Every Company Is a Story: What Sapiens Teaches Us About AI Governance
- Joseph Noujaim
- May 30
- 7 min read
The story no one wrote down for them
Every company runs on a story. That story is what lets a junior analyst in São Paulo and a senior partner in Singapore act in the same firm's name, on the same day, and still cohere. It rarely lives in the handbook. It lives in the side comments at the all hands, the corrections in the hallway, the things that get praised and the things that get quietly walked back. Human employees absorb it slowly, almost without noticing. The firm holds together because they do.
A new kind of employee has arrived, and that story has not been written down for it. AI agents now close tickets, move money, draft contracts, and answer customers in the company's name, at speeds and scales the old rituals were never designed for. They have never sat through the meeting where the firm decided what it does and does not do. The numbers look good. The applause is polite. And somewhere in the room, a quieter question is forming: by what story did the agent know what good work was? That question is the subject of this essay, and the answer is sharper than the applause suggests.
A feeling without a name
Most organizations cooperate at a scale that should not, on paper, be possible. Tens of thousands of strangers across dozens of countries make decisions every day in the name of the same enterprise, and the decisions cohere well enough for the enterprise to function. The official explanation points to policies, processes, controls, and training. The honest explanation is stranger and more interesting.
The reason a junior analyst in São Paulo and a senior partner in Singapore can act on behalf of the same firm is not that they have read the same handbook. It is that both of them have absorbed, gradually and largely unconsciously, the same story about what the firm is for, what counts as appropriate conduct, and how to fill in the blanks when the rulebook runs out. The story is invisible most days, and it is what holds the cooperation together.
AI agents now act on behalf of that same firm, at scale, in real time. The question of whether they have access to the story, in any form they can actually read, has not yet been seriously answered.
Harari and the trick that built civilizations
Yuval Noah Harari, in Sapiens, offers a striking explanation for how human beings ended up running the planet. The advantage was not bigger brains, sharper tools, or stronger bodies. The advantage was a specific cognitive capability: the ability to invent, share, and continuously revise fictions that exist only in the collective imagination, and to coordinate the behaviour of large numbers of strangers around those fictions.
Tribes, gods, nations, currencies, human rights, and limited liability companies are all examples of the same underlying mechanism. None of them exist in any way that could be touched, weighed, or photographed. All of them produce real, observable behaviour in the physical world, because enough people behave as if they are real.
The story is not a lie. It is a coordinating device, and it is the device on which civilization runs.
Peugeot, and the company that lives in a document
Harari's favourite illustration is Peugeot SA. The company has factories, employees, and a logo, but the company itself is not any of those things. The factories could burn down and the employees could be replaced. The company would continue to exist, because the company is a legal fiction sustained by ritual. A lawyer signs an article of incorporation, a state recognizes a new corporate person, courts agree to enforce contracts in its name, and from that moment onward, millions of people act as if the entity is real.
Every enterprise runs on a variant of this trick. Beyond the legal charter, there is a thicker, less visible artifact: the operating story of the firm. What it stands for. What it does not do, even when it would be profitable to do it. How its people are expected to read a situation that no policy covers. This thicker story is what controls fall back on whenever rules run out, and rules always run out.
Human employees inherit that story slowly. They watch colleagues get gently corrected in meetings, they hear what gets praised at the all hands, and over time they internalize a working model of how the firm wants to be represented in the world. The story becomes part of how they decide.
An AI agent does not have that runway. It arrives with the prompt, the tools, and the API keys, and it begins acting in the name of an entity whose operating story has rarely been written down in any form it can actually read.
The 150 threshold, and the limits of informal trust
Harari also points out a more mechanical fact. Below a population of roughly 150, human groups can hold themselves together through gossip, personal acquaintance, and direct observation. Above that number, intimate ties stop being enough, and coordination begins to require explicit shared fictions: law, money, ideology, ...
Organizations encountered this threshold long ago. Small firms run on personal trust and proximity. Large firms run on codified roles, written policies, and a layered control architecture that takes the place of the conversation a founder used to have over the workshop floor. Behaviour controls, output controls, and clan controls are all attempts to replace, at scale, the intimate-band coordination that does not scale.
AI agents now operate well above that threshold, often from day one. They are deployed across business units, geographies, and time zones, acting on behalf of an organization whose codified controls assume a human reader behind every action. The threshold Harari names has been quietly crossed again, and the mechanism that worked for the last crossing, codified shared fictions absorbed through tenure, is the mechanism the new actor cannot use.
The Neanderthal problem, in reverse
There is a particular passage in Sapiens that lands hard when read with AI in mind. Harari argues that Neanderthals lost not because they were weaker, but because they could not participate in shared fictions at scale. They could probably warn each other about predators. They likely could not organize fifty strangers around a story about a tribal spirit. That small cognitive gap was decisive.
AI agents present the mirror image of that problem. Where the Neanderthal could be enrolled socially but not at scale, the AI agent can be enrolled at scale but not socially. It can act in the name of the enterprise across millions of transactions, and it cannot be enrolled in the story that makes those actions legitimate. Execution without enrollment is exactly the configuration the legal and managerial inventions of the last few centuries were designed to prevent for human agents.
Naming the inversion clarifies the engineering task. AI agents will not be made more trustworthy by being made more human. That path is closed. They will be made governable by writing the story in a form they can read, against which their actions can be checked, and through which deviations can be surfaced for human judgement.
Fictions are kept alive by ritual
Harari makes one further observation that matters here. Shared fictions do not stay alive on their own. They are continuously re-enacted. The court renews the corporate person each time it enforces a contract. The flag renews the nation each time it is raised. The all hands renews the firm's story each time the CEO reaffirms what the firm does and does not do.
Without the recurring ritual, the fiction fades. With it, the fiction stays usable. The same will be true of any computable mandate placed in front of an AI agent. A static document on a shared drive will not bind a system that acts a thousand times a day. The mandate has to be invoked at deployment, checked at runtime, enforced when crossed, and revised when the environment shifts. The ritual moves from the boardroom into the loop, but the structure is recognisable.
This is the practical implication of taking Harari seriously. Governance of AI agents is not primarily a question of better filters. It is a question of whether the story the enterprise lives by has been made legible to a new kind of actor, and whether the ritual that keeps that story alive has been extended to include it.
A question the field is now equipped to ask
This line of inquiry is not a new invention. It sits inside a long tradition of organization control theory and agency theory, which have spent decades describing how enterprises grant authority to actors and verify that authority is honoured in practice. Those traditions assumed, reasonably, that the actors in question were human, socialized, and reachable through ordinary managerial means.
The arrival of capable AI agents introduces a new class of actor into that tradition, one that does not learn norms through tenure and does not absorb correction through social signals. Current research, including the doctoral programme behind this essay, aims to extend organization control theory to that new actor, and to build practical methods enterprises can adopt before drift becomes the kind that requires cleanup rather than course correction. A working methodology already exists in draft form, and it will be introduced properly in posts that follow.
An invitation for AI Governance research
The simplest way to summarize the argument is to borrow Harari's own framing.
Every company is a story. The story is what allows strangers to act in its name. The story has always been kept alive by ritual, codified in artifacts, and renewed by the people who care enough to keep renewing it.
A new kind of stranger now acts in the company's name, at speeds and scales the old rituals were not designed for. That stranger cannot be socialized, cannot be gossiped about, and cannot be expected to read the story between the lines. The story will have to be written down for it, kept alive in front of it, and revised when it begins to drift.
Readers drawn to that framing are warmly invited to follow along, to challenge the parallels drawn here, and to share where their own organizations have already met this problem under a different name. The research is being written in public on purpose, so that the field can sharpen the answer together.
Every company is a story. Someone is going to have to write the next chapter for the agents. Join the aigovernanceblog.com and contribute to the research.