progress, updating roadmaps, and synthesizing feedback from customer calls. Now, GitHub Copilot and AI agents handle the coordination. They monitor pull requests, aggregate status from project tracking tools and chat platforms, generate stakeholder updates, and flag blockers automatically.

She spends her time differently now: 40% on strategic product decisions, 30% on customer discovery and competitive analysis, 20% reviewing and refining agent outputs, and 10% on exception handling when agents escalate edge cases.

Her output? The team ships features 3x faster. But her role is unrecognizable compared to two years ago. She’s managing a system that manages the work.

Now multiply this across an organization of thousands of employees.

Questions for Reflection

These questions have no universal answers. Each organization will find its own path. But the questions themselves are unavoidable.

Estimation: If tasks complete in minutes instead of weeks, how do you estimate project timelines? What planning horizons make sense? Do traditional velocity metrics mean anything when velocity is no longer a constraint? Should we abandon estimation entirely, or do we need new models built from first principles?

Risk: Where does risk now concentrate? In specification quality? Agent reliability? Integration points between AI-generated code and existing systems? Human review capacity becoming the bottleneck? If agents execute work autonomously, what failure modes must you plan for that did not exist before?

Accountability: Who is accountable when an agent produces work? The employee who directed the agent? The manager who approved the specification? The organization that deployed the system? How do you assign responsibility for outputs you did not directly create? Traditional accountability frameworks assume human authorship. What happens when that assumption breaks?

Responsibility: If agents handle execution, what are employees responsible for? Quality validation? Strategic refinement? Continuous learning and improvement of agent systems? How do you define contribution when the traditional definition of “work performed” no longer applies?

Trust: Do managers trust employees to effectively direct agentic systems without constant oversight? Do employees trust managers to lead this transition without reverting to Taylorist control and micromanagement? Trust is the foundation of organizational change. Where might trust break under this shift? How do you rebuild it?

The Failure Points

Organizations that fail to adapt will break at predictable points. These are not distant theoretical risks. They are happening now.

Management structures designed for Taylorism will fail first. When management clings to old operating rhythms while teams operate at machine speed, management becomes the bottleneck.

Specific failure modes:

• Weekly status meetings when agents complete work in hours

• Multi-week approval cycles when iteration happens in days

• Quarterly planning when market conditions shift monthly

• Stage-gate processes designed for human execution timelines

• Reporting structures that require manual aggregation of data agents already synthesize

The symptoms are consistent: teams wait for approvals, decisions lag behind execution, work accumulates in review queues, and employees route around formal processes to maintain velocity. The organization’s speed collapses to the speed of its slowest management layer. If your processes assume human execution speed, they will throttle AI-enabled teams back to human speed.

Employee confidence in leadership will erode second. This erosion accelerates when managers and senior leaders do not use agentic AI themselves. Employees who use AI to multiply their output while their managers operate with traditional methods creates a credibility gap that undermines transformation efforts. If leadership demands adoption but does not model it, the message is clear: this is real enough to change your job, but not real enough to change mine. People follow leaders who understand where they are going.

Measurement and incentive systems will produce perverse outcomes third. KPIs designed for human productivity: story points, velocity, lines of code written. These are metrics optimized for the old world. Organizations that continue measuring the old metrics will optimize for the wrong outcomes.

Cultural inertia will be the final obstacle. One hundred years of Taylorism has embedded assumptions deep in organizational DNA. “Good managers assign clear tasks.” “Good employees execute reliably.” “Productivity means hours worked.” These beliefs are not consciously held. They are instinctive. Unlearning is harder than learning.

The Unlearning Challenge

The shift from human execution to agentic execution requires unlearning what we know about managing people and projects. This is not incremental improvement. This is a wholesale replacement of the operating model.

What must be unlearned:

• That project duration is determined by team size and developer productivity

• That managers plan and workers execute

• That granular task breakdowns alone drive project success

• That the iron triangle defines project trade-offs

• That steady, incremental progress is always superior to concentrated delivery

What must be learned is not yet fully defined. We are at the frontier. Early adopters are discovering new patterns through experimentation. Some will fail. Some will find sustainable models. The organizations that succeed will set the standards that others must follow.

Call to Action

If you lead software development or manage project teams, you face a decision: adapt now or inherit technical debt and organizational misalignment you did not create.

The pace of change will not slow. Jack Welch, the legendary CEO of General Electric, once said: “If the rate of change on the outside exceeds the rate of change on the inside, the end is near.” That warning is no longer theoretical. The world is accelerating. Agentic AI is accelerating faster. Your competitors are adapting. Your customers expect machine-speed delivery.

A Framework for Transformation: Adopt → Infuse → Invent

Organizations that successfully manage this transition tend to follow a three-phase pattern. This is not mandatory, but it is highly recommended based on early adopter experience. The activities listed in each phase are examples. Every organization must adapt this framework to their specific company context and sector requirements.

Phase 1: Adopt (Continuous Foundation)

This phase never stops. It runs continuously alongside the other phases.

• Deploy AI tools and make them accessible to all employees

• Build comprehensive training and upskilling programs

• Change recruitment criteria to assess AI capability alongside traditional skills

• Create space for experimentation and learning

• Measure adoption rates and capability growth

The goal is organizational literacy. Everyone learns to operate agentic systems effectively.

Adopt fails when AI remains optional. If employees can bypass AI tools without consequence, adoption stalls at enthusiasts and never reaches critical mass.

Phase 2: Infuse (Short-Term Integration)

AI moves from the sidelines to the center. It becomes embedded in workflows and cannot be bypassed.

• Redesign processes around AI capabilities, not traditional human workflows

• Integrate agentic systems into core tools and platforms

• Establish new operating rhythms that assume AI participation

The critical insight: infusion cannot be purely top-down. Employees must discover and embed AI into their own work. Managers empower this process rather than control it. Employees understand the value through direct experience and find ways to integrate AI that managers could not have prescribed. This requires trust, experimentation space, and tolerance for failure.

Short-term snowball effects emerge here. Productivity multiplies. Bottlenecks shift. Teams discover new capabilities.

Infuse fails when managers override agent-driven workflows to restore familiar control patterns. The old operating rhythm reasserts itself and throttles the new capacity.

Phase 3: Invent (Long-Term Transformation)

Organizations move beyond using existing AI tools to creating new structures and opportunities.

• Develop custom internal tools tailored to specific workflows

• Build new external products and services enabled by AI capacity

• Evolve role definitions as work changes beyond recognition

• Design new organizational structures that leverage multiplied capacity

• Identify new business models and growth opportunities

Long-term snowball effects compound here. New roles emerge. Current roles transform beyond recognition. The organization operates in ways that were not possible before.

Invent fails when organizational structure remains static. If roles, hierarchies, and incentives do not evolve to match the new capacity, the organization captures tool-level gains but misses the transformational opportunity.

Immediate Actions

Here are the steps to take now.

Audit your constraints. Identify where your organization still optimizes for human coding capacity, the iron triangle, or Taylorist structures. These are misalignments. They will cause friction and slow you down.

Experiment with compressed timelines. Select a project. Front-load specification work. Deploy agentic tools. Measure how much time compresses. Identify the new bottlenecks that emerge. Learn what breaks and what holds.

Redefine manager roles explicitly. If employees become manager-operators, what do managers do? Answer this question with specificity. Test your answer with pilot teams. Iterate based on what you learn.

Challenge your estimation models. If velocity is no longer the constraint, what drives project timelines now? Rebuild estimation from first principles. The old models will mislead you.

Invest in organizational readiness. The technology is ready. The tools exist. The question is whether your culture, management practices, and trust structures can absorb the change. This is where most organizations will struggle.

What Success Looks Like

The organizations that succeed in this transition will not be the ones with the best AI tools. Those tools are available to everyone. Success will be defined by three markers:

Management operates at the same speed as AI-enabled teams. Decisions, approvals, and coordination happen in hours or days, not weeks or quarters. Management rhythm matches execution rhythm.

Employees trust that leadership understands the change. Managers demonstrate competence by using agentic systems themselves, not just mandating their use. The credibility gap closes.

New capabilities emerge that were not possible before. The organization invents products, services, or operating models that exploit multiplied capacity in ways competitors cannot match. This is the true test.

The Choice

Your backlog is ready. Your agents are listening. Your competitors are moving.

The question is not whether AI will transform how projects are managed. That transformation is already happening. The question is whether your organization will lead it, follow it, or be disrupted by it.

The default outcome is not stagnation. It is a widening gap between what your technology can deliver and what your organization can absorb. Every week that management structures remain calibrated to human execution speed, that gap grows.

The compressed hockey stick is here. The man-month is dead. Everyone is becoming a manager. Will your management model evolve at the speed your technology demands?

For the last 50 years, software development has evolved significantly, with new programming languages that have enabled us to do more with every line of code; new architectures have allowed us to build increasingly complex and robust software; new methodologies have streamlined larger and more diverse software teams.

We have transitioned from monolithic applications to Service-Oriented Architecture (SOA) and then to microservices. We have moved from closed organizational networks to the Internet, enabling remote and shared access to resources across the globe. We have shifted from on-premises infrastructure to cloud-based applications, introducing Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS). The mobile revolution has brought a much more capable edge.

Yet, despite these advances, the fundamental process of software development has remained the same — people still write the code and build the software.

The Mythical Man-Month and the Limits of Scaling Software Development

In 1975, Frederick P. Brooks published The Mythical Man-Month, a seminal book that identified fundamental limitations in scaling software development. One of his key observations was that simply adding more developers to a task often incurs higher managerial overhead, making the project more complex and potentially slower rather than faster.

He also introduced a seemingly “magical” number: the average number of lines of code a developer writes per day. This number was estimated to be between 9 and 12 lines of debugged, production-quality code per day. While I was the CTO in a previous company in the mid-2010s, I conducted an internal analysis and found that this number was still 9 lines per day. It was a disappointing but revealing realization.

But it should not have been surprising. Software development is not just about writing code — it involves communication, planning, design, and coordination. Developers must design APIs and interfaces, ensure modularity, and plan for team-wide scalability. Microservices architectures require careful orchestration to ensure autonomy among teams. Software must scale not just in production, but also in how it is developed and maintained. Parallelization and decoupling of development teams are essential.

Generative AI and the Productivity Breakthrough

When generative AI emerged in late 2022, one of its first major applications was enhancing software developer productivity. Microsoft had already pioneered this space with GitHub Copilot (GHCP) in July 2021, an AI-powered “pair programmer.” Research showed that developers using GHCP experienced a 56% productivity increase, which is a staggering figure compared to decades of marginal improvements.

Using Brooks’ number as a baseline, this means that instead of producing 12 lines of code per day, developers can now write approximately 18. However, the world is no longer experiencing linear improvements; we are seeing exponential advancements. The software industry, which should be “eating the world,” has historically only seen single-digit gains in productivity over 50 years.

But now, AI is enabling an unprecedented leap, moving from 11 lines of code per day to 11,000. Here is why.

The Democratization of Software Development

Generative AI has fundamentally changed the act of programming. For the first time, we no longer need an intermediary language to translate between human intent and machine execution. Natural language is now a viable programming interface.

Recently, we have seen the rise of “vibe coding”, coined by Andrej Karpathy, a movement where platforms like Lovable and GitHub Spark allow individuals and organizations to build software using AI without requiring professional developers.

Powerful IDE extensions like Cline and Roo now leverage multiple large language models to plan, design, and generate software autonomously. Their success signals a deeper inflection point, which is the transition from assistive AI to agentic coding, where the tool is no longer a smart autocomplete but a self-directed teammate.

Agentic coding means giving an AI an end-to-end goal such as “migrate this module” or “remove deprecated APIs” and letting it plan, execute, test, and iterate with minimal supervision. Cline’s “Plan→Act” workflow, for example, drafts a strategy composed of clear steps and then edits files and runs commands inside VS Code. GitHub Copilot’s new Agent Mode operates on a repository scale. It analyzes dependencies, applies code changes, fixes build errors, and opens a pull request that is ready for review, all from a single prompt. Together, they mark the moment the “man-month” dissolves into millisecond model calls.

The Death of the “Man-Month”

In the last few months, we’ve seen glimpses into the future of software development and the potential of generative AI to transform it. In recent projects I observed, tasks that historically took two to three weeks were completed in minutes using agentic workflows. The pattern repeats when the repository has tests, continuous integration, and a clear goal.

There is no more “man-month.” For the first time, we are witnessing a three-order-of-magnitude increase in software development productivity. New startups are being launched without dedicated engineering teams. Today, two-person founding teams can build entire software products and onboard paying customers without hiring a single engineer.

If software development is no longer constrained by human coding capacity, where will the next bottlenecks emerge? Will it be management? Or will it be our software build, deployment and pipeline infrastructure and tools?

The Acceleration Challenge

A decade ago, when Agile methodologies became dominant, we introduced “velocity” as a measure of team productivity. But velocity is static, and we now live in a world where software development is experiencing continuous acceleration.

Jack Welch, the legendary CEO of General Electric, once said: “If the rate of change on the outside exceeds the rate of change on the inside, the end is near.”

That warning is no longer theoretical.

The world is accelerating, and fast. If you are a software company or a large enterprise with a large engineering organization, you need to ask yourselves: Are you prepared for this revolution?

Also, you should probably be aware that your current ways of working, coding, building, and shipping software are no longer sustainable.

What If?

1. What if each engineer could write 11,000 lines of code per day?

2. What if you could complete an entire quarter’s backlog worth of software development in one week? What would we do with the remaining time?

3. What if the biggest constraint was not engineering capacity, but your infrastructure, deployment pipelines, or even your management layers?

4. What if, instead of 20 product groups, we had 200 or 2,000?

5. What if software development teams only needed two engineers and a product manager to build and ship software end-to-end?

Call to Action: A Pilot for the Future

To address this shift and understand how your organization should evolve, I propose designing and launching pilot programs. This pilot will involve the creation of four “SWAT” teams (Software engineering With Advanced Tools). Each team will consist of three engineers and one product manager, tasked with building end-to-end products using only enterprise-approved AI-assisted coding tools at your organization, such as GitHub Copilot and Cline.

The pilot will:

• Measure how quickly AI-first teams can develop complete, production-ready software.

• Identify the emerging challenges and new bottlenecks in AI-driven software development.

• Validate the feasibility of smaller, highly autonomous teams replacing traditional development structures.

• Map out the deployment challenges within your organizations that these new development approaches and tools will encounter.

• Understand the end-to-end product development challenges beyond time constraints, including scalability, security, integration, and long-term maintainability.

This will provide you and your organizations with actionable insights into the future of software engineering and help your management redefine its approach to building software in an era of generative AI.

The New Software Revolution

Brooks showed that adding people adds friction; generative AI removes it. Teams that adopt agentic coding today will set the tempo everyone else must follow, while those who wait will inherit technical debt they never wrote. Your backlog is ready, your models are listening, and your customers will not slow down. Start the pilot, measure the results, learn at the pace of this new epoch. The “man-month” is finished; software now moves at machine time.

Will your organization move with it?