Ask any client advisor at a bank how eager they are to document their work and their client book, and the answer will be clear: not particularly excited. The advisor’s knowledge of who calls whom, which negotiation is stuck, and which client tolerates which terms is bargaining power over their employer, social capital with their clients, and finally economic capital which is portable across jobs. Sharing it strips them of all three. So the book and the advisor’s most relevant information stay in their head, and the organization stays closed. I recently experienced this first-hand over drinks with friends who work in such a bank.
In large corporations, employees keep secrets from each other everywhere. Executives plan restructurings behind closed doors, salespeople hide leads and the status of negotiations from one another, and one person may hold deeply rooted, old implicit knowledge about legacy software or certain operating procedures. In each case, the result is a lot of implicit knowledge distributed across an organization, inaccessible to outsiders, other departments, newcomers, or the technically unfamiliar.
It has long been thought that this closed organization was a “moral failing” that technology could not fix. Generative AI now promises to fix it. Before that, Ray Dalio’s radical transparency paradigm at Bridgewater Associates demanded a culture where feedback is immediate, frequent and radically direct, and where coworkers codify knowledge in a manner accessible to almost everyone [1]. Most companies never went that far; the discipline was unaffordable. Now it seems self-evident that companies should create a unified knowledge base through AI to be faster, reduce mistakes, and become more efficient. Companies such as Palantir implement data layers that serve as a digital twin of the organization [2]. Cleansing and codifying existing data becomes much simpler, and agents can scan large amounts of data from the far ends of the organization to make sense of the problem at hand: past deals with one entity help predict which prospects will close; managers use the results of previous reorganizations to inform the next wave of restructuring. Tobi Lütke of Shopify has an AI chief of staff, which he says is now common among tech people [3].
But this optimism sees only one part of the problem. The closed organization is a power problem wearing a technical costume, and deploying agents will not solve it. AI sharpens it on both sides.
Consider first the individual. Organizational data, especially non-public data, has gained relevance since the value of high-quality human data for model training and reinforcement learning became clear. Buying the data remains of bankrupt companies is an old practice now gaining traction with AI companies [4]; data brokerage has become a hot topic after prominent cases like the pending sale of Spirit Airlines’ data [5] or the settlement over the mass download of pirated books by Anthropic [6]. The more valuable the organizational data layer becomes, the more expensive it becomes for the individual to feed it. Refusing to document is rational for the client advisor, because codification is expropriation. AI strengthens the incentive to hoard knowledge as it raises the price of the hoarded good.
Consider next the organization. Some secrecy is legitimate. Among individual contributors, shielding knowledge originates in motives of personal gain. In management it can be necessary: existential scenarios and decisions are discussed there (new business projects, shutting down departments or projects, potential competitive moves), and there, shielding information is also in the interest of the company, as are the necessary Chinese walls between public and non-public information in an investment bank. The organization that opens everything is defenceless, because vital information leaks too widely.
Therefore, the agentic organization’s central question is who may know what. Information that should not be broadly accessible will end up in the organizational data layer anyway, through the carelessness of individuals, model failures, and the cost of screening so much data. Agents that structure knowledge also review their own conversations and rewrite their own instructions at night (Shopify calls this process dreaming [3]) and may publish sensitive knowledge in unexpected places along the way. Yet, a “ground truth” across the organization is the whole point of the agentic layer; without a way to establish it, the layer loses much of its use. Model safeguards, as we know them from biosecurity or cybersecurity, do not solve this: when too many requests are rejected, the agentic layer becomes useless. The type 1 / type 2 error trade-off is unavoidable. In the long run, information will be far more accessible within a company unless we establish clear rules of access and make judgements about who may know what.
There will be so many of these judgements that agents will have to make most of them. But in my view, machines are bad at exercising judgement. Lütke says they can inform a decision but cannot take responsibility for it [3]. Management consultants play the same role for executives. One emerging field of study will be the cognitive biases of LLMs [7], just as Kahneman and Tversky explored those of humans [8]. People already apply the heuristics of human teams to LLMs. Letting several agents debate sounds like an attempt to avoid “groupthink” and “bring in diverse perspectives” [9]. If there were one truth to be surfaced through research, we would only need one agent. But there is still ample judgement needed, so we mix and scramble diverse perspectives to enable our own.
Is this a human way to enable judgement? More broadly: what is judgement, really? One definition is that judgement is choosing one pattern among competing heuristics, in a complex system without clear rules, by reacting to implicit cues. If that is even roughly right, then “who may know what” is just the kind of decision judgement exists for. And soon, machines might have to make it millions of times, although they neither hold the pattern nor bear the responsibility.
Organizations entering the agentic age should therefore settle two questions before they structure their data: who decides who may know what, and who is accountable when the machine gets it wrong. This decision needs an owner, because “no” is the engineered exception while “yes” is the machine’s natural state: an agent’s whole purpose is to surface relevant knowledge, and it declines a request only where someone deliberately built and tuned rules to decline it. Even Palantir, the vendor of the organizational digital twin, sells mainly this: an ontology that governs how agents, models, and users are allowed to read or change data [10]. Those rules will in turn be tuned towards “yes”, because an agent that rejects too many requests is an agent nobody uses. If no one owns the answer, the machine will decide instead, and it will usually say yes.