What health insurance claims taught me about operational systems

A claim is a small object that touches a dozen systems before it resolves: the EHR that produces it, the clearinghouse that transmits it, the carrier that adjudicates it, the patient record that links back to it, the AR pipeline that books the payment when it arrives. Each touch is a chance for the data to drift.

What I learned working with claims data at Infinedi is that most "data quality" problems are actually identity-resolution problems. Two records look like the same person but aren’t. Two records look different but are. The patient changed addresses, the carrier changed providers, the encounter was re-keyed under a slightly different date of birth. Software that doesn’t handle this carefully creates downstream chaos.

The lesson generalizes far past healthcare. Any operational system that touches multiple parties has to assume the identifiers don’t agree. The cheap fix is to require a perfect join key. The right fix is to make the system robust to imperfect ones — and to surface, explicitly, the cases the system isn’t sure about.