Why AI needs a supply-side facts layer for care
As agents start acting on people’s behalf, the cost of acting on wrong provider data rises. Here’s why source-linked, timestamped facts become infrastructure.
For a decade, the supply side of care — who provides what, where, under what license, at what price — has been “good enough” as messy, scraped, star-rated data. Good enough for a human skimming a directory. Not good enough for software that acts.
That threshold is moving. When an AI agent books a telehealth consult or routes a patient to a clinic, a wrong or stale fact isn’t a bad search result — it’s a bad action. The tolerance for unsourced data collapses.
Three properties that matter
A facts layer for regulated care has to be source-linked (you can see where each field came from), timestamped (you can see when it was last checked), and honest about gaps (an unknown field is empty, not invented).
Those three properties are unglamorous. They are also exactly what lets a regulated product or a grounded model cite the data instead of guessing.
Why this is hard, and why that’s the point
The official records exist — NPPES, LegitScript, state boards, the FDA — but they’re scattered and inconsistently structured. Turning them into one queryable schema, with provenance preserved, is real work. That work is the moat.