“Damn it, they’re good”

Why Epic’s Dominance Threatens to Overwhelm Healthcare’s AI Marketplace

“We’re an Epic-first organization.” When I ask healthcare organizations where they’re getting their AI tools, this is the answer I hear most often, including at my own hospital.

This doesn’t mean that a healthcare delivery system that employs Epic’s electronic health record (EHR) will use Epic-built AI tools exclusively. Rather, it means that our default setting is to stick with Epic’s AI, even when we know there are better tools out there. Epic has become healthcare’s version of IBM in the 1980s, when the mantra in IT departments everywhere was, “Nobody ever lost their job going with IBM.”

While there are real advantages to using AI tools that are automatically integrated into the EHR, this trend threatens to stifle innovation. Let’s explore why, and – as Julia Adler-Milstein, Sara Murray, and I describe in a recent JAMA article – whether anything can (and should) be done about it.

Epic’s Unlikely Rise to Dominance

While there are plenty of Epic haters (and even some Epic conspiracy theorists) out there, I’m not one of them. As I describe in A Giant Leap, I believe Epic is an extraordinary company that achieved market dominance largely by having the best product at a time (2008-2015) when every health system was installing EHRs – massively complex software packages that transformed every clinical, operational, and financial transaction. Epic has also forged strong relationships with its customers and improved its system over time in response to user feedback and changing demands.

The Barn, Epic Headquarters, Verona, Wisconsin

In my book, I describe Epic’s unlikely ascendence and why the EHR – Epic’s and everybody else’s – failed to deliver on the promise of making healthcare better, safer, more convenient, and less expensive. In the end, EHRs turned out to be mostly expensive digital filing cabinets – a necessary ingredient for the transformation of healthcare, but not the final answer. Moreover, by relentlessly demanding piles of additional documentation and facilitating in-box messages from patients, EHRs were also powerful contributors to clinician burnout.

Real transformation, it seemed, would require new tools that would take data from the EHR, analyze it, and serve as the scaffolding for new systems and processes to remake the experience of clinicians and patients. This requirement for additional insight and action layers is not unique to healthcare. The development of the internet didn’t automatically give rise to Airbnb, Netflix, Uber, or DoorDash. But none of these companies could have emerged without the digital foundation supplied by the internet and cloud storage.

Generative AI, particularly large language models (LLMs), turned the tide – enabling the conversion of EHR data into clinical and business intelligence, and facilitating new ways of organizing our work. This is because so much of healthcare’s data is unstructured (e.g., clinician notes) and so many of our processes are driven by conversations (e.g., collecting the patient’s history, asking a consultant for advice on a tough case). For the first time, LLMs gave us the skills we needed to deliver on the promise of digital transformation.

I won’t talk much about the role of the behemoths in this sweepstakes – legacy giants like Google, Amazon, and Microsoft, or newer ones like OpenAI and Anthropic. Suffice it to say that all these companies are rushing to offer tools that address key healthcare use cases – either by going directly to patients (ChatGPT Health, Claude for Health, Amazon Health AI) or providing AI infrastructure for healthcare systems, often as new capabilities embedded in their corporate clouds.

To me, the more interesting battle is the one between venture capital-funded AI startups and the EHR vendors themselves. As the dominant EHR vendor in the U.S., the central character in this latter category is Epic.

When AI Tools Are “Good Enough”

Epic recognized the opportunity to build AI functionality into its EHR well before the 2022 public release of ChatGPT brought generative AI to public attention. For its first major foray into the world of AI, Epic chose a sepsis prediction model – one that would signal when a patient had a high probability of having a life-threatening infection. This choice made sense – sepsis can be fatal, early recognition is key, and the rapid delivery of evidence-based therapies can save lives.

There was one problem with Epic’s sepsis prediction model, which the company rolled out in 2017: the tool’s performance was abysmal. In one study, University of Michigan researchers found that the tool threw off sepsis alerts for more than 100 patients who didn’t have sepsis for every patient who did, leading to lots of unnecessary wolf-crying and scores of patients receiving inappropriate therapies (e.g., antibiotics for patients with blood clots, IV fluids in patients with heart failure and fluid overload).

The 2021 Michigan study of Epic’s sepsis decision-support model

Impressively, despite the massive misstep, Epic didn’t shy away from AI – in fact, the company doubled down. Over the past few years, Epic has rolled out boatloads of AI tools – more than 100 in all – that do everything from predicting the risk of readmissions to helping patients interpret their lab data to summarizing a 500-page medical record.

A recent systematic review of Epic’s AI clinical decision support tools found fair, but not excellent, performance, with none achieving an area under the ROC curve above 0.79. While it’s hard to find comparable performance data for the thousands of AI tools being built by start-ups, it’s logical to assume that a tool built by an AI company that focuses on a discrete set of use cases might have real advantages over a tool built by an EHR company working on hundreds of different problems. Since more than 85% of Epic’s EHR customers are now using at least one of Epic’s AI tools, if that assumption is correct, it would mean that millions of patients are being exposed to AI tools – potentially high-stakes ones suggesting diagnoses or therapies – that are producing suboptimal results.

Many Systems Lean Toward Epic for Their AI

Even if Epic’s tools are imperfect, many healthcare systems find the company’s advantages to be compelling. These advantages include flip-on-a-switch integration with the EHR, the preexisting trusting relationship with Epic, and complete confidence that Epic will be in business in five years – something no start-up can guarantee. Moreover, choosing an Epic tool saves a health system’s IT department from wading through dozens of vendor pitches, struggling to sort reality from hype.

These advantages have profoundly tilted the playing field. One AI administrator at a large healthcare system told me, “If a third-party tool is an A-, and an Epic tool is a B, we’ll still go with Epic.”

Even when an AI tool is not ready for release, Epic typically announces it 6-12 months in advance at the huge summer user meeting at its Verona, Wisconsin headquarters – thereby freezing much of the market. “Epic is like the kid who licks all the cookies before his friends arrive at the birthday party,” one health system AI leader told me. A leader at an AI start-up that sometimes finds itself on the short end of the competition with Epic begrudgingly told me, “Damn it, they’re good.” I believe he was referring to both the quality of Epic’s recently released AI tools – there have been no sepsis-like missteps – and the company’s competitive chops in the AI marketplace.

(Note that when many people in healthcare say anything – particularly anything remotely critical – about Epic, they often insist on anonymity, another indication of the hold that the company has on the industry.)

A 2025 study found that among hospitals implementing AI, 79% chose tools supplied by their EHR vendor, most commonly Epic.

How Epic’s Culture and History Have Influenced Its AI Strategy

Epic’s founder, Judy Faulkner, started the company in 1979 with three employees in the basement of an off-campus apartment near the University of Wisconsin-Madison. The privately held company now employs 14,000 people and generates annual revenues of more than $5 billion.

Epic CEO Judy Faulkner

From the start, Faulkner’s theory of the case was that only by maintaining complete control of everything could Epic produce an integrated product that reliably served its users’ needs. Over the years, she has rebuffed myriad offers to buy the company, take it public, or accept Wall Street financing. Before AI, this meant that, to ensure seamless integration, the various components of the EHR (the clinician-facing system, the lab system, the billing system, and the patient portal (MyChart), which contains medical records on three out of every four Americans) had to be built by Epic, and Epic alone.

Interestingly, when generative AI came along, Epic departed from its usual “control everything” mantra. Epic decided not to build its own AI scribe, instead striking partnerships with two third-party vendors: a start-up named Abridge, and DAX/Nuance, owned by Microsoft. Epic even invested in Abridge, something unheard of in Epic’s long history of self-sufficiency. The integration with Epic gave Abridge and DAX an advantage over other AI scribe start-ups, as the companies were able to tout easier integration with the Epic EHR (a claim disputed by the other scribe start-ups, who saw the Epic partnerships as more marketing buzz than practical advantage).

But if anyone interpreted the AI scribe partnership as evidence of a new Epic, poised to play well with others, they would have been quickly disappointed. Over the next two years, Epic rolled out dozens of other AI tools, demonstrating to itself and the market that it could compete successfully with VC-funded start-ups. The company also released new pricing models that, in some cases, allowed clients to purchase a bundle of AI tools at a single price, rather than buying a la carte, undercutting many start-ups on cost.

Finally, in a move that made it clear the company’s control-everything DNA had not changed, in 2025 Epic announced it would release its own AI scribe tool. Soon after Epic’s announcement, Abridge CEO Shiv Rao tried to reassure jittery clients and investors, touting his company’s efforts to go well beyond scribing to develop an integrated platform of clinical AI tools. “There’s a big difference between a party trick and a platform,” he said. “But being able to translate that into a scalable enterprise-grade platform that can actually work on a daily basis, but also improve with continual feedback? That’s a totally different story.” His tone may have been measured, but it’s hard to imagine he wasn’t somewhere between disappointed and livid.

Can Anything Be Done to Thwart Epic’s Market Dominance?

In last week’s JAMA, my colleagues Julia Adler-Milstein, Sara Murray, and I discuss why Epic’s stranglehold on the healthcare system AI market is likely to stifle investment and innovation. We also enumerate policy maneuvers that could create a more level playing field.

The most important changes would be those that make it easier to integrate third-party tools into the EHR. Perhaps to get ahead of this critique, last month Epic announced the upcoming release of Agent Factory, a platform that allows customers to build, customize, and monitor various AI agents. Whether Agent Factory will facilitate the testing and integration of third-party AI tools (as opposed to homegrown tools developed by health systems) is unclear; if so, it would address one of our pain points, as summarized in this table from the JAMA article.

While these are steps that could be taken, the question is whether anything more should be done to dampen Epic’s potential AI dominance. It’s not an easy call. As Seth Joseph wrote in a 2024 profile of the company in Forbes, “A big question ultimately comes down to whether Epic’s conduct can be interpreted as fair play – simply part of the ‘rough and tumble of the marketplace’ – or exclusionary and anti-competitive in nature, pursued primarily for the purpose of advancing Epic’s control.”

Where Will All This End Up?

The healthcare marketplace is large enough to support multiple winners, but Epic’s dominance means that successful start-ups will need to convince purchasers not simply that they are better than the native Epic tool, but that they are better enough to be worth some additional risk – of integration, corporate failure, and potential hassles – and maybe some additional cost. They’ll also need to move well beyond single use cases to become the platform of choice across a wide variety of adjacent use cases, since no health system is going to tolerate implementing dozens of standalone tools.

I am rooting for high-profile successes among the start-ups (and their investors) since, in the end, a vibrant, competitive marketplace will be good for patients, clinicians, and the overall healthcare system. Yet, absent major policy interventions or successful antitrust actions, which seem unlikely given the current administration’s anti-regulatory stance, Epic seems poised to continue to win more than its share of battles.

As everyone who has benefited from the tight linkages between Gmail, Google Docs, Google Calendar, and Google Chrome knows, there are real advantages to living inside a single digital ecosystem. But healthcare is not productivity software, and “good enough” is a higher-stakes verdict here than in virtually any other human pursuit.

A note regarding conflicts: I advise several 3^rd party companies (Notable, Commure, and Arbiter) that may compete with Epic in the AI space. I’m also an Epic user in my clinical role at UCSF.