Editor's note:
Welcome to Tinsight, Tabbit's column for in-depth writing worth spending time with. We have previously shared essays from offbook.press, including
The Real Divide in the AI Era and Rebuilding Individual Learning.
Our current series is The Breakdown of Organizational Forms. The full essay runs to around 20,000 Chinese characters, published in six installments. This is chapter two. In the previous installment, the author explored why companies exist and argued that AI is redrawing their boundaries more precisely rather than simply tearing them down. This chapter asks a further question: if the company is no longer the default structure, how do independent individuals, small teams, and loose networks actually collaborate?
Author: Dawei Geng
Original essay: The Breakdown of Organizational Forms
Chapter Two · Collaboration no longer happens on the same plane
After asking why companies exist, the next question is how people collaborate within these new organizational forms: independent individuals, small teams, and loosely connected networks.
The answer is counterintuitive. Collaboration in the AI era is becoming polarized rather than simply easier.
The pool of collaborators is shrinking
In the past, choosing a collaborator involved several dimensions:
Does this person have relevant experience?
Are they proficient at the work?
Do they have the necessary knowledge?
Do they have industry information and connections?
Can they exercise judgment?
These five dimensions used to be largely independent. Someone could be strong in some and weaker in others. A highly proficient person with 15 years of experience and extensive knowledge, but only average judgment, used to be a valuable collaborator. You could handle decisions yourself while relying on their execution, experience, and ability to integrate information. Their value was real.
The change in the AI era is not that their value falls to zero. It is that AI has absorbed almost all of the first four components.
Experience → AI knows the pitfalls across industries and explains them when asked.
Proficiency → AI does not get tired, make mistakes, or take leave.
Knowledge → AI is a living knowledge base.
Information and connections → Model training data covers everything publicly available.
Once these four are absorbed, what remains is cognitive decoupling: the ability to exercise genuine judgment.
If someone has strong cognitive decoupling skills, they remain extremely valuable, perhaps even more so, because judgment previously buried in execution can now be released.
If that ability is only average, their accumulated value collapses because AI has replaced its supporting pillars.
The calculation behind collaboration has therefore changed entirely:
Before: their experience + proficiency + information + knowledge + average judgment = a useful collaborator.
Now: AI already handles the first four + their average judgment = less useful than working on my own with AI.
The right side of that equation is decisive. The comparison is with yourself plus AI. That option did not previously exist. It is now powerful enough to make most experience-based partnerships less worthwhile.
The rising cost is screening, not collaboration
This is often misread as meaning that collaboration itself becomes more expensive in the AI era. That interpretation is wrong.
Working with the right person is smoother than before. Both people use AI and have strong judgment, so they can work asynchronously, use AI as a shared intermediary, and run frequent cycles.
What becomes more expensive is finding qualified collaborators.
Previously, many dimensions could qualify someone, so the candidate pool was large and finding a suitable person was inexpensive. Now the acceptable dimensions narrow and the pool shrinks sharply. Finding the right person in a small pool naturally costs exponentially more.
This is not growing arrogance or higher personal expectations. The underlying definition of a worthwhile collaborator has changed. Suppose 70% of previously qualified people relied on experience and proficiency. Their absolute value has not changed: they still know what they know. But AI has reduced their relative value until collaboration is no longer worthwhile.
You are screening a redefined pool, rather than simply becoming more selective.
Production systems become incommensurable
An analogy makes this change clearer.
Two textile workers could once collaborate well. They shared a production logic: feeling thread with their hands, inspecting fabric with their eyes, and taking hours to finish a bolt of cloth. They could understand, complement, and check each other's work.
After production lines arrive, a maintenance technician and a hand weaver occupy different production systems. The technician works at the scale of machines: inspecting, adjusting, and optimizing, with the output of keeping a hundred machines running well. The hand weaver works at the scale of a bolt of cloth, producing one bolt a day. Their timescales, objects of judgment, and units of output all differ.
Asking them to collaborate is a category error, like asking a farmer and an air traffic controller to grow crops together. The difficulty is that they cannot align their work, rather than one person simply slowing the other down.
The AI era is producing the same divide.
An independent person using AI may complete a cycle in an afternoon and a full product in weeks. Someone embedded in traditional processes without AI may work on a weekly meeting and a quarterly release. A twofold or threefold efficiency gap still allows collaboration. A gap of more than an order of magnitude becomes a production-system problem: timescales, decision objects, and units of output occupy different coordinate systems.
Ask someone who delivers once a month whether to ship a feature tonight, and your senses of time do not align. Tell someone who discusses every decision in a meeting that you just decided to replace the architecture, and your decision scales do not align. Tell someone waiting for requirements from a PM that you make product judgments yourself, and your models of responsibility do not align.
This is a systemic issue. After the steam engine appeared, factory engineers and blacksmiths were performing different kinds of work, rather than merely having trouble communicating.
Redefining the subordinate
The shrinking collaboration pool is particularly visible in the relationship between a manager and a subordinate.
This structure has always existed: a manager sets the direction while a subordinate works independently within a narrower scope, reports progress, and accepts review. It predates AI, extending through industrial, preindustrial, and imperial eras.
AI changes the definition of a qualified subordinate while leaving the structure itself intact.
A qualified subordinate used to look like this:
Relevant experience, allowing them to get started quickly.
Reliable execution, completing assigned work correctly.
Some judgment, avoiding problems within a limited scope.
Responsive without being unthinking, raising issues without overstepping authority.
Established industry contacts and resources.
Before the AI era, such people were plentiful. Recruitment, HR, universities, and professional training all helped produce them. They were relatively easy to find, with low screening costs.
After AI, most of those capabilities are absorbed by AI. The remaining contribution is the ability to run a judgment cycle independently: within a manager-defined scope, completing the full loop of judgment, execution, feedback, and adjustment without the manager intervening at every step.
Such people have been scarce in every era. Industrial organizations could also use many people who fell short of that standard but compensated with experience, proficiency, or execution. Processes, supervision, and middle management allowed them to contribute.
AI absorbs the contribution made by those supporting capabilities. What remains is the ability to run a complete cycle independently. The scarcity of people with that ability has not changed. What changes is that they become the only qualified subordinates, rather than exceptionally good ones.
A team could once consist of ten people who partly met the standard plus one or two who could run a full cycle. Now it can use only those able to run cycles: AI does the other work better and more cheaply.
Two kinds of collaborators
On this analysis, people worth collaborating with in the AI era fall into two groups:
The first is the peer collaborator: someone who can independently run a complete judgment cycle and participate as an autonomous node in a larger effort. The screening criteria are demanding: independent judgment, broadly aligned direction, compatible working rhythms, responsibility for consequences, and no need for alignment rituals.
The second is the qualified subordinate. They may not set their own overall direction, but they can run a full cycle within a defined scope. A manager provides direction and boundaries; they break down the problem, work with AI, assess its output, decide what comes next, execute, and report back. Assessing AI output is crucial. Someone who cannot tell whether AI is right or wrong can cause harm by passing hallucinations upward as facts.
Both groups are much scarcer than before. The first was always rare; most people who used to qualify for the second no longer meet the new standard.
That is what a shrinking collaboration pool means: previously qualified subordinates no longer qualify. The standard has changed, rather than the people becoming worse.
To be continued…
Chapter One · The overlooked question: what problem did companies originally solve?
Chapter Two · Collaboration no longer happens on the same plane
Chapter Three · Why AI productivity gains differ across organizational forms
Chapter Four · Old management fails; new management is still unknown
Chapter Five · The distillation curve of roles and knowledge
Chapter Six · Closing thoughts
