Preface

The first two essays were about individuals: cognitive re-stratification in the AI era (Issue 1), and the failure and rebuilding of individual learning (Issue 2).

This essay is about organisations. Once individuals are re-stratified by AI, what happens to the firms, teams, and collaboration relationships built out of those individuals?

The colder question is this: which parts of existing organisational forms, management methods, and collaboration relationships have already structurally failed in the AI era, while we continue describing that failure with obsolete words?

This is difficult to discuss because the entire vocabulary of organisation, management, and collaboration (“firm,” “employee,” “manager,” “team,” “collaboration,” “efficiency”) comes from the industrial era. The words carry industrial-era assumptions. Using them to discuss AI-era phenomena is like discussing electric cars with the vocabulary of coal-burning efficiency. The issue is not only that the discussion becomes unclear. The words themselves are wrong.

But we do not yet have better words. So this essay does only one thing: use old vocabulary as precisely as possible to describe what is happening, while admitting the limits of the vocabulary itself.

No solution. An honest diagnosis.

Chapter 1 · First Ask the Skipped Question: What Did Firms Originally Solve?

Before discussing how firm structures should evolve in the AI era, we must ask the question everyone skips: what problem did the firm originally exist to solve?

Most people treat the firm as a natural state: a group of people doing things together. But the firm is not natural. It is an organisational solution that appeared under specific historical conditions to solve specific problems. If those problems are solved or changed by new technology, the necessity of the firm becomes less obvious.

Coase’s question

In 1937, Ronald Coase wrote “The Nature of the Firm” and asked a core question:

If markets are so efficient, if prices allocate resources, why do firms exist? Why doesn’t every economic activity happen through market transactions one by one? If I need A, I buy A; if I need B, I buy B.

Coase’s answer: market transactions themselves have costs. Search, negotiation, contracting, enforcement, accountability, and breach management can be so expensive that doing something through the market becomes uneconomical. In those cases, putting the activity inside an organisation is cheaper, because inside the organisation, commands and processes replace repeated market negotiation.

So the firm is, at bottom, a solution to the problem of excessive market transaction costs. When the cost of doing something through the market exceeds the cost of doing it inside an organisation, the activity is pulled into the firm.

Oliver Williamson and others later deepened Coase’s argument into the foundation of new institutional economics. This is not business-school decoration. It is the clearest starting point for answering why firms exist.

The eight problems firms solve

Expand Coase’s frame and firms solve eight concrete transaction-cost problems.

I. Search and matching costs

Complex work used to require finding many kinds of people: engineers, designers, salespeople, finance, legal. Searching for them in the market every time was expensive. You did not know who was reliable; they did not know whether you were reliable. Trust took time.

The firm solves this by binding a group of people together for the long term. Once inside the firm, engineers, designers, and salespeople become ready resources. They do not need to be searched for again each time.

II. Information asymmetry

In a market transaction, it is hard to know in advance whether the other party will do a good job. A freelancer says they can write good code, but how do you know? You discover quality only after paying or wasting time.

The firm uses long-term employment, reputation, and observable daily performance to reduce this asymmetry. If you spend three years inside a company, your manager knows your ability in a way no one-off market transaction can.

III. Asset specificity

Some investments only have value inside a specific relationship. If you spend three months learning a client’s internal system for a custom project, that learning may be useful only for that client. In a one-off transaction, you bear the risk of that specific investment.

The firm solves this through long-term relationship. Company-specific investments, such as learning internal processes, building internal relationships, and mastering local know-how, become worthwhile because the relationship lasts.

IV. Coordination costs

Complex work requires many people to coordinate. It is impossible to coordinate every step through market transactions. Imagine signing a contract every time an engineer needs to discuss a detail with a designer. The firm replaces repeated coordination with hierarchy and process: reporting lines, routines, and shared norms.

V. Risk sharing

In markets, each independent actor bears their own risk. Firms create risk sharing by binding many people together. One project fails, losses are distributed. One employee gets sick, others cover temporarily. One client leaves, the firm remains.

VI. Capital concentration

Some activities require large upfront investment: factories, R&D, equipment. Individuals or small teams cannot assemble the capital. As a legal entity able to raise money, the firm concentrates dispersed capital for large-scale production.

VII. Legal-contractual power

Many activities require a legal entity: signing large long-term contracts, bearing bankruptcy risk, paying employee benefits, issuing options, appearing in court, accepting regulation, obtaining licences.

Individuals either cannot do these things, because they lack the legal identity, or can do them only with excessive personal risk. The firm provides legal personhood. Shareholders’ liability is limited; the firm itself can contract, borrow, be sued, and go bankrupt. This personhood is not a side effect of the firm. It is one of the core reasons some things must be done through firms.

VIII. Brand and trust backing

Some situations require a credible third party, especially when consequences are high and quality is hard to verify in advance.

A hospital surgery, a law firm’s litigation, a bank’s custody service, a consulting firm’s strategy advice: the client cannot evaluate output quality beforehand. They are buying the promise that if something goes wrong, an institution stands behind the result. The weight of that promise comes from the institution’s brand, history, scale, and the incentive not to put all of that at risk.

An independent individual may be equally capable, but cannot provide the same weight of commitment. They do not have as much reputation or institutional substance pledged behind the promise, and the client does not have the same target to sue. In these situations, half of what the client buys is capability, and half is institutional backing. The second half is hard for individuals to replace.

Where the eight problems stand in the AI era

Once we understand what firms originally solved, we can ask the real question: do these eight problems still exist in the AI era, and is the firm still the best solution?

Take them one by one.

I. Search and matching costs: weakened

Finding a reliable engineer, designer, or lawyer used to be expensive. Now AI replaces a large share of these roles’ work, so fewer people are needed. For the human work that remains, remote platforms, Upwork, LinkedIn, X, and Discord communities reduce search costs. AI’s general capability also means not every role needs to be an expert. A generalist who can use AI to write ordinary code, produce ordinary design, or handle ordinary legal work can replace what previously required a team.

So the cost-effectiveness of binding a group of people together long-term is falling quickly. Many things can now be done by temporary combinations or by an individual plus AI.

II. Information asymmetry: partly weakened

Long observation used to be the only way to judge whether someone could do the work. Now public work is a stronger signal than old résumés: GitHub, X, personal sites, past projects. AI also makes trial tasks and evaluation cheaper. But deep trust is not solved. Complex collaboration requiring long-term investment still needs long relationships.

This item is weakened, not eliminated.

III. Asset specificity: unchanged in essence, changed in form

Specific investment remains necessary for deep work. But the nature of the investment changes. Before, it often meant learning one company’s internal process, valuable only in that company. Now it more often means understanding a domain, mastering a tool stack, and building a distinctive perspective. These investments can be reused across relationships.

So an individual’s specific investment becomes less tied to a single employer. What you know is no longer “the firm’s know-how.” It becomes your own know-how, strengthening bargaining power and independence.

IV. Coordination costs: sharply reduced

This is the most dramatic change. Complex work once required meetings, processes, hierarchy, and pipelines. Now asynchronous tools reduce meetings, AI acts as a patient collaborator for documentation and follow-up, and one person plus AI can do work that previously required five people.

Small-scale collaboration has reached historically low cost. Three people plus AI can often produce more than ten people did before.

V. Risk sharing: still present, changed in form

Risk sharing remains necessary. Individuals still get sick, burn out, and make mistakes. But the old model of long-term employment plus benefits is increasingly rigid for both firms and individuals. New risk-sharing forms appear: flexible collaboration, mutual-aid networks among independents, communities, and AI as a tireless backup. In some countries, state welfare, insurance, and healthcare reduce personal dependence on firms.

The need remains, but the firm is no longer the only or best solution.

VI. Capital concentration: unnecessary for much software, content, and service work

Much work used to require large upfront investment. Now software can be built by one person, a computer, and AI. Independent creators can use AI to produce output that once required a content team. Consulting, design, writing, and translation can be done by individuals.

But capital concentration remains necessary for chips, rockets, drugs, large hardware, and physical infrastructure. In those fields, the necessity of firms has not changed.

VII. Legal-contractual power: not weakened, in some ways strengthened

AI has not changed contract law, company law, bankruptcy protection, compliance, or tax regulation. These still require legal entities.

AI also adds new legal complexity. Who is responsible for damage caused by AI output? Who pays for training-data infringement? Whose decision is an AI-made decision? These questions deepen the need for an entity to bear liability.

So this item is not weakened in the AI era. It may be strengthened.

VIII. Brand and trust backing: surface weakened, depth strengthened

AI cuts both ways. On one side, individuals can cheaply produce professional-looking deliverables. The surface gap in brand presentation is flattened.

On the other side, AI-generated content explodes, making trust scarcer. Clients increasingly cannot tell whether a proposal is the result of deep human thinking or mass-produced AI output. In high-consequence settings, clients need an institution to promise that this is not merely AI slop.

So the surface layer of brand is weakened, because polished deliverables are no longer rare. The deeper layer is strengthened. What is scarce is not polish, but real responsibility for outcomes. Individuals still struggle to replace institutions here.

A structural inference

Put the analysis together.

• Two categories are sharply weakened: search/matching and coordination.
• Two are partly weakened: information asymmetry and asset specificity.
• Two still exist, but the firm is no longer the best solution in many cases: risk sharing and capital concentration for software-like work.
• Two are not weakened and may be strengthened: legal-contractual power and brand/trust backing.

This is more cautious than the story that “firms are being destroyed.” The firm is not collapsing everywhere. Its boundary is contracting unevenly.

The contraction is sharpest in cognition-intensive, low-capital, low-contract-risk, low-regulation fields that do not require heavy brand backing: software, content, consulting, design, and some services. In these fields, one person plus AI can do work that used to require a team; three to five people plus AI can do what used to require a medium-sized company.

The contraction is weakest, and may reverse, in capital-intensive, heavily regulated, high-contract-risk, high-consequence fields: finance, medicine, infrastructure, manufacturing, energy, aerospace, government contracting. In these fields, firms may become more necessary because AI adds legal risk and trust scarcity.

So the accurate conclusion is not “the firm is disappearing.” It is: the applicable boundary of the firm is being redrawn. The firm used to be the default form for almost all collaboration. Now it remains necessary, even more necessary, in some areas, while becoming too heavy, slow, and expensive in others.

The rest of this essay discusses mainly the area of sharp contraction: organisational change in cognition-intensive work. If you are in the other area, the analysis below applies only to the cognition-intensive part of your work.

And our entire vocabulary for discussing organisation is built on the assumption that the firm is the default. On the contracting side, that vocabulary is about to hit its boundary.

Chapter 2 · Collaboration Is No Longer on the Same Plane

After asking why firms exist, the next question is: what does collaboration look like inside new organisational forms: independents, small teams, loose networks?

The answer is counterintuitive. AI does not simply make collaboration easier. It polarises collaboration.

The collaboration pool is shrinking

In the past, “who should I collaborate with?” was judged across many dimensions:

• relevant experience
• fluency and skill
• knowledge
• industry information and relationships
• judgment

These dimensions were roughly independent. A person with fifteen years of experience, strong fluency, rich knowledge, and ordinary judgment was a very useful collaborator. You could keep the judgment-heavy part yourself and give them execution, experience recall, and information synthesis. Their value was real.

AI does not make their value zero. It eats almost all of the first four components of that value.

• Experience: AI knows the common traps across industries and can state them instantly.
• Fluency: AI does not tire, miss work, or make routine mistakes.
• Knowledge: AI is a living knowledge base.
• Information: model training has absorbed much of the public world.

After those four are eaten, what remains is cognitive decoupling: the ability to make real judgments.

If that ability is strong, the person remains extremely valuable, perhaps more valuable than before because their judgment is freed from execution. If it is ordinary, the stacked value they used to have collapses, because the pillars supporting it have been replaced by AI.

The collaboration equation changes:

Before: experience + fluency + information + knowledge + ordinary judgment = useful collaborator

Now: AI already does the first four + ordinary judgment = worse than me + AI

That right-hand side is decisive. You are no longer comparing the person with yourself alone. You are comparing them with yourself plus AI. That option did not exist before. Now it is strong enough to make many “experienced partners” uneconomical.

The cost is not collaboration, but screening

This is often misread as “collaboration has become more expensive.” That is wrong.

With the right people, collaboration is smoother than before. Both sides use AI, both have high-bandwidth judgment, both can work asynchronously, both can use AI as a shared intermediary, and cycles can be fast.

What has become expensive is screening for qualified collaborators.

The acceptable dimensions were broader before, so the pool was larger. Now the acceptable standard is narrower, and the pool shrinks. Finding people inside a smaller pool is naturally much harder.

This is not arrogance or raised expectations. It is that the definition of “someone worth collaborating with” has changed at the root. Many people previously counted as qualified because experience and fluency supported them. Those abilities still exist in them, but their relative value has dropped below the threshold for collaboration.

You are not screening a fussier pool. You are screening a redefined pool.

Incommensurable production systems

An analogy makes the change clearer.

Two textile workers in the past could collaborate well. They were inside the same production logic: touch the thread, look at the cloth, finish a piece over several hours.

After the assembly line appears, a line-maintenance engineer and a hand weaver are not facing “higher collaboration cost.” They are in different production systems. The engineer works at the scale of machines: inspect, adjust, optimise, keep a hundred machines running. The weaver works at the scale of one cloth. Their time scale, object of judgment, and unit of output differ.

Asking them to collaborate is a category error. It is like asking a farmer and an air-traffic controller to collaborate on planting crops.

The same split appears in the AI era.

An independent using AI as leverage runs at the rhythm of “one loop in an afternoon, one product in weeks.” A worker embedded in traditional process runs at “one meeting per week, one release per quarter.” This is not an efficiency gap. A twofold or threefold gap can still collaborate. Once the gap exceeds an order of magnitude, it becomes a production-system problem. Time scale, judgment object, and output unit no longer share a coordinate system.

Discussing “should we ship this feature tonight?” with someone who produces once a month cannot align. Talking about “I just decided to change the whole architecture” with someone whose decisions require meetings cannot align. Saying “I am the product judgment” to someone who needs a PM to write requirements cannot align.

This is not communication. It is system mismatch.

Redefining “subordinate”

The contraction of the collaboration pool appears most clearly in the relationship called subordinate.

That structure is old: a superior makes directional judgments; a subordinate works independently within a narrower scope, reports upward, and accepts review.

AI does not change the structure. It changes the definition of a qualified subordinate.

Before AI, a qualified subordinate had relevant experience, reliable execution, some judgment, obedience without stupidity, and accumulated industry resources. Society produced such people in large numbers. Universities, HR systems, training, and labour markets all supported that profile.

After AI, most abilities in that profile are eaten. The remaining valuable ability is running a judgment loop independently within a given scope: receive direction and boundaries, decompose the problem, collaborate with AI, judge AI output quality, decide next steps, execute, and feed back.

This kind of subordinate has always been rare. In the industrial era, many “near-qualified” subordinates could still be used because experience, fluency, and execution compensated. Process, supervision, and middle management made them productive.

AI removes the compensating abilities. What remains is the ability to run the loop. The rarity level did not change; what changed is that these people are now the only qualified subordinates, not merely excellent ones.

Previously a team could contain ten near-qualified subordinates and one or two who could truly run loops. Now only the loop-runners are worth keeping for cognition-intensive work. Everyone else is done better and cheaper by AI.

Two kinds of collaborators

In the AI era, worthwhile collaborators fall into two classes.

Parallel collaborators are people who can run full judgment loops independently. They join collaboration not to be commanded, but as independent nodes in a larger system. The filter is severe: independent judgment, aligned direction, matching cycle rhythm, consequence-bearing, no alignment ritual.

Qualified subordinates are not independent directional thinkers, but can run a complete loop inside a given scope. A superior gives direction and boundary; they decompose, work with AI, judge output quality, choose the next move, execute, and feed back. The key is that they can review AI output quality. A subordinate who cannot judge whether AI is right is harmful, because they pass hallucination upward as reality.

Both categories are much rarer than before. The first was always rare. The second is rarer because many formerly qualified subordinates no longer qualify under the new standard.

This is what “the collaboration pool is shrinking” means: subordinates who used to qualify no longer do. Not because they got worse, but because the qualification standard changed.

Chapter 3 · Why AI Efficiency Differs Across Organisational Forms

The first two chapters discussed the boundary of the firm and the re-screening of collaboration. This chapter turns to a concrete phenomenon: AI produces vastly different gains in different organisational forms.

The difference is not marginal. It is an order-of-magnitude difference. The same AI tool can make an independent person’s whole loop run, while inside a large company it often only speeds up isolated tasks, leaving overall output almost unchanged.

Most discussion says this is about who uses AI better. That is wrong. The difference is not tool use. It is organisational form.

The real mechanism of AI productivity

Most people treat “AI productivity” as “AI makes a task faster.” Code that took hours now takes minutes. This task-level acceleration is real and exists everywhere.

But task-level acceleration is not whole-output acceleration.

The qualitative change comes not from tasks becoming faster, but from the gaps between tasks collapsing toward zero.

An independent wants to change a product architecture. They imagine an approach, ask AI about side effects, change the idea, let AI implement a test, run it, see the result, adjust, and continue if the direction works. In one afternoon they may complete several judgment-execution-feedback-adjustment loops.

In a large company, the same loop requires a proposal, manager review, cross-team review, alignment meetings, follow-ups, priority queues, development, and more review. The loop may take weeks or months.

This is not an efficiency gap. It is a loop-count gap. For independents, AI accelerates the loop itself. For large companies, AI accelerates a small step inside the loop while the total loop length barely changes, because the bottleneck is not execution. It is decision and coordination.

So “AI productivity” points to different things in different forms. For independents it means loop acceleration. For large companies it means task acceleration. The first changes output by orders of magnitude; the second is marginal optimisation.

Four structural reasons AI fails to transform large companies

Once loop acceleration is understood, the reasons become precise.

I. Hierarchical decision chains

Large-company decisions require multiple layers, stakeholder alignment, risk assessment, compliance review, and priority coordination. AI cannot accelerate these because they are not steps that “produce a thing.” They are steps that get multiple people to agree.

AI can write a prettier proposal. It cannot make five stakeholders agree faster.

Even if AI speeds every execution step 5x, the project remains slow because execution may be only 20% of the total time. The other 80% is waiting: decision, alignment, review. Accelerating the 20% changes little.

This is Amdahl’s law applied to organisations.

II. Incentives make AI accelerate “looking like work”

At sufficient scale, it becomes hard to separate real contribution from performance. Evaluation systems degrade into proxies: reports, meeting performance, clear expression, alignment with a manager’s direction. These things look like output.

AI lowers the cost of producing them to near zero.

An employee who previously spent a day writing a report can now produce a better one in thirty minutes. The saved seven hours do not necessarily go to real work. They may go to more reports, because colleagues are doing the same and not doing it makes one look behind.

The result: more reports, more meetings, prettier slides, longer emails, denser Slack. The organisation looks busier, while real output stays flat or falls. Somebody has to consume all this AI-produced “work-like” material, so real work is crowded out.

An existing inefficiency begins operating faster.

III. Information silos keep AI from seeing the whole judgment problem

An independent’s use of AI is: put all relevant information into one context and ask AI to think about the whole problem with me.

A large company cannot do this. Information is scattered across tools, permissions, private spaces, and people’s heads. Access is layered. Many decisions depend on informal information: a superior’s mood, a competitor signal, a customer hint.

AI can only see fragments. It can write code, summarise a meeting, or polish an email. It cannot help with real judgment because the information required for judgment is unavailable.

An independent is different. They are the information node. Their mind, files, and AI conversations contain the relevant whole. AI can touch everything, so it can participate in judgment.

IV. AI antibodies

This item is under-discussed and has two sides.

One side is real risk management. Large-company caution around AI is often reasonable. Hallucinations at thousands-of-employees scale can cause real loss. Customer data in public models can cause compliance accidents. AI decision liability is unclear. Multi-tenant data isolation is genuinely hard. These are not excuses.

The other side is role self-protection. Many jobs inside large companies derive value from coordination, review, reporting, alignment, and compliance. These jobs act as information filters and decision intermediaries. AI threatens their core value. People in these roles, consciously or not, make AI harder to use.

The two sides merge into a familiar picture: legitimate risk concerns are amplified by role-protection incentives into tool-blocking processes. Risk, compliance, hallucination, and data-security issues are expanded into strict approvals. Official enterprise tools lag public tools by generations. “AI output needs human review” becomes a way to route AI back into human process.

The sign of healthy risk management is specificity: “in this business, the data-boundary problems are X, Y, Z.” The sign of role protection is endless process: “wait until compliance finishes reviewing.”

So AI is weakened in large companies by both technical/compliance reasons and organisational politics. They are hard to separate.

Why independents benefit so strongly

The independent individual is the physical limit of AI productivity because all obstacles disappear:

• Zero decision bottleneck. One person decides.
• No incentive to look busy. Output is everything.
• Information is concentrated. The person is the node.
• No AI antibodies. You do not stop yourself from using AI.

So the judgment-execution-feedback-adjustment loop can run densely and continuously. In the same time window, the number of completed loops can far exceed a large company’s. That is why one person plus AI can build products that used to require a medium-sized company. Not because the person is smarter, but because they inhabit an organisational form in which AI can actually work.

Keep the boundary strict

This does not mean large companies will be replaced.

Large companies may have moats, network effects, capital barriers, brand trust, distribution, regulatory capacity, and long-term customer relationships. AI productivity does not touch all of these. They may remain stable for years.

This chapter asserts only one limited claim: on the dimension of AI productivity, independents and small teams have a structural advantage over large companies. It cannot be extended into a prediction about the overall fate of large companies.

Even this limited claim reveals something many people have not seen: AI-era organisational competition is not about who uses AI. Everyone uses AI. It is about whose organisational form lets AI actually function. Most large companies cannot answer that question, because changing the form means dismantling what they are built on.

Chapter 4 · Old Management Fails, New Management Is Unknown

The first three chapters discussed organisational boundaries, collaboration, and AI productivity across forms. This chapter goes deeper: the industrial-era thing called management is no longer sufficient for AI-era cognition-intensive work, and the replacement has not yet formed.

The boundary from the previous chapter still applies. The discussion concerns cognition-intensive work on the contracting side. In capital-intensive, heavily regulated, large-scale physical operations such as manufacturing, medicine, infrastructure, energy, and government, Taylorist-bureaucratic management may remain necessary.

Every “failure” below is inside that boundary.

The four assumptions of the Taylorist-bureaucratic paradigm

The management paradigm formed after the Industrial Revolution: standardised tasks, division of labour, supervision, KPI, hierarchy, job descriptions. Most companies still use this skeleton. Its core assumptions are four.

I. Work can be decomposed into repeatable tasks

Therefore work can be standardised, measured, and supervised. Assembly-line actions can be broken down precisely.

II. Decision-makers and executors are different people

Decision-makers think: what to do, how, why. Executors do. Management transmits the decision-maker’s thought to executors and ensures execution.

III. Efficiency comes from division of labour plus supervision

Each person does one small part. Someone ensures each part is correct and fits together. Efficiency comes from fine division and effective supervision, not from everyone’s global judgment.

IV. Motivation comes from external reward and punishment

Employees are driven by pay, bonuses, promotion, and punishment. Management designs these mechanisms so behaviour aligns with organisational goals.

Together, these assumptions form the core of industrial-era management. MBA programs, consulting, organisational behaviour, KPI, OKR, and Scrum largely sit on top of them.

Why the assumptions fail in AI-era cognition-intensive work

I. Work can no longer be decomposed into repeatable tasks

AI has taken the repeatable parts. The human remainder is non-repeatable, high-judgment, context-sensitive work. Each case is unique and cannot be specified in advance.

“Should we change the product direction?” is not a task that can be decomposed mechanically. It requires market, users, competition, capability, and team status to be judged together.

II. The decision-maker/executor distinction collapses

In the AI era, anyone who can use AI is both decision-maker and executor. They judge, execute, receive feedback, judge again. Decision and execution become a continuous motion inside one person or small team.

The management-intermediary role that transmits decisions to executors loses its object.

III. Efficiency no longer comes from division plus supervision

High-judgment work cannot be divided without losing the whole. You cannot give person A the first half of a judgment and person B the second half.

It also cannot be supervised externally. Someone may think for two hours; from outside, you cannot see whether the thinking is deep or valuable. Supervision observes behaviour, but observable behaviour and judgment quality have decoupled.

IV. Motivation no longer comes primarily from external reward

The quality of high-judgment work depends more on intrinsic motivation, attention, and curiosity. A person genuinely curious about a problem invests far more than a KPI can require. A person without curiosity will not produce good judgment for any bonus.

External incentives work for low-judgment work. For high-judgment work, they are often ineffective or harmful, because they encourage performance rather than good work.

Old management loses its object

When these assumptions disappear, Taylorist-bureaucratic management loses the object it describes.

Its five core verbs, planning, organising, supervising, coordinating, motivating, each lose their target:

• Planning: work cannot be pre-planned into tasks.
• Organising: executors and decision-makers are no longer separate.
• Supervising: judgment quality cannot be observed externally.
• Coordinating: tiny teams and individuals coordinate internally through their work.
• Motivating: external incentives lose force for high-judgment work.

In efficient AI-era small-scale cognition-intensive work, old management words do not map cleanly to reality.

What does an independent “manage” every day? Not AI, because they collaborate with it. Not themselves, because that word is odd at one-person scale. Not a project, because they are simply doing it. The vocabulary does not appear in the practice.

A 2- or 3-person team does not “manage” one another. They collaborate, divide, sync, and collide. These are not synonyms for management. Management implies power gradient and supervision. Collaboration and collision are horizontal.

This is not a decline in cost-effectiveness. The conceptual frame has lost its premise.

Five specific failure zones

Inside the contracting zone, old management fails in identifiable places.

I. Periodic coordination rituals

Weekly meetings, biweekly meetings, monthly meetings, quarterly reviews assume work rhythm is linear and roughly predictable.

AI-era output is nonlinear. One morning may finish what used to take a week, then two days may be spent stuck on a hard problem or building a tool for the next loop. The rhythm is determined by problem structure, not calendar.

Periodic rituals impose external rhythm on real output. During high-output periods they interrupt loops. During hard periods they create pseudo-progress.

The failure: they treat time as the coordination unit. In the AI era, the coordination unit should be loop completion, not a week passing.

II. Performance evaluation based on observable behaviour

Traditional management uses observable proxies: attendance, response speed, meeting participation, task completion, visible activity.

These were reasonable proxies for repeatable work. In high-judgment, nonlinear work, observable behaviour and real output are almost completely decoupled. One person may stare at a screen for eight hours and produce nothing. Another may walk for three hours and return with the decisive judgment.

Such evaluation does not only fail. It creates reverse incentives by rewarding people best at producing observable activity.

III. Project management built on task decomposition

Scrum, Kanban, Jira, and ticket systems assume work can be pre-decomposed into tasks with inputs, outputs, and measurable progress.

This holds for known, repeatable work. It fails for high-judgment work, where decomposition is itself part of the work, task boundaries are unclear, and progress cannot be measured by task count.

For AI-era small teams, forcing Scrum can reduce output. Real work becomes compressed into ticket-shaped tasks, while the crucial work, judgment, trade-off, structural reconstruction, cannot be recognised.

IV. Collaboration processes aimed at alignment

Much management seeks alignment: everyone understands the same thing, moves in the same direction, shares priorities.

This was reasonable when 500 people had to work on one thing. In small high-judgment units, over-alignment is harmful. Each person is making real judgments. If everyone is aligned into one direction, only one judgment remains active. The organisation loses the value of plural judgment.

High-quality collaboration is not alignment. It is collision. Different judgments collide and produce new judgment. Alignment kills collision.

V. Vertical responsibility allocation

Traditional management allocates responsibility vertically. A manager is responsible for a business area; subordinates are responsible for assigned tasks.

In AI-era small teams, responsibility cannot be vertical. Everyone is both judge and executor. AI participates in almost every step. “Who is responsible for this decision?” becomes harder to state. With small teams and no middle layer, there may be no object of upward reporting.

Accountability is not cancelled. It is replaced by self-accountability and outcome accountability. Each person is accountable for their judgment quality; the team is accountable for final output.

But this replacement cannot be expressed in the old management frame. You cannot draw “this person is accountable for the quality of their own judgment” on a traditional org chart.

The old fails; the new is unknown

We have to admit one thing honestly: the failure zones are visible, but we do not know what should replace them.

There have been candidates: Holacracy, self-organising teams, Spotify’s tribes/squads/chapters, Haier’s RenDanHeYi, independent-node networks. Each has its story, and each has worked in some places while failing to replicate broadly.

The reason is simple. Most replacements still adjust old words: team, manager, alignment, OKR. They do not replace the underlying frame. A real replacement would require new organisational primitives, tools, legal forms, education systems, and language. None are ready.

So the honest statement is not “new management is X.” It is we do not know what new management is. We know:

• old actions are failing: weekly rituals, observable-behaviour metrics, task decomposition, forced alignment, vertical accountability
• new actions are emerging: loop-based coordination, judgment-quality review, outcome accountability, autonomous unit collaboration
• they have not yet condensed into a named system
• anything claiming to have found “the AI-era management method” should be treated with suspicion

This is not pessimism. Around 1900, factories had already replaced workshops for decades, but Drucker-style modern management theory would not appear until 1954. Decades of attempts and failures came in between. We are in a similar position.

Admitting that the new is unknown is closer to reality than forcing a systematic-looking answer.

The limit of language

This chapter had to use old words to describe the failure of old management. That itself is a symptom.

“Coordination” is more horizontal than “management,” but does not include judgment-quality review. “Sync” is more reciprocal than “reporting,” but does not include directional decision. “Judgment bandwidth” is more accurate than “KPI,” but remains descriptive rather than operational.

We are experiencing a lag in language. New organisational forms are emerging, but the vocabulary has not been invented. Every analysis using old words is partly inaccurate, including this one.

This is not a personal failure. It is the cognitive debt of the era. We are discussing AI-era organisations with industrial-era vocabulary.

Admitting the limitation is more honest than inventing new words too early. Forced neologisms hide confusion. What we need is a clear awareness of the confusion, letting language form slowly as reality reshapes it.

Chapter 5 · The Distillation Curve of Jobs and Knowledge

The previous chapters covered contracting organisational boundaries, collaboration-pool screening, AI productivity differences, and obsolete management concepts. This chapter asks a more concrete question: how does the value of jobs inside organisations change in the AI era?

The answer is counterintuitive and sharp.

They remain, but not necessarily for the reasons people think

Look at any company today. Most employees are still in their roles: working, completing tasks, delivering output, attending meetings, receiving pay.

Ask a manager what value these people have, and the answer will be experience, fluency, knowledge of company processes, client relationships, industry familiarity.

This answer is not wrong. Some of that value is real and hard for AI to replace: tacit knowledge not verbalised or documented, real trust relationships, political sense inside an organisation, the intuition that “something here is off.”

But a large share of the answer is domain knowledge AI has not fully distilled yet, and that share is often mistaken for the first category.

The two must be separated.

Industry rules, process details, public cases: AI already knows much of the public version, and is learning the organisation-specific context quickly. This is distillable value.

Fluency: AI does not tire or make routine mistakes, and its cost falls. Much structured work is already beyond median human fluency.

External relationships: some parts can be documented, such as who someone is, communication style, history. Some parts cannot: real trust, tacit cooperation, willingness to do things outside contract. The first is being distilled into organisational knowledge bases; the second is hard-to-distil value.

Process understanding: “how the process runs” is distillable. “How to handle grey zones, when to use informal routes, who can bypass whom” is harder.

So people remain in jobs partly because of genuinely hard-to-replace value, and partly because the distillable knowledge attached to them has not yet been extracted. Organisations and employees often mistake the latter for the former.

This is not an insult. These are qualified industrial-era workers. In stable systems doing repeatable, supervised, measurable work, their value was real. They are not stupid or lazy. They fit the role.

But for many roles, the role itself is disappearing, especially where distillable value is high and hard-to-distil value is low.

The mechanism of distillation

Their knowledge is being distilled through several simultaneous mechanisms.

I. They distil themselves when using AI

Every time they use AI for work, even as a search engine, they make tacit knowledge explicit. Their prompts, questions, and edits expose what was previously inside their head.

They know, for example, that a certain client likes concise writing. When prompting AI to draft an email, they add “this client likes concise messages and dislikes long ones.” That tacit knowledge becomes text.

Once turned into text, it can be recorded, learned, and copied. The next person can give AI the same context without needing the original employee.

Each use of AI leaks a bit of their value. They do not feel it because it seems like providing context for better output. Structurally, it is a net loss of role value.

II. They distil themselves while collaborating

Explaining things to colleagues, describing situations in meetings, and documenting processes all make tacit knowledge explicit. If documents, meeting recordings, and Slack messages enter the company knowledge base, the knowledge becomes retrievable and learnable by AI.

III. They are distilled through documentation pushes

Many companies now push “knowledge management,” “process documentation,” and “tribal knowledge capture.” The stated reason is retaining organisational knowledge. The effect is extracting knowledge once bound to people and storing it in systems the organisation can use. Once extraction is complete, the people formerly binding that knowledge become more replaceable.

This is not conspiracy. HR and operations teams in large companies are doing it deliberately, even if they do not fully see the final effect.

IV. Industry-level data is collected by AI companies

One company’s knowledge may be local, but industry-level knowledge is being collected through financial reports, industry reports, professional forums, public discussions, contracts, and cases. Each year, foundation models understand industries more deeply.

So even if employees do not distil themselves, their industry’s collective knowledge is being distilled. Once a model’s understanding exceeds the average practitioner in a field, the value of the practitioner’s “domain knowledge” goes to zero.

Different knowledge distils at different speeds

The exact number of years cannot be stated precisely. Model iteration, multimodal capability, robotics, and industry depth are all changing. But the relative order is stable.

First to be distilled:

• procedural knowledge
• details of public rules, regulation, tax, compliance
• common-question answers, customer support, routine consulting
• template work: standard contracts, reports, designs

Next:

• domain judgment experience
• common customer-specific knowledge
• industry common sense known to insiders

Still valuable for longer:

• grey judgment in complex situations
• political judgment involving multiple interests
• trust work built through long relationships
• precise work in the physical world

Hardest for now:

• true creative judgment
• relationships themselves
• deep interaction with the physical world

One caveat: the physical-world moat has an expiry date. In 2026 it is real because embodied intelligence, general robotics, and precision manipulation are not mature enough. But it is not permanent. Multimodal models and robotics are progressing faster than many expect. Physical work should be seen as relatively safe for a period, not as a permanent refuge.

Different jobs face different time windows. Textual procedural work has the shortest. Complex grey judgment and relationship work have longer windows. True creative work is largely outside the current distillation threat. Physical work is safe for now, but must watch the next technology wave.

A counterintuitive implication

The framework yields a counterintuitive implication:

High-end knowledge work may be hit more directly by AI.

The common assumption is that AI replaces from low to high: simple work first, complex work later, high-end work last.

That is wrong. AI replaces from textifiable to non-textifiable.

Many high-end knowledge jobs are highly textifiable: strategy analysis, investment research, legal documents, consulting reports, academic research. Their outputs and processes are textual, so their knowledge is easy to distil.

Many apparently “low-end” jobs are harder to distil: a senior salesperson’s feel for a client, a mechanic’s bodily sense of machinery, a nurse’s intuition about patient status, a boss’s observation of employee state. Much of this knowledge is tacit, embodied, and situational.

So the middle manager, a high-education, high-salary, high-status role, may be one of the fastest value-decaying roles in the AI era. Much of the work is textifiable judgment: reading reports, making comments, writing feedback, holding meetings, sending emails. The AI-replaceable share is expanding quickly.

By contrast, a genuinely creative designer, a deep relationship-based salesperson, or a precision technician may be more resilient.

Again, the physical-world safety claim is time-bound to 2026. Embodied intelligence and robotics are still far from replacing skilled technicians, but the distance is shrinking. “Embodied = safe” is not a permanent conclusion.

This flips the value map of organisations. Value is no longer distributed by rank. It is distributed by whether the knowledge is distillable. Most organisations, and many employees, have not noticed.

Time can no longer be ignored when evaluating a role

The distillation curve adds a simple observation: current output is no longer an independent variable when evaluating a job.

In the past, a role was evaluated by current value. The hidden assumption was that this value would remain roughly stable.

In the AI era, the assumption fails. Two roles may create the same value today. If one is sitting on fast-distilling knowledge and the other on stable hard-to-distil value, their real value differs completely. The former is a depleting inventory; the latter is a sustainable cash flow.

No precise formula is offered. Remaining time windows cannot be quantified accurately, and maintenance cost includes salary, occupied position, crowding out future roles, and organisational psychology. A fake formula would create false precision.

But the time dimension must enter the discussion:

• A role that still appears valuable but derives most value from distillable knowledge may have book value far above long-term value.
• A less glamorous role built on hard-to-distil ability may be systematically undervalued.
• Organisations often handle the two backward, keeping the first because it “still produces” and marginalising the second because its output is less visible.

Admitting that the time window cannot be calculated is more honest than inventing precision.

The reverse mechanism: after distillation, what remains rises in value

Distillation devalues many jobs. It also creates an opposite mechanism: after a class of knowledge has been widely distilled, what remains undistilled becomes scarcer and more valuable.

Several forces overlap.

I. Recursive digestion of AI training data

Foundation models have consumed most public text, code, and professional content. Further progress requires rarer data: firsthand judgments in real situations, nonstandard operations in complex contexts, tacit knowledge never publicly written down. People who can generate such data hold an appreciating asset.

II. AI content explosion turns “not yet distilled” into a signal

AI-generated text, code, reports, and design saturate markets. When medium-quality textifiable output is almost free, “this clearly did not come from AI” becomes a scarce signal. Not because AI technically cannot produce it, but because it comes from an undigested angle, an undistilled judgment path, or concrete experience without public samples.

III. The other side of distillation is anti-distillation

As more capacities become AI-mediated, their opposites become independently valuable: hand-made, personally judged, actually present. This is already visible in cultural products, high-end services, and trust situations. It is not nostalgia. It is scarcity redistribution.

This does not refute the distillation analysis. It completes it. Most distillable work devalues; the undistilled remainder becomes scarcer and more valuable. The two ends pull apart. They are not symmetric: distillable value falls faster than undistilled value rises. From society’s total view, most jobs face net devaluation. For a few roles, value rises.

This explains a strange observation: some top roles become more expensive in the AI era. Not because AI did not hit them, but because AI flattened the surrounding replaceable positions, leaving the genuinely irreplaceable role scarcer.

What it means for people

From the individual’s perspective:

You are not replaced by AI because you are not hard-working or not excellent enough. You are replaced at a speed determined by how textifiable your knowledge is. This is not a moral judgment. It is a property of the knowledge.

It is distributed not by dignity, effort, or goodness, but by a colder dimension: how much of what you do can be turned into text.

The “assembly-line worker” of the future will increasingly be a white-collar worker doing textifiable judgment: customer support, junior analysts, middle execution, clerical roles, ordinary programmers.

They are not “wrong.” They occupy jobs that were reasonable when they entered them, and those jobs begin disappearing in mid-career.

The honest response is not to pretend those positions will last, not to beautify remaining value, and not to push people into vague “adapt to AI” rhetoric that often cannot work. It is to give them enough time and resources to prepare before the position disappears.

That is not softness. It is honesty closer to respect than comforting fiction.

Chapter 6 · Conclusion

The three essays describe one thing at three levels.

Issue 1: individual cognitive divergence. Most people mistook knowledge and experience for cognitive ability. When AI flattens the former and exposes the latter, a cognitive inequality long hidden becomes visible.

Issue 2: how individuals locate themselves inside that divergence. Not a methodology, but a return to learning’s essence: continuously exposing judgment to reality’s feedback and using AI to accelerate that process rather than its disguises.

Issue 3: how that divergence becomes deeper structural change through organisational form. In the contracting zone of cognition-intensive work, collaboration pools shrink, old management fails, and job value is distilled; while in capital-intensive, regulated, high-contract-risk zones, firms may become more necessary.

All three are one thing. At bottom: AI dismantles the external mechanisms that bound ability and output together in the industrial era: credentials, jobs, processes, experience, fluency, and being well-informed. It exposes, for the first time, the underlying ability that determines judgment quality.

Individuals see whether their cognitive ability is real (Issue 1) and how it may be built (Issue 2). Organisations see whether their form can still operate (Issue 3).

The two interact. Individuals with real cognitive ability will move toward organisational forms that amplify them. Organisational forms that cannot let AI function will gradually fail to retain people with judgment. This will not happen overnight. But it has begun.

Questions this essay intentionally does not answer

For completeness, it is necessary to list what this essay does not answer.

I. If large companies are structurally disadvantaged in AI productivity, why are many still growing?

Because AI productivity is only one competitive dimension. Distribution, network effects, regulation, brand, long-term customers, and capital barriers may remain strong. “Large companies lose on AI productivity” does not mean “large companies collapse overall.” These are different axes.

II. If I work in such a company, or manage dozens of people, what should I do?

The honest answer: this essay cannot answer that.

Any universal three-step answer is likely false. The real situation depends on organisation, industry, role, and personal leverage. This essay can provide diagnosis: which zone you are in, how much job value comes from distillable knowledge, which moat your organisation is strengthening. Choosing from diagnosis is more reliable than choosing from a packaged solution.

III. What is “new management”?

We do not know. Old actions are failing. New actions are appearing. They have not formed a named system. This may last a long time, just as modern management theory appeared decades after factories replaced workshops.

IV. What should individuals prepare?

Issue 2 gave the part it can: rebuild learning, expose judgment to feedback, avoid confusing AI fluency with ability. Anything further depends on each person’s starting point, resources, and risk tolerance.

Admitting unanswered questions is more honest than manufacturing answers.

This series offers no solution

The three essays offer no solution.

• No chicken soup about becoming an AI-era winner.
• No consulting advice about transforming your company.
• No checklist for AI-era learning.
• No prediction of “the future.”

Their job is not that. Their job is to describe what is happening, as precisely and honestly as possible, while acknowledging what cannot be inferred.

After reading, you may not know what to do. You may see your current reality more clearly. Decisions based on clearer reality are more likely to be right than decisions based on solution packages.

That is the only thing this series hopes to provide: a clearer description of reality.

One last question

Return to the final self-test from Issue 1:

When was the last time you genuinely changed an important view, and why did it change?

Now add another:

The organisation around you, your team, and the people you collaborate with: which era’s logic are they operating under? How much time does that logic have left?

If you can answer concretely, you have at least seen reality.

If you cannot, you may still be outside it.

Either way, time is moving.

Appendix · Starter Reading List

Several core lines in this essay have classic readings: why firms exist, the history of management paradigms, and attempts at alternative organisational forms. Grouped by topic and roughly ordered from easier to harder.

Why firms exist, Chapter 1’s Coase framework

• “The Nature of the Firm,” Ronald Coase (1937), short and dense.
• The Economic Institutions of Capitalism, Oliver Williamson.
• The Visible Hand: The Managerial Revolution in American Business, Alfred D. Chandler.

The formation and limits of management paradigms, Chapter 4

• The Principles of Scientific Management, Frederick W. Taylor.
• The Practice of Management, Peter Drucker.
• Out of the Crisis, W. Edwards Deming.
• The Fifth Discipline, Peter Senge.
• The Effective Executive, Peter Drucker.

Alternative organisational forms

• Reinventing Organizations, Frederic Laloux.
• Holacracy: The New Management System for a Rapidly Changing World, Brian J. Robertson.
• Materials on Haier’s RenDanHeYi model.
• Maverick, Ricardo Semler.

Structural problems of change

• The Innovator’s Dilemma, Clayton Christensen.
• Seeing Like a State, James C. Scott.
• Out of Control, Kevin Kelly.

Deep work and the organisation of knowledge work

• Deep Work, Cal Newport.
• A World Without Email, Cal Newport.

Practical note: these books are not solution manuals to copy directly. Their value is helping you see which paradigm’s vocabulary your organisation is living inside, and what assumptions that paradigm carries. Seeing that is more important than finding a “new management method,” because the new paradigm does not yet exist.

References and Sources

Sources for the main factual claims in this essay.

The theoretical basis for firms, Chapter 1

• Ronald H. Coase, “The Nature of the Firm,” Economica 4, no. 16 (1937): 386-405. Introduces the transaction-cost framework.
• Oliver E. Williamson, The Economic Institutions of Capitalism (Free Press, 1985). Develops asset specificity, opportunism, and other core concepts in new institutional economics.
• Coase received the 1991 Nobel Prize in Economics; Williamson received the 2009 Nobel Prize.
• “Legal-contractual power” and “brand/trust backing” are present in the broader Coase/Williamson frame but separated here as modern organisational-practice extensions.

Resistance to AI implementation in large organisations, Chapter 3

• Amdahl’s law: Gene Amdahl, “Validity of the Single Processor Approach to Achieving Large-Scale Computing Capabilities,” AFIPS Conference Proceedings 30 (1967): 483-485. Originally about parallel computing; used here for the general rule that system acceleration is bounded by the share of the accelerated component.
• McKinsey, The State of AI annual reports (2024-2026), documenting large-enterprise AI deployment patterns in which task-level gains often fail to translate into whole-organisation output change.
• “AI antibodies” is not an academic term. It names an empirical phenomenon here, partly inspired by Clayton Christensen’s analysis of how large companies resist disruptive innovation in The Innovator’s Dilemma.

History of management paradigms, Chapter 4

• Frederick W. Taylor, The Principles of Scientific Management (Harper & Brothers, 1911).
• Peter F. Drucker, The Practice of Management (Harper & Row, 1954).
• Peter M. Senge, The Fifth Discipline (Doubleday, 1990).
• The rough 1900-1954 interval from factory dominance to Drucker is used as an analogy for the possible gap between “old management failing” and “new management condensing.”

Alternative organisational forms

• Brian J. Robertson, Holacracy: The New Management System for a Rapidly Changing World (Henry Holt, 2015).
• Frederic Laloux, Reinventing Organizations (Nelson Parker, 2014).
• Henrik Kniberg & Anders Ivarsson, “Scaling Agile @ Spotify” (Spotify internal whitepaper, 2012). Spotify later acknowledged serious problems with scaling the model.
• Haier’s RenDanHeYi: Zhang Ruimin’s writings and Harvard/California Management Review case studies, including Frynas et al., “Haier’s RenDanHeYi,” California Management Review (2018).
• Ricardo Semler, Maverick (Warner Books, 1993).

Documentation and distillation of knowledge work, Chapter 5

• Ikujiro Nonaka & Hirotaka Takeuchi, The Knowledge-Creating Company (Oxford, 1995), a classic account of tacit and explicit knowledge in organisations.
• Michael Polanyi, The Tacit Dimension (1966), the source for the tacit-knowledge distinction used here.
• Villalobos et al., “Will we run out of data? An analysis of the limits of scaling datasets in Machine Learning” (arXiv:2211.04325, 2022 and 2024 updates), on high-quality text data exhaustion and recursive data digestion.

The innovator’s dilemma and organisational self-binding

• Clayton M. Christensen, The Innovator’s Dilemma (Harvard Business Review Press, 1997). The analysis of why large companies systematically fail under disruptive technology informs Chapter 3’s account of AI inside large organisations.

Coordination costs and modern collaboration tools

• Cal Newport, A World Without Email (Portfolio, 2021), on how modern white-collar work is submerged by collaboration rituals.
• GitLab’s remote work handbook (about.gitlab.com/handbook/) as a source for asynchronous collaboration practice.