From volume to selection: the trap that AI agents are forcing organizations to solve
There is a belief that runs through the corridors of almost every organization that has invested in artificial intelligence over the past eight years. The belief that the problem is always one of quantity. More data. More tokens. More coverage. More stored history. As if intelligence were proportional to volume, and the solution to any system failure were simply to add more.
This conviction was not born out of naivety. It was born from the era of big data, a time when accumulating information was technically difficult, expensive, and therefore valuable in itself. Whoever had more data had an advantage. Whoever could process it, even more so. The model was simple and had clear market logic.
What is happening now in organizations that are deploying AI agents in production is forcing an uncomfortable revision of that premise. The problem is no longer data scarcity. Medium-sized companies in mature sectors have accumulated trillions of tokens across CRMs, databases, documents, emails, support tickets, internal communication threads, and legacy systems. The problem is that agents do not know what to do with that volume. Not because they are incapable of processing it, but because no one has taught them to filter. And that inability to select is not a technical problem. It is a problem of organizational design that companies have been evading for years with the excuse that they first needed more data.
The illusion that more context is better context
There is a structural difference between feeding a model everything available and giving it access to the exact fragment it needs to act well at this specific moment. The first option seems safer because it feels complete. The second requires that a difficult decision has already been made: knowing what matters and what does not.
That decision is costly because it forces someone in the organization to commit to a hierarchy of relevance. And committing to a hierarchy of relevance means accepting that some things do not matter as much as we thought, that some data we have spent years collecting does not change the outcome, that some sources a given area defends as critical are in practice nothing but noise.
Not many organizations are willing to have that conversation. Not because they cannot. But because it carries an internal political cost that no one wants to assume. The result is that agents receive bloated contexts, with contradictory information, without a clear hierarchy, and they produce responses that are technically plausible but operationally useless. The failure is attributed to the model. The conversation that was never had remains intact.
What is emerging as a response to this problem has a technical name: context engineering. It is not a prompt optimization practice, although on the surface it may appear to be. It is the discipline of deciding, with organizational judgment, what information an agent receives to execute a specific task. It involves structured search to extract precise facts from formal systems, semantic search to recover meaning from unstructured content, and inverted indexing to locate exact identifiers in real time. Three distinct retrieval layers, each with a different function. None of them replaces the other. Together, they convert accumulated knowledge into usable context.
The problem is that implementing this correctly requires that someone in the organization has previously defined what is relevant for what type of task. And that is not an engineering problem. It is a knowledge governance problem that most organizations have never resolved in an explicit way.
What context graphs reveal about organizational maturity
The next frontier in enterprise agent architecture carries another name: context graphs. The distinction from conventional knowledge graphs is precise and worth pausing over.
A knowledge graph models what exists: entities, relationships, taxonomies, ontologies. It tells the agent how the organization's conceptual world is structured. It is useful, but insufficient. An agent that knows an exceptions approval process exists does not, by that knowledge alone, know how those exceptions are resolved in practice, who has real authority to approve them in ambiguous situations, what informal conversation thread generated the decision that is today encoded as policy, or what workaround the operations team has been using for two years because the formal process does not work.
Context graphs capture that procedural layer. They record decision traces: who approved what, in what order, using what tools, with what result. They build a persistent organizational memory that includes not only the current state of things, but the path that led there.
The implication is significant for those who lead organizations, not only for those who design them technically. An organization that can build useful context graphs is an organization that has been capable of making its own decision-making process visible. That has named its real approval flows, its habitual exceptions, its escalation patterns. That has had the conversation about how decisions are actually made, not just how the organizational chart says they should be made.
Many organizations cannot build that layer because they have never articulated it. Not because the information does not exist, but because it exists distributed across informal conversations, in the memory of specific individuals, in undocumented practices that no one has had an interest in making explicit, because making them explicit would also make them auditable. And therein lies a tension that agentic AI projects are bringing to the surface with greater clarity than any previous process consultancy ever did.
The AI agent cannot operate with what the organization refuses to name. And the refusal to name is not always technical. It is often political. It is the protection of discretionary spaces that certain areas or individuals do not want to see formalized, because doing so would cost them a share of power or autonomy.
Why the pace of adoption predicts who will have the advantage, not who has it today
Gartner projects that more than 50% of AI agent systems in enterprise environments will use context graphs before 2028. It is a figure worth reading carefully, because it does not say that all organizations will use them well. It says that the majority will use them in some form.
The difference between using them in some form and using them well depends on something that cannot be resolved with a technology budget. It depends on whether the organization has been capable of doing the prior work of articulating how it makes decisions in a granular and honest way. Organizations that arrive at 2028 with context graphs built on top of formal processes that no one actually follows will have sophisticated agents that replicate dysfunctions with greater efficiency. Organizations that have done the uncomfortable work of mapping their real flows — including the informal ones, the ones nobody documents because they are convenient precisely because they are opaque — will have something qualitatively different: an institutional memory that can learn.
The competitive advantage in AI agents will not belong to whoever deployed the most models or whoever has the most tokens stored. It will belong to whoever knew how to filter first. Whoever built systems capable of identifying the exact fragment of context that changes the result of a specific decision. And that, in practice, is an organizational capability before it is a technological one.
It is worth considering what happens in the opposite scenario. An organization with hundreds of agents operating in parallel, each one building its own fragmented and inconsistent view of how the company works, generates a kind of chaos that is not immediately visible but is structurally corrosive. The agents contradict each other. The decisions made by one are not coherent with those made by another. Institutional memory does not accumulate — it fragments. And when something goes wrong, no one can clearly trace what context a given agent received and why it acted as it did. Governance collapses at precisely the moment it is most needed.
Selection is the discipline that organizations have not yet learned
There is something that the evolution of the past eight years in enterprise AI confirms with considerable consistency. The problem was never data scarcity. It was the resistance to deciding what matters.
Deciding what matters carries a cost. It means that some areas receive less attention from the system than others. It means that some data sources representing years of accumulated work do not enter the operational context of the agents. It means that someone has to commit to a hierarchy and defend it in the face of those who disagree.
That conversation, in most of the organizations I know, never took place explicitly in the context of AI strategy. It was avoided with the implicit promise that the system could handle everything if it were given sufficient computing capacity. What AI agents are now making evident is that this promise was never viable. Not because the computing power is insufficient, but because the intelligence that an agent can deploy is limited by the quality of the context it receives, and the quality of that context is not a function of volume. It is a function of the clarity with which the organization has been capable of articulating what it knows and how it uses that knowledge.
The organizations that manage to build that clarity will not do so because they found the right technology platform. They will do so because someone in a leadership position had the will to force the conversation that others were avoiding, to name what the system preferred to leave unnamed, to commit to a hierarchy of relevance that carries a real and visible political cost. That is the capability that cannot be purchased with an infrastructure budget. And it is, for now, the scarcest of all.
