From the average to the case.

What data does

Data, however vast, describes populations. It tells you the average breach probability, the average treatment effect, the average loss for a portfolio of this composition. Statistical models — including the LLMs trained on enormous text corpora — are designed to recover that average. They do it well.

What decisions ask

The question that arrives in front of a decision-maker is never about the average. It is about this claim, this patient, this portfolio, this control. The averaging-out is what makes the model useful for general questions and silent on the one in your inbox.

What closes the gap

Causal structure does. A causal model encodes which variable causes which — the mechanism the data is silent about. With that structure in hand, the model can reason from the population the data describes to the individual the decision concerns: condition on everything specific to this case, ask what would happen under a different choice, and return the answer for this case rather than the typical one.

Where the structure comes from

Not from data. From people — the underwriters, clinicians, engineers, and analysts who have learned, over careers, what causes what in their domain. The structure is theirs; the model encodes it. When they retire, it stays.

Read further

For the long-form version of this argument — with the gap, what does and doesn't close it, and Pearl's three-step procedure for reasoning about an individual case — see What Data Cannot Tell You.

For a worked case where it changes the decision, see Insurance Attribution — The Expert Who Was Right for the Wrong Reason.

For the formalism, see Rung 3 — Counterfactual Reasoning — The Three-Step Procedure.