Making Assumptions Explicit: The Most Underrated Step in Enterprise AI
Why reasoning systems require a new approach to governance
Every model is a collection of assumptions. This is not a controversial statement, but it is a deeply uncomfortable one for most enterprises. We like to think of our models as objective, data-driven artifacts, as mathematical machines that ingest data and output truth. But the reality is that every model, from a simple linear regression to a complex deep learning network, is built on a foundation of human beliefs about the world. We assume that the future will look something like the past. We assume that the variables we have chosen are the ones that matter. We assume that the relationships we have modeled are stable. We assume that the data we have collected is a faithful representation of reality.
For the most part, we get away with ignoring these assumptions. In a stable, predictable world, the assumptions that held yesterday are likely to hold today. But in a world of rapid change, of market shifts, of black swan events, those hidden assumptions become a ticking time bomb. A model that was once a reliable guide can become a source of catastrophic error, not because the math is wrong, but because the world has changed in a way that invalidates the model’s core beliefs.
This is the central challenge of building intelligent systems in the enterprise: it is not a technical problem, but an epistemological one. It is the problem of how we know what we know, and how we adapt when what we know is no longer true. And it requires a new set of tools and a new way of thinking about governance.
Traditional model governance is focused on the model itself. We have model validation teams, we have model risk management frameworks, we have processes for monitoring model performance. But we have very little in the way of assumption governance. We have no systematic way of surfacing, documenting, and testing the beliefs that are encoded in our models. Those beliefs are often hidden in the code, buried in the documentation, or, most dangerously, live only in the heads of the people who built the model.
This is where the tools of causal inference, particularly Directed Acyclic Graphs (DAGs), can be surprisingly powerful, not just as a modeling technique, but as a governance tool. A DAG is more than just a picture of a model; it is a picture of our beliefs about the world. It is a formal, explicit representation of our assumptions about what causes what. When we draw an arrow from “underwriting standards” to “loss ratio,” we are making a clear, testable claim about the causal structure of our business. When we omit an arrow between “marketing spend” and “customer retention,” we are making an equally strong claim: that we believe these two things are not directly related.
Viewed in this light, a DAG is a kind of organizational mirror. It reflects the collective beliefs of the team that built it. The process of building a DAG forces a conversation that rarely happens in most organizations: a structured, rigorous debate about the way the business actually works. It forces us to move from vague, qualitative statements to precise, formal claims. It forces us to confront the areas where we are uncertain, where we disagree, where our mental models diverge.
But simply drawing the DAG is not enough. The world changes. The assumptions that were true last year may not be true this year. This is the problem of assumption drift, and it is a far more insidious threat than the more commonly discussed problem of data drift. Your data may be perfectly stable, but if the underlying causal relationships in your business have changed, your model is still wrong. A pricing model built before a major regulatory change, a claims model built before a global pandemic, an underwriting model built before a sudden shift in consumer behavior—all of these are examples of assumption drift in action.
How do we manage this?
We need to start treating our assumptions as first-class citizens in our data stack. We need to build systems for belief management. This means surfacing assumptions—every new model or analysis should begin with a formal process of articulating the underlying assumptions. This could be a structured document, a collaborative whiteboarding session, or, ideally, a formal representation like a DAG.
It means versioning beliefs—just as we version our code and our data, we need to version our beliefs. When we update a model, we should be able to clearly see which assumptions have changed. This creates an auditable history of our thinking, not just our code. It means testing mental models—we need to build a culture of actively trying to break our own models. This means running sensitivity analyses, testing for the impact of unobserved confounders, and constantly asking the question: “What would have to be true for this model to be wrong?”
This is where the agentic architecture we discussed last week becomes so powerful. An agentic system can help to formalize this process. The Hypothesis Agent can proactively suggest areas where our assumptions may be breaking down. The Causal Agent can build models based on different sets of assumptions, allowing us to compare them. The Validation Agent can automatically run a battery of tests to stress-test our beliefs. And the Narrative Agent can communicate the results of this process in a clear, uncertainty-aware way.
Consider a practical example in underwriting. For years, a commercial property insurer has operated on the assumption that the age of a building is a primary driver of fire risk. This assumption is encoded in all of their pricing models. But a series of large losses in newer buildings calls this assumption into question. In a traditional environment, this might trigger a one-off analysis that takes months. In an agentic system, the process is continuous. The Data Agent flags the anomaly. The Hypothesis Agent suggests that the relationship between building age and fire risk may have changed. The Causal Agent builds a new model that incorporates a new variable: the type of electrical wiring. The Validation Agent confirms that the new model is more predictive. The Narrative Agent communicates the finding: “Our long-held belief that building age is a primary driver of fire risk is no longer supported by the data. The evidence suggests that the type of electrical wiring is a more significant factor, particularly in buildings constructed in the last 10 years.”
This is more than just a better model. It is a better way of thinking. It is a system that is designed to adapt, to learn, and to challenge its own assumptions. It is a system that treats epistemology not as an academic curiosity, but as a core business function.
The work of building intelligent systems is not just about data and algorithms. It is about the careful, disciplined management of our own beliefs. It is about building organizations that are not just data-driven, but also self-aware. And it is about recognizing that the most valuable asset we have is not our data, but our ability to reason with it in a world that is constantly changing.