The Real Job of AI Is to Kill Bad Hypotheses Fast
We started from a common hypothesis while building a pharmacy location model on medical data: places crowded with pharmacies are oversaturated and risky. We had an AI agent test it against tens of millions of prescriptions, closure histories, and population statistics. The verdict was blunt. Density explained neither pharmacy revenue nor closures, and it held across three resolutions: district, neighborhood, and radius. The explanatory power was essentially zero.
The 0.93 accuracy the demo bragged about was an illusion created by spatial leakage. The AI stripped that illusion away with spatial generalization and survival analysis. It honestly settled in a few hours what a person would have to spend days doubting.
AI Does Not Know Where to Look
Had we stopped there, the conclusion would simply have been: it does not work. What changed direction was one remark from someone who knows the field: pharmacy revenue ultimately comes from prescriptions at the clinic next door, and prescription volume differs by department.
With that one sentence we rebuilt the logic. Instead of density, a prescription-demand index: the prescription strength of nearby clinics divided by the number of competing pharmacies. Run again, revenue separated cleanly, a twofold gap between the bottom and top tiers.
The Formula: The Expert Picks Where, the AI Checks Whether It Is Real
A common way to misunderstand AX is to expect the AI to do everything. In practice it is the opposite. The expert knows where to sell. The AI quickly verifies whether it is real. Neither alone is enough.
The side effect was large too. We assembled the whole analysis pipeline, from a cloud database through public APIs to survival analysis, in a single session. The friction cost of forming and testing a hypothesis approaches zero. So you get to be wrong more often, and faster.