The Commune Returns: Why AI Needs Human Scale

What happens when a company can produce almost anything, but can no longer decide what is worth producing?

This is the strange organizational reality emerging around artificial intelligence. Software agents can draft, research, code, analyze, summarize, and execute. Yet some of the most aggressive users of these systems are not shrinking their teams. They are hiring. One highly automated company grew from four people to thirty and continued recruiting, even as agents became woven into nearly every workflow.

At first glance, this seems like a failure of automation. If machines do more work, should humans not do less?

The deeper answer is that automation does not eliminate work evenly. It makes certain forms of competence abundant while making judgment, coordination, taste, trust, and responsibility more important. When yesterday’s difficult output becomes cheap, the organization is flooded with possibilities, drafts, experiments, and nearly acceptable results. The scarce resource is no longer production. It is knowing what to pursue, what to reject, and how to turn machine capability into something other people can rely on.

This is where an unlikely organizational tradition becomes relevant: the commune.

The historical commune was often organized around three ideals: egalitarianism, human scale, and resistance to bureaucracy. These ideas were attempts to solve a problem created by industrial society. Factories could coordinate more people and produce more goods, but they could also make human beings feel interchangeable, distant from decisions, and trapped inside administrative machinery.

Artificial intelligence creates a similar tension in reverse. It can make production intimate and instantaneous while making the surrounding system vastly more complex. The question is no longer simply how to automate work. It is this:

How do we build organizations powerful enough to use abundance without becoming too large, opaque, and bureaucratic for human beings to understand?

The answer may require treating the commune not as a political relic, but as a design pattern for the age of intelligent machines.

Automation Makes Output Cheap and Meaning Expensive

Imagine a small publishing company before generative AI. Producing an article required research, drafting, editing, design, and distribution. Each stage imposed friction. That friction limited how many ideas the company could pursue. The team had to choose carefully because every project consumed scarce human hours.

Now imagine the same company with a dozen specialized agents. One can generate research notes, another can propose angles, another can draft, and others can create images, social posts, and audience analyses. The cost of starting a project collapses. The company can pursue ten ideas where it once pursued one.

But the cost of evaluating those ideas does not collapse at the same rate.

In fact, it may rise. The team now has to inspect ten drafts, compare five strategies, verify dozens of claims, resolve conflicts between agents, and decide which output is merely plausible and which is genuinely useful. The organization has replaced a production bottleneck with a selection bottleneck.

This is why more automation can create more human work. The machines are not necessarily failing. They are succeeding at the wrong layer of the problem. They generate outputs, but outputs are not outcomes. A plausible memo is not a good decision. A functioning feature is not a valuable product. A fluent argument is not a true one.

A useful way to model this is to divide organizational work into four layers:

1. Generation: producing possible answers, artifacts, or actions.
2. Verification: testing whether they are accurate, safe, and internally consistent.
3. Selection: choosing which possibilities deserve resources.
4. Integration: embedding the choice in relationships, systems, and consequences.

AI is rapidly lowering the cost of generation. It is improving at verification, but remains uneven, especially where context and responsibility matter. Selection and integration remain deeply human because they involve competing values, institutional memory, and consequences that cannot be inferred from a prompt alone.

The mistake is to assume that an organization’s workload is proportional to the amount of visible production. In reality, production is only one part of the system. When production becomes abundant, the other parts become more prominent.

The more cheaply an organization can create possibilities, the more carefully it must govern attention.

This is also why the distance between a person and an agent matters. An agent that works closely with a human can borrow context, receive correction, and inherit standards. An agent operating far from human judgment may generate a large quantity of work that is technically competent but strategically irrelevant. The result is not liberation from work. It is an expanding perimeter of supervision.

The Commune Was a Theory of Human Scale

The ideal of human scale is often mistaken for a preference for smallness. It is more precise than that. A human scale institution is one in which people can still perceive the connection between their actions, the decisions being made, and the consequences that follow.

A village may be small but authoritarian. A large institution may contain teams that operate with remarkable autonomy and mutual understanding. Scale is not only a headcount. It is a question of legibility.

Can a person understand who decides? Can they see how resources move? Can they challenge a bad judgment? Can they identify the owner of a mistake? Can they feel that their contribution changes the whole rather than disappearing into an anonymous process?

AI complicates these questions because it allows a small number of people to command an enormous amount of computational labor. Ten employees may effectively operate hundreds of agents, tools, and automated processes. In terms of productive capacity, the organization becomes large. In terms of human membership, it remains small.

This creates a new kind of scale mismatch. The organization may have the output surface of a corporation but the social structure of a studio. If its internal systems are designed like those of a large bureaucracy, it will drown in approvals, dashboards, and rituals that were created to coordinate humans who lacked direct access to one another. If it has no systems at all, it will drown in inconsistency and hidden dependencies.

The solution is not to reject scale. It is to separate operational scale from social scale.

Operational scale is the amount of work an organization can perform. Social scale is the number of people who must remain mutually intelligible for the organization to function. Intelligent agents can dramatically increase operational scale without requiring social scale to expand at the same rate. That is an opportunity, but only if the organization protects direct relationships and clear accountability.

Consider a software team with eight people and fifty agents. The agents write code, run tests, monitor systems, and prepare documentation. The team can accomplish what once required dozens of employees. But when something breaks, someone still needs to answer questions such as: Why was this feature built? Which tradeoff was intentional? What risk was accepted? Who has the authority to reverse the decision?

If the answers are buried in automated logs or scattered across tools, the team has become operationally powerful but socially unintelligible. It has achieved machine scale without human scale.

The commune’s lesson is not that every organization should remain tiny. It is that every organization needs a visible circle of responsibility. People should know the community to which they belong, the standards it upholds, and the decisions they are empowered to make.

Egalitarianism Becomes a Knowledge Problem

The egalitarian ideal also acquires a new meaning in an AI saturated organization. Traditionally, hierarchy helped institutions allocate authority. Senior people knew more, controlled more resources, and made more decisions. But AI weakens the connection between rank and access to competence.

A junior employee with strong tool fluency may investigate a market, build a prototype, or analyze a complex dataset faster than a senior executive. A small team can challenge an established expert by producing alternative models in an afternoon. The organization’s knowledge is no longer concentrated in a few people, and the old hierarchy becomes less reliable as a map of actual capability.

This does not mean hierarchy disappears. It means hierarchy must justify itself differently.

In an AI enabled organization, authority should increasingly be based on context, judgment, and accountability, not merely on possession of information. Information is becoming cheap. Responsible interpretation is not.

This distinction supports a healthier form of egalitarianism. Equality does not mean everyone has identical authority in every situation. It means that status should not determine who is allowed to see reality, question assumptions, or propose a better path. A person close to the work may have the best understanding of a problem, even if they occupy a lower formal position.

The practical risk is that organizations will respond to abundant machine capability by centralizing control. If agents can act quickly, leaders may be tempted to impose more permissions, more review gates, and more centralized policy. That creates a paradox: the tools make action easier, but the institution makes action harder.

A better pattern is distributed execution with explicit boundaries. Teams and individuals should be able to make many decisions locally, provided that they understand the constraints, the quality standards, and the situations that require escalation.

For example, a customer support team might allow agents to resolve routine cases, issue small credits, and identify recurring problems. Humans retain authority over unusual complaints, policy exceptions, and high consequence decisions. The organization does not need to approve every action. It needs to define the edges of legitimate action and make those edges easy to understand.

This is the difference between bureaucracy and governance. Bureaucracy tries to prevent every possible error by adding procedures. Governance creates shared principles that let people act without asking permission at every step.

Anti Bureaucracy Does Not Mean No Structure

The phrase “anti bureaucratic” can sound romantic until one remembers why bureaucracies exist. They preserve knowledge, coordinate strangers, ensure consistency, and create records. Removing all structure from an AI native organization would not produce freedom. It would produce invisible bureaucracy, where decisions are made by defaults, prompts, permissions, and undocumented habits.

The real enemy is not structure. It is structure that exists primarily to protect the structure itself.

AI makes this distinction urgent because automated systems can multiply procedural complexity at almost no immediate cost. A company can add another approval rule, another monitoring agent, another classification layer, and another reporting dashboard. Each seems reasonable in isolation. Together, they create an organization where no one can explain why work takes so long.

A useful test for any process is to ask three questions:

1. What failure does this process prevent?
2. Who is empowered to change or remove it?
3. What evidence would show that it is no longer useful?

If no one can answer, the process has probably become ceremonial. It is consuming human attention without producing corresponding trust or safety.

The anti bureaucratic organization therefore needs fewer rules, but better ones. Its core operating documents should be short enough to remember and specific enough to guide action. They should state what the organization values, what kinds of errors are tolerable, what kinds are unacceptable, and where human judgment is mandatory.

This also suggests a new role for agents. They should not merely automate tasks. They should make organizational reasoning more visible. An agent can record why a decision was made, identify which assumptions changed, surface disagreements between teams, and point out when a process has accumulated exceptions. Used this way, automation reduces bureaucracy by improving institutional memory rather than by eliminating oversight.

The best system is not one in which no one has to think. It is one in which people spend their thinking on the decisions that deserve it.

A Design Pattern for the Human Scale AI Organization

The synthesis of these ideas can be expressed as a simple organizational model: small human communities commanding large machine capacity.

Such a community has five characteristics.

First, it has a clear purpose that cannot be reduced to output volume. If the goal is only to produce more, agents will fill every available surface with work. A human purpose gives selection a standard.

Second, it maintains a compact circle of mutual accountability. Everyone should know who is responsible for major decisions, which people represent the organization’s standards, and how disagreements are resolved.

Third, it gives agents narrow roles with broad observability. An agent may be allowed to perform a task, but its actions, assumptions, and handoffs should remain inspectable. Autonomy without visibility is not empowerment. It is hidden delegation.

Fourth, it treats review as a creative function rather than a final inspection. Human review should shape direction early, not merely correct machine output at the end. The human contribution is often most valuable when deciding the question, defining quality, and choosing among plausible answers.

Fifth, it periodically removes machinery. Every automated workflow should have an expiration date, a review owner, or a clear signal that would trigger reevaluation. Otherwise, automation accumulates like sediment and eventually becomes the bureaucracy it was intended to replace.

This model changes how leaders should think about hiring. The question is not simply whether a candidate can produce an artifact faster than an agent. It is whether they can improve the organization’s ability to choose, learn, coordinate, and take responsibility.

The most valuable people may be those who can move between levels: understanding a customer deeply, directing agents precisely, spotting weak assumptions, and translating a local insight into a better system. Their value lies not in competing with machines at generation, but in connecting generation to meaning.

Key Takeaways

1. Measure selection load, not just production capacity. If AI increases the number of drafts, options, or experiments, track how much human attention is required to evaluate them. Automation is not successful if it merely moves the bottleneck downstream.

2. Design for social scale separately from operational scale. A small team can command enormous machine capacity, but it still needs clear relationships, decision rights, and visible accountability.

3. Replace status based authority with judgment based authority. Let proximity to the problem, quality of reasoning, and willingness to own consequences matter more than rank alone.

4. Write principles instead of multiplying procedures. Define the organization’s non negotiable standards and decision boundaries. Then let people and agents act locally within them.

5. Keep human review close to the beginning of the process. The highest value human contribution is often choosing the goal, framing the problem, and recognizing what good looks like, not polishing an already generated answer.

The Organization Is Becoming a Commune with Machines

The usual story of AI and work asks whether machines will replace people. That question is too blunt to be useful. The more important question is what kind of human association becomes possible when machines can perform so much of the productive labor.

One possibility is a larger bureaucracy: more layers of control, more automated surveillance, more distant decisions, and more people managing systems they cannot explain. Another is a return to human scale, not through technological retreat, but through technological leverage. Small communities could possess extraordinary productive power while preserving direct trust, shared purpose, and meaningful responsibility.

That future will not happen automatically. Machine abundance does not create humane institutions by itself. Without deliberate design, it simply magnifies the organization’s existing habits. A rigid company becomes more rigid at greater speed. A confused company produces confusion in greater volume. A thoughtful community gains leverage.

The deepest lesson is therefore not that automation creates more work. It is that automation changes what work is for. Once machines can generate nearly endless possibilities, the central human task becomes building a community capable of judging among them.

The winning organization may not be the one with the most agents, the fewest employees, or the fastest workflows. It may be the one that can remain small enough to understand itself while becoming powerful enough to shape the world.