Epic Is the Best EHR I Have Used. We Should Expect More.

Nov 12

The next medical record should help us learn from every patient—and let others compete to make that possible.

A recent Becker’s Hospital Review article described nine business rules Judy Faulkner refuses to bend. Epic’s founder has kept her company private, rejected acquisitions and insisted on developing an integrated product. I admire that independence. I also wonder whether the principles that built the dominant electronic health record should determine the limits of what comes next. [1]

I have used several electronic health records. Epic is, by far, the best. No system is perfect, and I would not pretend its competitors have solved the problems that frustrate me. But being the best available system should invite higher expectations.

My test is simple. I can ask Google or a conversational AI how many Chinese restaurants there are in New York City and quickly receive an answer. Yet when I wanted to know how many supracondylar elbow fractures I had pinned over the preceding year, obtaining that information from my own clinical practice was far less intuitive.

These are common childhood fractures just above the elbow. I had treated the children, performed the operations and entered the documentation. Asking the record to count those cases should have been straightforward.

Of course, a fast answer about restaurants may be wrong. A clinical answer needs a defined population, reliable records and a way to check the result. I want that rigor. I also want to be able to ask the question without first learning how the database thinks.

And the count is only the beginning. How many children needed another operation? Who recovered full movement? Did my results improve? What happened to those who never returned?

Those are questions about whether I am doing my job well.

The limits of integration

One rule in the Becker’s article particularly caught my attention: Faulkner’s refusal to sell software into a system design she considers wrong. She recalled turning away hospital leaders who wanted separate vendors for pharmacy, nursing and documentation. They eventually chose Epic’s integrated system. [1]

I understand the reasoning. As a surgeon, I want the medication list, operative record and discharge instructions to agree. I have no desire to replace a functioning system with a collection of incompatible products.

But I would distinguish integration from dependence on one supplier. We should be able to preserve a coherent clinical record while allowing different companies to compete over how best to use it.

The scale makes this question consequential. KLAS’s 2026 market analysis, as reported by Fierce Healthcare, placed Epic at 43.7% of US acute-care hospitals and 56.9% of beds, based on 2025 activity. This is substantial dominance, although I would not use those figures alone to declare a legal monopoly. [2]

Nor can we dismiss Epic as a billing system that works only in America. King’s College Hospital and Guy’s and St Thomas’ introduced it together in 2023. Its use in the NHS undercuts that familiar criticism. [3]

My concern is more specific: how much should healthcare’s capacity to learn depend on the choices of its leading software supplier?

Epic has an answer

It would be unfair to suggest Epic has ignored this problem. SlicerDicer offers self-service analysis, and its SideKick assistant can generate reports from plain-language questions. Cosmos combines data contributed by participating health systems to support research and clinical insights. Epic also documents interfaces for outside applications and large-scale data export. [4–6]

There is newer evidence of that direction. On October 21, Epic announced Connectors, designed to let its AI tools retrieve information and perform tasks in other applications. The company says developer-created Connectors will become available through its Showroom in March 2027. [7]

These are meaningful developments. They also sharpen the test. Connecting outside applications to Epic’s AI is one form of openness. Allowing an independently developed clinical system to compete on useful, authorized access to records is another. We should ask what each arrangement permits in practice.

My own difficulty does not prove a feature is absent. Local permissions, implementation and training may explain it. But a capability advertised by a vendor and one a physician can readily use are different measures of success. Hospitals share responsibility for closing that gap.

We should judge the result at the point of care: can I ask a sensible question, understand the answer and verify the evidence?

What the data must tell us

I want a record that helps us examine the consequences of our decisions. For each elbow fracture, it should connect the injury, treatment, recovery, complications and the family’s account of the child’s function. It should tell me where information is missing.

A child who never returned cannot automatically count as a successful result. A second operation performed elsewhere should not disappear from the assessment because another institution holds the record. These are basic requirements for the system I would want to build.

More data will not, by itself, produce better medicine. In a 2019 Science paper, Ziad Obermeyer and colleagues showed how a widely used health algorithm produced racial bias by using healthcare spending to predict need. Unequal spending made cost a poor substitute for illness. [8]

That lesson matters here. If we train systems to reproduce what healthcare has billed, we should not assume they have learned what patients need. We must define the outcomes that matter and test performance in the people we intend to serve.

I want AI to compare cases, identify patterns worth investigating and help us test whether changing care improves results. I do not want it to confuse an association with proof that one treatment caused a better outcome.

The promise is considerable. The proof must come from clinical results, access and cost—not the size of the dataset.

Privacy is part of the design

It is tempting to imagine letting a powerful AI search every medical record and discover what we have missed. The ambition is understandable. Unrestricted access is not a sound plan.

The record contains intimate information entrusted to healthcare. A patient’s permission to use information for treatment should not become a blank check for commercial model development.

We should build around defined purposes, appropriate authorization, limited access and records of who used the information. Methods such as federated learning can allow institutions to collaborate on model development while keeping raw records locally, although they do not remove every privacy or security risk. [9]

Nor should we describe Epic as simply owning all these data. Its software, hospital decisions and patients’ interests play distinct roles. The practical questions are who can authorize access, which uses are permitted, what access costs and whether another qualified developer can participate.

Those questions deserve clear answers. Privacy should impose real duties on anyone handling records, including the incumbent and its challengers.

What I would build

I would begin with a focused clinical service. Before a visit, it would gather the patient’s history and available records, identify missing information and help determine what needs prompt attention. After treatment, it would seek the outcomes needed to assess whether care helped.

This is the direction I want to pursue through my own work developing AI-enabled clinical services. The ambition is a record organized around the patient’s course, with useful information available before, during and after each encounter.

It would show clinicians where each answer came from. It would flag uncertainty. It would let patients contribute information and correct errors. And it would support approved tools from different developers, with the same expectations for safety and performance.

A challenger could start by proving that one such service improves care within existing systems. Over time, it might earn a larger role, including replacing functions now supplied by the EHR. Replacement should remain possible if a new system proves better.

By universal, I mean a record that can follow the patient across institutions and support different tools. I would not want to replace Epic’s dominance with a requirement that everyone depend on my company, or any other single company.

I believe AI will eventually perform much of the cognitive work physicians do today, and that robots will assume more physical tasks. That is my forecast, not a claim that autonomous medical practice is ready. Such systems will need reliable clinical histories, meaningful outcomes and evidence that they work safely.

The same requirements would improve medicine now.

Faulkner’s independence helped build a remarkable company. We should bring that same willingness to question convention to the system she built. Epic should have every opportunity to lead the next stage, and others should have a fair opportunity to do better.

I want to know how many children I treated, how they fared and what I should change. Their experience should help me care for the next child. That is what I expect a medical record to make possible.

Disclosure: I am developing Texican Signal, an AI-enabled clinical service. The capabilities described here are proposed goals.