Few interventions have altered the daily practice of medicine as thoroughly as the digitization of the medical record. The landscape of care is now mediated by “screens”. For most clinicians, the dominant interaction of the workday is no longer with a patient, but with a keyboard — and that shift arrived with enormous cost, substantial unintended consequences, and remarkably little prospective evaluation.

The electronic medical record (EMR) is not only a cautionary tale; it is the closest thing we have to a natural experiment when a transformative digital technology is introduced into health care by policy mandate, designed largely outside of clinical medicine, and paid for by the institutions compelled to adopt it. As artificial intelligence (AI) is now promoted with a nearly identical rhetorical package — it will end burnout, solve the workforce shortage, and cure disease — the EMR record deserves careful analysis.

A Mandate Assembled Without A Pilot Study

The Health Information Technology for Economic and Clinical Health (HITECH) Act, enacted in 2009 as part of the American Recovery and Reinvestment Act, created the Medicare and Medicaid EHR Incentive Programs and the concept of "meaningful use." Adoption was initially framed as “voluntary” and incentive-driven. In practice it was not: Stage 2 criteria and 2014 technology requirements arrived in 2014, and Medicare payment adjustments for eligible professionals who failed to demonstrate meaningful use began in 2015. Soon thereafter, non participation carried a payment penalty, a program of incentives became a de facto mandate.

On the narrow metric of adoption, the policy succeeded completely. Hospital use of certified EHRs rose from under 10% in 2008 to 99.4% by 2024 , with the disparities once seen across bed size, ownership, and geography effectively eliminated. The relevant question is not whether adoption occurred. It is what else the U.S. purchased along the way.

What the legislation did not include is as instructive as what it did. There was no requirement for prospective pilot testing in clinical settings, no usability or cognitive-load standard for certification, no mechanism to protect purchasers when a certified vendor failed, and comparatively little clinician input relative to the technology sector's lobbying presence. The specification that emerged optimized for structured data capture, billing integrity, and regulatory attestation. These are not the same objectives as clinical care, and the products reflect this discrepancy.

Record-Keeping Became A Permanent Cost Center

Before 2009, clinical documentation was not a meaningful line item on most balance sheets. It now represents a recurring capital and operating expense. Cloud-based EHR platforms commonly run in the range of $200 to $700 per provider per month , with functionality that most practices cannot operate without — telehealth modules, e-prescribing, laboratory interfaces — frequently priced as add-ons at an additional $30 to $100 per provider per month. Enterprise implementations at large systems are budgeted in the millions to tens of millions. Information technology departments have expanded simultaneously, adding personnel who deliver no direct patient care.

Health system dollars are exchangeable. Capital committed to licenses, servers, interface fees, and implementation consultants is capital not committed to imaging, laboratory capacity, surgical technology, facility maintenance, or salaries. In an already strained system, this is not a rounding error, and smaller facilities cannot absorb it.

The magnitude is visible even at institutions with substantial reserves. Memorial Sloan Kettering Cancer Center reported a fiscal-year 2025 deficiency of operating revenues over expenses of $47.9 million , which it attributed to planned one-time investments tied to its February 1, 2025 Epic go-live alongside higher personnel, pharmaceutical, and supply costs. Roughly $177 million of the institution's operating expense growth was associated with Epic go-live support and training; adjusting for those costs, operating expenses grew 6.6% rather than 8.0%. Sloan Kettering had also posted a $ 113.2 million operating loss in the first half of 2025 and experienced the temporary decline in patient activity in February and March that characteristically follows a large go-live.

If a transition of that caliber produces an operating loss at an institution with $8.5 billion in annual operating revenue, philanthropic depth, and a national referral base, the arithmetic facing a rural critical-access hospital or an independent five-physician practice is considerably less forgiving. Industry estimates place the cost of switching platforms at $50,000 to more than $500,000 for a practice, with six to eighteen months of operational disruption.

Vendor Attrition And The Stranded-Asset Problem

A mandated purchase in an unstable market transfers risk to the purchaser. The health IT vendor market has consolidated sharply since the incentive programs began. Industry analyses describe a contraction from more than 1,200 EHR vendors in 2014 to roughly 280 by 2024. The federal government's final hospital adoption data brief documented that three developers now supply more than 80% of U.S. hospitals , with a single vendor growing from approximately 8% of the hospital market in 2010 to nearly half by 2024.

Maintaining federal certification is expensive, and the regulatory floor keeps rising — most recently with requirements to support USCDI v3 through FHIR US Core profiles . Those costs fall hardest on small developers, and market exits follow. When a vendor is acquired, sunset, or wound down, its customers face an unplanned migration: data conversion, revenue-cycle disruption, and full retraining, often while still servicing debt on the original system. Practices and hospitals that borrowed to comply with a federal mandate can find themselves paying twice within a decade for the same regulatory obligation.

Concentration carries a second, less discussed cost. With the majority of U.S. care delivery running on one of two platforms, the sector has assembled a single-point-of-failure risk profile for cybersecurity , in which one successful intrusion could compromise documentation, billing, and clinical operations across thousands of sites simultaneously. State attorneys general and federal antitrust authorities have begun examining the market structure that produced it.

Consolidation, Closure and Access

No registry counts the practices that closed because of health IT costs, and no honest analysis can attribute closures to EMRs alone. Reimbursement pressure, staffing costs, and payer administrative burden are all in the mix. But the directional effect is not seriously disputed: high fixed technology costs, steep implementation demands, ongoing administrative load, and the risk that a vendor may disappear a few years after purchase all push independent physicians toward two exits — early retirement, or sale to a larger network. Large networks are concentrated in urban markets. Each acquisition or closure in a rural county lengthens the drive to care.

This is the part of the ledger that was never modeled in 2009. Mandating an expensive, recurring, unfunded technology obligation across an entire sector produced predictable second-order effects: reduced access, fewer independent practices, longer waits, less patient choice, and emergency department crowding as the release valve. A policy sold as improving care delivery has, in aggregate, reduced the number of places where care can be delivered.

The Clinician Cost: Burnout By Design

The workload data are unusually consistent. Direct observation and time-motion analysis across four specialties found that for every hour of direct clinical face time, physicians spent approximately two additional hours on EHR and desk work during the clinic day, plus one to two hours of personal time on documentation each night. EHR event-log analysis validated against direct observation found that clinicians spent 355 minutes — 5.9 hours — of an 11.4-hour workday inside the EHR: 4.5 hours during clinic hours and 1.4 hours after. Clerical and administrative tasks accounted for 44.2% of that time and inbox management for another 23.7%. Clerical burden and computerized order entry have been independently associated with higher burnout and lower professional satisfaction .

The physician “inbox” has since become its own crisis. A national cross-sectional analysis of Epic Cosmos data spanning 2,067 hospitals, approximately 47,100 clinicians, and roughly 139 million patients found that patient-authored portal messages rose from 0.99 to 2.5 per patient per year between 2020 and 2025 — a 153% increase — while messaging intensity among patients who sent any message rose 146%, from 2.2 to 5.4 messages per year. Clinician- and staff-authored messages rose 24%. Critically, office visits increased 17% over the same period: asynchronous messaging is additive, not substitutive. Nearly none of this additional work is scheduled, staffed, or reimbursed.

Secure clinical chat has produced a parallel problem. A question that once required a thirty-second verbal exchange now generates a threaded message directed at multiple recipients, several of whom may have no role in that patient’s care. Inpatient clinicians routinely field dozens to hundreds of such inquiries per shift. Each is an interruption, and interruption during clinical reasoning is a well-characterized contributor to clinical errors. The cumulative effect is an "always-on" expectation that no profession sustains indefinitely.

The Patient Cost: The Rise Of The “iPatient”

Ultimately, patients absorb the residual “cost.” Mandated and redundant data entry — the same finding captured three times in three fields for three different compliance purposes — displaces face-to-face attention and degrades clinical encounters. Nearly two decades ago, Dr. Abraham Verghese described the emergence of the "iPatient": the data construct that receives the team's attention while the actual patient waits in the bed. The EMR institutionalized that construct.

Something specific was also lost in nursing and interdisciplinary documentation. The subjective, objective, assessment, and plan (SOAP) narrative was a compact, transferable account of what actually happened to a patient over a shift, readable by anyone on the team. Structured data systems do not reward free text, and the narrative was largely replaced by checkbox templates and copy-forward blocks that satisfy an auditor and inform no one. Note length has grown while note content has thinned. Physicians, advanced practice clinicians, nurses, and therapists uniformly describe the patient encounter — not data entry — as the reason they entered the field of medicine.

Now Comes AI, With The Same Procurement Playbook

The case being made for AI in medicine is structurally the same case that was made for the EMR: it will relieve administrative burden, offset workforce attrition, extend access, and ultimately transform diagnosis and treatment. Some of this is plausible. Ambient documentation tools are among the first health IT products in twenty years that clinicians have adopted enthusiastically rather than under duress, and early data on documentation time are encouraging.

But the EMR did not fail because digitization was the wrong idea. It failed because of how it was deployed: mandated adoption, costs borne by the adopter, requirements written by payers and vendors rather than clinicians, no pilot phase, no usability standard, no exit strategy, and no plan for what happens when the market consolidates. Every one of those conditions is currently being reproduced.

The infrastructure preconditions are also not in place. Meaningful clinical AI at scale assumes reliable high-bandwidth connectivity that large parts of rural America still lack; data center implementation and projected capacity faces significant opposition, including in New York. Electricity generation and transmission capacity on a grid that is aging will take decades to modernize. Security and governance questions also remain unresolved: agentic systems capable of acting with limited human supervision introduce problematic interfaces that health care has not previously modeled—and models hosted offshore raise data sovereignty questions that no institution should answer casually on behalf of its patients. Regulatory frameworks for continuously updating clinical models remain poorly defined and developed.

What A Prudent Introduction Would Require

Health care needs technology. It does not need another decade of the 2014 procurement model. Six principles follow directly from the EMR experience:

1. Pilot before mandate. Prospective evaluation in real clinical environments, with workload, safety, and outcome endpoints — not vendor case studies.

2. If it is mandated, it is funded. A federal requirement should carry federal financing, with explicit provision for small, independent, and rural practices. This is the single clearest lesson of HITECH.

3. Clinicians write the requirements. Usability, cognitive load, and documentation burden belong in certification criteria, not in post-marketing complaints.

4. Portability and a survivable exit. Mandatory data export in a usable format, and migration standards that make vendor failure recoverable rather than ruinous.

5. Measure the burden as a quality metric. Documentation time, inbox volume, and after-hours work should be tracked and reported the way infection rates are.

6. Address market structure now. Antitrust and procurement scrutiny should precede consolidation of the AI layer, not follow it by a decade.

The EMR delivered precisely what it was asked to deliver: a digitized record and near-universal adoption. It also raised the cost of practicing medicine, contributed materially to clinician burnout, degraded the clinical narrative, accelerated the disappearance of independent practice, and produced a vendor market concentrated enough to attract antitrust attention. It enriched the technology sector and created an administrative apparatus that no facility can now function without.

None of that was inevitable. It was a direct result of the way of the technology was introduced. AI is arriving with far greater capability, far more capital, and considerably less regulatory clarity than the EMR offered. The lesson of 2014 is not that medicine should resist the tool. It is that medicine should refuse the terms.

Dr. Peter Papadakos , Professor of Anesthesiology and Critical Care at University of Rochester Medical Center, is a contributor to this article.