Post 3 of 4 in The Narrowing EMR series.
KEY TAKEAWAYS
- 85% of healthcare leaders were exploring or implementing generative AI as of Q4 2024, up from 72% earlier that year — a 13-point jump in roughly three quarters.
- 31.5% of US nonfederal hospitals deployed generative AI inside their EHR by end of 2024; another 24.7% planned deployment within 12 months — with no federal subsidy or regulatory mandate.
- EHR adoption under HITECH took a decade and tens of billions in federal incentive payments to move from 9% (2008) to 84% (2015). Generative AI in hospitals covered comparable ground in 24 months with no subsidies. These are not the same reference class.
- Healthcare moves slowly at the application-of-record layer (EHR replacement: 5-7 year capital project) and fast at the workflow-adjacent layer (ambient AI scribes: from pilot to meaningful-scale deployment in 3 years). The distinction matters for strategic planning.
- The pattern across SAP, Salesforce, ServiceNow, and Microsoft is consistent: vertically integrated vendors have kept the operational application of record while formally opening the data, analytics, and AI layer through partnerships with model-agnostic platforms.
- The Apple counter-argument does not apply to Epic: Apple’s customers have no regulatory obligation to interoperate. Health systems do. CMS–0057-F, TEFCA, and HHS-OIG information blocking enforcement all push healthcare toward more standardized, cross-organizational data exchange. CMS–0057-F specifically requires FHIR APIs for covered payer use cases beginning in 2027.
The strongest objection to the argument I’ve made over the last two posts is one I hear often, and I want to take it seriously. It goes something like this. The pattern you’re describing, where incumbent application vendors lose the data and analytics layer to open-platform competitors, may hold in generic enterprise tech, but healthcare is different. Healthcare is slow. Give this twenty-five years, not five. The HITECH Act took the better part of a decade and tens of billions of federal dollars to drag hospitals from negligible EHR adoption to ubiquity. Why would AI agents and data platforms move any faster?
That is not a strawman. The history is real, and the instinct is grounded in a generation of operational experience. But I think the 2024-2026 data tells a different story, and the relevant question for health system leadership today is not whether the pattern arrives. It is how late it arrives.
The 2024-2026 Data Says Healthcare AI Is Not the Right Reference Class
Start with what we can actually measure. McKinsey’s Q4 2024 healthcare survey found that 85% of healthcare leaders, across payers, health systems, and healthcare services and technology organizations, were exploring or already implementing generative AI. That was up from 72% earlier in 2024. A thirteen-point jump in roughly three quarters. JAMA Network Open’s December 2025 study, which I cited in Post 1, found that across 2,174 nonfederal US hospitals, 31.5% had deployed generative AI inside their EHR by the end of 2024. Another 24.7% were planning deployment within twelve months. Roughly 56% of US hospitals deployed or actively planning, in 2024, with no federal subsidy and no regulatory mandate driving them.
Now compare against the historical reference class. ONC data shows basic EHR adoption in nonfederal acute care hospitals rose from 9% in 2008 to 84% by 2015, with certified EHR possession reaching 96% by 2017. That trajectory required the HITECH Act of 2009 and the meaningful-use incentive program, with tens of billions in federal payments distributed over roughly a decade. Generative AI inside the EHR went from effectively zero to roughly one-third of US hospitals in twenty-four months, with no subsidy, no regulatory floor, and no meaningful-use program pulling adoption forward.
EHR adoption is the right reference class for “healthcare is slow.” AI adoption inside the EHR is not. Healthcare is moving on this faster than it has moved on anything in fifteen years.
Why Healthcare Moves Fast on AI and Slow on Infrastructure: Two Different Categories
The “healthcare is slow” framing collapses two different categories that move at very different speeds. Let me pull them apart.
Healthcare is genuinely slow on infrastructure rearchitecture. Replacing a core EHR is a five-to-seven year capital project. Migrating a major clinical workflow to a new platform takes years. Swapping out the underlying clinical data store of a thirty-hospital system is a generational decision. On those, the historical reputation is earned.
Healthcare moves much faster on workflow-adjacent automation that has visible ROI and does not require ripping out the spine. Ambient AI scribes are a clean example of how quickly healthcare can move when the technology augments an existing workflow rather than replacing the system of record. Three years ago, Abridge, Nuance DAX, and Suki were vendor pilots. Today they are deployed at meaningful scale across the provider market, moving from pilot to production faster than core EHR transitions of comparable scope have historically run.
The pattern is consistent. When the change is application-replacement, healthcare is slow. When the change is application-adjacent, augmenting an existing workflow with measurable yield, healthcare absorbs as fast as any industry. The real insight is that healthcare is slow at exactly the layer where the EHR vendors operate (the application of record), and fast at the layers where this series’ central question actually lives: data, analytics, AI agents, and governance.
A Directional Pattern from Adjacent Enterprise Categories: SAP, Salesforce, ServiceNow, Microsoft
The same pattern has played out across adjacent enterprise categories over the last fifteen years. I’m not claiming this is a law of history. But in every major category where an incumbent application vendor has tried to hold customers inside a vertically integrated stack, the data and AI layer has gradually migrated toward open, model-agnostic platforms. The incumbents have not collapsed. What they have done is open the data layer through partnerships, acknowledging the gravity of where the analytics and AI work is headed.
Oracle EBS and SAP held the operational ERP application but have moved toward an open data-platform posture. SAP itself formalized that posture through zero-copy data sharing with Snowflake, following an earlier partnership with Databricks inside SAP Business Data Cloud. Those announcements signal that SAP customers want to run a substantial share of their analytics and AI workloads on platforms that sit alongside SAP rather than entirely within it. The ERP is still where the transactions live; the analytics and AI plane increasingly extends into adjacent platforms.
Salesforce ran a similar play. Salesforce formalized bidirectional zero-copy data sharing with Snowflake and Databricks rather than fighting them, which says something about where the strategic gravity is and where Salesforce sees its customers wanting to run their analytics and AI work.
ServiceNow is on the same arc. ServiceNow’s internal analytics remain useful for ITSM reporting, and ServiceNow announced its own zero-copy partnership with Snowflake in October 2024 and a similar zero-copy integration with Databricks shortly after, both on the premise that customers want to combine ServiceNow data with broader enterprise data outside the platform.
Microsoft is the most instructive case, and the lesson worth pulling is broader than Microsoft Fabric specifically. In September 2025, Microsoft announced expanded model choice in Microsoft 365 Copilot, bringing Anthropic’s Claude models alongside OpenAI inside Copilot. That is a signal: enterprise customers want optionality, not single-model lock-in. Even the credible vertically integrated player is making explicit that a closed AI stack is not where the enterprise customer wants to land.
The pattern across these four cases runs consistently in one direction. The vertically integrated vendor has kept the operational application of record. The vendor has formally opened the data, analytics, and AI layer through partnerships with model-agnostic platforms. None of these incumbents has been displaced. All of them are now coexisting with open platforms in their customers’ stacks.
Epic is in a position that rhymes with where Oracle EBS was in 2010. The healthcare market has not run this experiment yet, and there is no guarantee it runs the same way. The point is not that healthcare will replay the ERP story step-for-step. It is that the same structural pressures are beginning to appear here.
Why the Apple Counter-Argument Does Not Apply to Healthcare
Anyone who has thought carefully about vertical integration will reach for Apple as the counterexample. Apple holds a vertically integrated stack against open-platform competitors and has done so successfully for two decades. Why doesn’t that apply to Epic?
Two reasons it doesn’t. Apple is a consumer business. The buyer is choosing a personal device based on aesthetics, ergonomics, and ecosystem convenience. Enterprise IT decisions get made by committees that prioritize interoperability, vendor leverage, and total cost of ownership. The strategic logic is different.
More importantly, Apple’s customers have no regulatory obligation to interoperate with anyone else’s stack. Health systems do. CMS-0057-F, TEFCA, and the HHS-OIG information blocking enforcement posture I described in Post 1 all push in the same direction. Clinical data is increasingly expected, and in some use cases required, to flow across organizational boundaries through standardized exchange mechanisms. That is the opposite of an Apple-style closed stack. The Apple analogy fails not on Epic’s product strategy, but on the operating environment Epic operates in. Apple wins by closing. Healthcare cannot close.
The Semantic Layer Objection: Why Epic’s Clinical Logic Is a Real Moat
The second objection is more substantive, and it deserves a direct response. It runs roughly like this. Even granting that the data and AI plane moves to an open platform, Epic has thirty-plus years of encoded clinical logic, configuration, decision support rules, and workflow customization that lives inside the application tier. That semantic layer cannot be reconstructed elsewhere. The competitive moat is not Chronicles or Hyperdrive. It is the accumulated configuration that makes the product clinically usable.
That’s real, and I’m not dismissing it. Epic’s clinical logic represents a substantial body of institutional knowledge refined over decades of customer deployments. Reproducing it from scratch is not an option on any near-term timeline.
I want to be explicit that what follows is forecast, not current-state fact: this is a timeline problem, not a capability problem. My read is that AI-assisted business-rule extraction and code comprehension are improving quickly, though enterprise-grade results on legacy application configurations are still uneven and the work is harder than the demo cases suggest. Within a five-to-ten year window, AI-assisted extraction of Epic-equivalent semantic layers into a database-native form is plausible, in my view. Aggressive, but buildable. That is a forecast, not a settled outcome.
The practical implication for procurement decisions is straightforward. A health system making a 2026 decision cannot reasonably bet on the semantic layer gap closing by go-live. A health system making a 2028 or 2030 decision can. At Abundant, what we see in client conversations is that the most sophisticated organizations are already watching for proof-of-concept work at organizations like Mayo, Kaiser, and Intermountain. If those proofs land in the next two to four years, my expectation is that the middle of the market will follow. If they don’t, the timeline of this argument extends accordingly. This series is a late-decade argument, not a near-term procurement argument. That distinction matters.
What’s next in the series
If the pattern holds, and it will not hold overnight, the interesting question is what the EMR application tier actually becomes. Not whether it disappears. It does not. In Post 4 I’ll close out the series with three things: three caveats stated honestly, because the argument needs them to be taken seriously; the bifurcation I think actually plays out, where Epic’s role narrows but does not collapse; and three questions a CIO, CMIO, or CTO should be asking right now to position their organization for the late-decade environment, regardless of whether they end up running Epic, Oracle, or both.
Continue Reading
This article is part of our four-part series.
➡️ Part 1: The Safe Path Has Inverted: Why an EHR-First Strategy Became Riskier in 2026
Frequently Asked Questions
Healthcare is not uniformly slow at technology adoption – it is slow at specific types of change and fast at others. Healthcare is genuinely slow at infrastructure rearchitecture (replacing a core EHR is a 5-7 year capital project) and fast at workflow-adjacent automation with visible ROI (ambient AI scribes went from pilot to meaningful-scale deployment in roughly 3 years). The ‘healthcare is slow’ framing also reflects the legacy of HITECH: EHR adoption required a decade and tens of billions in federal incentive payments. Generative AI inside the EHR covered comparable penetration in 24 months with no subsidy. Healthcare’s speed depends entirely on what is changing and at which layer of the stack.
An ambient AI scribe is an AI tool that listens to clinical conversations and automatically generates clinical documentation, reducing the documentation burden on physicians and clinicians. Leading products include Nuance DAX, Abridge, and Suki. As of 2025, ambient AI scribes have moved from vendor pilots to meaningful-scale deployment across the US provider market in roughly three years – a deployment pace faster than core EHR transitions of comparable scope. Northwell Health deployed Abridge across 28 hospitals. Several major academic medical centers and regional health systems have followed. Adoption is continuing to expand as reimbursement models and workflow integration mature.
The biggest barriers to AI adoption in healthcare are governance and data access – not technology capability. As Post 2 in this series documents, most digital governance committees stall on AI tool approvals not because tools lack merit, but because no one has articulated who has decision authority over clinical data access and what controls apply. Secondary barriers include regulatory uncertainty (FDA oversight of AI as a medical device), integration complexity with legacy EHR systems, clinician workflow disruption concerns, and liability questions around AI-assisted clinical decisions. Technology cost and availability are no longer the primary constraints for most health systems.
Epic’s primary competitive moat in 2026 is not its technology architecture – it is the accumulated clinical logic, workflow configuration, and decision support rules refined across thirty-plus years of customer deployments. This semantic layer represents decades of institutional knowledge that cannot be reproduced from scratch on any near-term timeline. Secondary moats include Epic’s training infrastructure, user group culture (KLAS, UGM), research community (Cosmos), and deep integration with payer and referral partner workflows. Epic also holds the governance layer for first-party AI agents and the deepest workflow paths in its deployed base. The argument of this series is not that Epic’s moat disappears – it is that the strategic center of gravity is shifting from the application tier (where the moat is strongest) to the data and AI governance layer (where it is weakest).
The EHR industry in 2026 rhymes structurally with where the ERP market was around 2010–2015. SAP and Oracle held the operational application of record but gradually opened their data and analytics layers through partnerships with model-agnostic platforms such as Snowflake and Databricks. Salesforce and ServiceNow followed similar zero-copy and data-sharing patterns. In each case, the incumbent was not displaced; it coexists with open platforms in its customers’ stacks. The key difference for healthcare is the regulatory layer. CMS–0057-F requires certain FHIR APIs for impacted payers, while TEFCA and HHS-OIG information-blocking enforcement push the market toward broader, more standardized exchange of clinical data. That interoperability pressure accelerates the opening of the data layer in a way that did not exist for ERP or CRM vendors.
The Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 authorized approximately $27 billion in Medicare and Medicaid incentive payments to accelerate EHR adoption in US hospitals and physician practices, through the Meaningful Use program. It took roughly a decade and this federal investment to move basic EHR adoption from 9% of nonfederal acute care hospitals (2008) to 84% (2015), with certified EHR possession reaching 96% by 2017. HITECH matters for understanding AI adoption because it establishes the reference class for ‘healthcare is slow’: a regulatory mandate plus tens of billions in incentives driving a decade of adoption. Generative AI inside the EHR reached roughly 31.5% of US hospitals in 24 months with no equivalent mandate or subsidy. These are not the same phenomenon.
About the Author
Ryan Kent is the founder of Abundant Healthcare Strategies, a healthcare IT advisory firm that helps health systems navigate strategic IT decisions, digital transformation, and AI governance. With over 10 years in healthcare IT consulting, Ryan works with CIOs, CMIOs, and CTOs at health systems navigating the transition from EHR-centric IT strategy to organization-controlled data and AI architecture. The Narrowing EMR series reflects his ongoing advisory work with health systems planning their 2026–2030 technology strategy.