By STEPHEN FARBER

Part one argued that patient access was never the finish line; it was the foundation. This is about what we build now that we are standing on it.

Every good game show has a second round, and the second round is always harder. The first round was about getting health information out of the buildings that held it. Regulation, standards, and a decade of unglamorous plumbing have brought us to an important threshold. Individuals can increasingly pull records from multiple institutions into one place. The board paid out. The audience cheered.

Then a quieter and stranger question arrives. You are holding years of visits, laboratory results, medications, and discharge summaries, and you realize that having your history in one place is not the same as understanding it. Assembly was the first hard problem, but it was not the last.

The record is finally readable

This is where artificial intelligence walks onto the stage, and the promise is real. Few clinicians have time to read a patient’s entire history every time a decision must be made. Appropriately safeguarded AI can summarize years of encounters, support medication reconciliation, explain a laboratory trend in ordinary language, and surface a pattern buried across multiple providers. Large consumer platforms have already begun connecting AI assistants to medical records; Amazon’s Health AI, for example, draws on a person’s One Medical history to provide personalized explanations and guidance. Once the record can be assembled, a capable assistant can finally make more of it legible.

Legibility, however, is not reliability. The farther AI moves from explaining a record toward acting on it, the thinner the evidence becomes. A 2026 scoping review of agentic AI in healthcare found that most of the available studies were exploratory and lacked robust clinical validation, and that only one trial involved patients. An assistant working from an incomplete, duplicated, or subtly wrong record can produce a fluent and plausible conclusion that is still wrong. AI does not make data quality less important. It makes the consequences of poor data quality harder to ignore.

AI cannot read what was never written

There is a more fundamental limitation, and it has less to do with the model than with the record itself. A longitudinal record can tell us what happened, but it rarely tells us enough about why. It preserves the encounter, the diagnosis, the result, and the prescription while losing much of the reasoning that surrounded them. Why was one treatment chosen over another? What was the person hoping to preserve at that point in life? Who else was helping? What changed after the visit, and what did a family member notice before anyone thought to schedule another one?