Inside the Walled Garden: What AI Unlocks in the EHR

Almost ten years ago, I got into a heated discussion with a VC who told me he would never invest in a founder trying to build a new EHR. The discussion occurred just after the NYTimes had published a controversial piece on Epic's burgeoning empire and "Willy Wonka"-like campus. The VC made a compelling argument that investing in any company trying to build an EHR was a doomsday scenario. That large-scale hospital systems would never rip and replace after spending so much on implementing it in the first place. Of course, there have been verticalized counter-scenarios such as Flatiron, which was acquired in 2018. Yet, on net, the VC might have been right - the top EHR companies are still gaining market share.

For a sector that holds on to legacy solutions from the 1990s, the rapid adoption of AI in healthcare at 2x the rate of the broader economy is consequential. Not because we are going to see EHRs be ripped and replaced tomorrow. But because there are a number of companies developing agents that report back to the system of record in a way that was extremely difficult to do, until now.

‍

I. How Epic became the dominant

EHR

Headquartered in a Wisconsin suburb, Epic was founded by Judy Faulkner in 1979. More than its location, the company is an outlier from the typical Silicon Valley profile - Epic has shunned venture capital (Judy owns 42%), clients have to be invited to buy the technology, the company has never made an acquisition, and Faulkner, who remains CEO at 82 years old, has kept a relatively low profile. Over the years, many have criticized Epic as the unlikely beneficiary of the $30Bn spent by the government to bring medical records online, starting with GW Bush's executive order mandating that most Americans have interoperable EHRs within 10 years, and Congress's HITECH Act of 2009, which provided the financial windfall for healthcare providers to adopt EHRs.

Epic was not listed among the top 20 hospital EHR vendors until 2003, when it signed on with Kaiser Permanente and its 8M members. The company was selected over IBM and Cerner. The cost to implement Epic at Kaiser: $4Bn. At the time, Epic had 400 employees and 73 clients. By 2017, Epic had 8,100 employees, 315 clients, and $2Bn in revenue. Today, Epic has become the de facto system of record. The company has 14,000 employees, runs in 3,500 hospitals (including over 600 large healthcare organizations like Memorial Sloan Kettering and the Mayo Clinic), and generates $6Bn in annual revenue. Epic maintains records for ~90% of all patients in the United States, and controls 67% of acute multispecialty hospital beds. More than 190M Americans manage their own care via MyChart, the patient-facing app.

Most publications have largely criticized Epic as a walled garden, and the reason why patients can't easily access their health records when changing hospitals. The truth is that Epic, along with Cerner and other large enterprise solutions, has built something that healthcare decision-makers want: an all-encompassing product that does almost everything, everywhere (from prior auth to claims management to medication and laboratory ordering). It requires a unified database and a closely managed system that allows patients to be seen everywhere - from the oncology center to the NICU clinic to the gastroenterologist's office. Epic's system covers most of these critical boxes.

Moreover, Epic services its clients' incentives. Hospital systems care about connecting with the patient, but only to a certain extent. For example, scheduling appointments and messaging. But not so much when revenue walks out the door. Take NYU, for example, which has spent 12 years building out its healthcare IT as a core competitive advantage. Their stated identity is "one patient, one chart." A central data warehouse where all mission data flows. If you go to NYU for trauma surgery, it would be more beneficial if you get your MRI done in their clinic too. Hospital systems do not want patients to shop around for the cheapest MRI, or CT scan. These outpatient services have the highest profit margins in healthcare and effectively subsidize lower-margin services such as emergency departments and intensive care units.

So how does Epic help its clients? Its current system creates friction for patients who want to shop around. Small health systems cannot buy Epic directly, but they can "sublet" the software from a local, large hospital system that already uses it. It's touted as cost-effective because the small practices avoid the multi-million-dollar independent setup fee. But in reality, a clinic that chooses to stay outside of a major health system's medical orbit may find itself faxing papers in order to exchange information with that hospital.

Recent regulatory changes, such as the Hospital Price Transparency Rule, have been enacted to protect the patient and give them options. Hospitals and insurance companies are now forced to publish their negotiated cash and insurance rates, and a crop of startups has emerged over the last 5 years to help patients find alternatives. A few are listed below:

Of course, all of this still leaves the problem of behavioral friction. Providers and administrators write referrals that automatically point inward to their own facilities via Epic workflows. A patient would have to specifically request a referral or go to an independent facility to bypass this process.

II. Digital Tech 1.0: build around EHRs instead of within them

Several well-known companies have bypassed the massive walled garden and built outside EHRs instead over the last decade. Large health systems were already locked into all-in-one EHRs - to the tune of billions of dollars - and were fairly satisfied with a convenient option for managing their workflows, particularly the billable events. While administration time ballooned, costs skyrocketed, and doctor burnout reached unprecedented levels, most argued that bureaucracy was the root cause, not the EHR.

Left with few alternatives, the most successful technology companies in this era were those that turned to the new buyer: employers. Faced with low-interest-rate environments and skyrocketing healthcare costs, employers sought to retain employees, reduce costs, and gain greater visibility into their employee population. They decided to bypass insurance carriers and self-insure. By selling to employers and reporting strong ROI, companies like Spring Health ($3Bn valuation), Hinge Health ($5Bn), Sword Health ($4Bn), Maven Clinic ($2Bn), and Livongo (acquired for $18.5Bn) emerged as multi-billion-dollar companies.

The problem, outlined on this substack a few years ago, is that employers began to experience vendor fatigue (which has since led to the rise of Transcarent, Accolade, and Quantum Health). And, two parallel systems were built:‍

Two parallel systems: inside vs.

outside the system of record

The EHR holds billable encounters and the unified record; wearables, employer-benefits, chronic-disease data, and social determinants sit outside.

‍

III. What AI unlocks: core workflows

on top of the system of record

(healthcare administration agents)

For the past 20 years, "the source of truth" has lived in the EHR, in the billable event. True, parallel systems are gathering more data, but the revenue flows through the EHR. Healthcare Information Exchanges ("HIEs") emerged as a way to fix interoperability. However, government grants for the initiative eventually ran out, information was blocked due to the aforementioned competitive dynamics (42% of hospitals reported perceiving information blocking by other providers in 2021), and most importantly, the US never established a national patient ID largely due to privacy concerns in Congress. This meant HIEs had to do probabilistic matching across systems -- first name, last name, date of birth, address -- and often got it wrong.

With AI, companies can now use ML to predict probabilistic matches by identifying name permutations and address variations-for instance, recognizing that "Maria Perla, DOB 1/5/75" and "Maria E. Perla, DOB 01/05/1975" are the same individual. This creates a reliable patient identity across systems while maintaining privacy. Datavant, for example, uses machine learning to connect more than 60 million healthcare records across thousands of healthcare organizations.

That's why the rapid adoption of scribes has been one of the most consequential shifts in healthcare since the foundation of the internet (although the ultimate ROI is still up for debate). AI now sits at the source of truth, leaving the EHR primed for safe, efficient access. While major health systems remain entrenched in the "innovators' dilemma," where unlocking data might threaten in-network revenue, the rest of the market is ripe for an entire ecosystem of what we'll call Healthcare Administration Agents ("HAAs"). These agents will not displace EHRs per se at first. Instead, they'll focus on high-ROI administrative workflows that require writing data back into the EHR for billable events.

By managing core workflows that cause administrative burnout, these agents earn the right to expand their influence within the system of record. The EHR continues to hold the actual billable events, such as prescriptions or lab orders, but the identity of who is writing those records shifts. As an example, at Dria, we recently invested in a company called Psyrin, which leverages a proprietary voice AI tool to shorten the behavioral health clinic intake process from 2 hours to 30 minutes. Behavioral health workflows are quite specialized, with forms such as PHQ-9 or GAD-7 required. For serious mental illness, the consequence of getting the intake wrong is extremely consequential, requiring hyper-specialized training data. And the intake is a billable event with a set CPT code.

The same framework can be applied to the ortho specialization of Assort Health, which recently raised $120M to scale its agents.

This framework of 1. Specialized; 2. Consequential; and 3. Billable (i.e., written back into the EHR) is one that can signal an HAA with a compounding moat. In Psyrin's case, the more intakes they complete, the more data they collect, and the more integral they become in identifying and assisting with the treatment of high-risk patients. In Assort's, the more acute injuries they triage, the more integral they become at managing the audit trail of imaging, consults, and post-op visits.

The relationship between Epic and AI-driven health startups is analogous to that of established LLM providers and many YC-batches. If you build it, Epic will build it too. That is why the HAAs that can nail the right to become a core workflow will be among the most differentiated and highly scaled companies we have seen in healthcare. They can transform incumbent EHRs into what the original Healthcare IT Act promised - APIs that builders can access. Founders are already navigating this transition, which remains highly friction-intensive at the moment. One founder said it will take 4-6 months to develop the ability to write to an EHR with which they have already signed a definitive agreement. The technology is there, yet the incentives at the top are misaligned. On the lower end, the market pull for medium-sized, specialty, and independent health systems to adopt HAAs continues to grow. Practices can stay with their current EHR / system of record. The core workflows will just become defined, agentic, and simplified.

The timeline for this across each part of the Healthcare Administrative Agentic spectrum seems not too far ahead.

‍