Health systems are racing to keep up with their EHR vendor’s quarterly cadence. They should be asking whether that roadmap leads where clinical AI actually needs to go.
A recent Becker’s piece on EHR upgrade strategy reads like a dispatch from a parallel universe, one where billion-dollar Epic implementations, single-digit downtime records, and Gold Star designations are treated as the front line of AI-driven healthcare. They aren’t. They’re health systems optimizing their position on a treadmill someone else controls.
Here’s what’s actually happening. EHR vendors have convinced health systems that the right response to the AI era is to become better customers of their vendor’s AI roadmap. Take 75% of new enhancements. Run quarterly upgrades. Earn your Gold Star. The implicit message: the path to AI leadership runs through your EHR vendor’s product cycle. It doesn’t.
When CIOs frame success as adopting 75% of new enhancements, the question nobody asks is, which 75%? Whose clinical priorities shaped those features? Whose commercial interests determined the roadmap, and who profits when the algorithm is wrong? When digital officers celebrate generative AI embedded in the platform, it’s worth asking who embedded it, what it was trained on, who is monitoring its behavior now that it’s deployed, and whether it was ever calibrated to a patient population that resembles ours.
The AI that matters in healthcare isn’t the AI that comes pre-packaged in your next upgrade. It’s the AI that your institution’s clinicians validated, trained on your patient population, audited for bias, and can defend to a plaintiff’s attorney when something goes wrong. Those are not the same products.
The sovereignty argument isn’t theoretical. In our cardiac surgery program, the decisions that matter happen in milliseconds: intraoperative anatomical mapping during robotic AF ablation, perfusion dynamics during ECMO, real-time arrhythmia burden inference from implantable loop recorders, post-operative trajectory prediction from continuous physiological signals. None of that runs through an EHR upgrade cycle. It runs on local compute, on validated pipelines, on algorithms the clinicians using them helped build. The people writing the inference logic and the people standing at the operating table are, often, the same people. That is the only configuration in which anyone in the room can actually be held accountable for what the algorithm recommends.
None of this is a knock on Epic. Epic is an excellent administrative and documentation platform. The mistake is treating it as an AI strategy. Health systems that conflate EHR optimisation with AI leadership are building on borrowed architecture, and they’ll discover the limitation at exactly the wrong moment.
So what does a real AI-era infrastructure posture look like?
First, separate the administrative layer from the intelligence layer. Your EHR handles documentation, orders, and billing. Your AI infrastructure handles inference, prediction and decision support. These have different governance requirements, different upgrade cadences and different accountability frameworks. Conflating them because they share a vendor isn’t a strategy. It’s an abdication.
Second, own your data and watch it. The moment your institution’s clinical data trains a vendor’s model, you’ve handed over the most valuable asset in your AI future. Sovereign data infrastructure isn’t a luxury for academic medical centers. It’s the foundation of institutional AI that is defensible, auditable and improvable. And ownership is only the start. Someone in the institution has to be monitoring that data continuously, because algorithms drift, patient populations shift, and a tool validated on last year’s cohort isn’t automatically valid on this year’s. Set and forget is not a clinical AI strategy. It’s a malpractice case waiting to happen.
Third, close the gap between the people who build the AI and the people who use it. The most dangerous architecture in clinical AI is the one where the engineers who created the algorithm never watch it run on a real patient, and the clinicians running it on real patients never saw how it was built. That handoff is where understanding evaporates. The people building these tools should also be the people deploying them, and the clinicians using them need to actually understand them. Not at a marketing-deck level. At a level where they can articulate what the algorithm knows, what it doesn’t, what it was trained on, and where it will fail. Clinicians who can’t answer those questions don’t have an AI program. They have a vendor relationship and a liability exposure.
The cautious voice in the Becker’s piece deserves more credit than it received. The IT leader who pointed out that there has to be a reason to introduce change, that you can’t just roll in new code on a quarterly basis when it doesn’t solve a real clinical problem, is the most clinically honest voice in the entire article. Her peers framed that position as conservative. I’d call it the only one that would survive peer review.
The implicit argument in the Becker’s piece is that the health systems adopting the most features fastest are winning. In product management, that’s plausible. In clinical AI, it’s backwards. The system that adopts fewer, better-validated AI tools, with genuine clinical governance and clinician-builders behind each one, will produce better outcomes than the one with a 90% feature adoption rate and a Gold Star.
A vendor’s quarterly cadence is not a clinical AI strategy. Health systems that mistake one for the other will discover the difference when their algorithm makes a recommendation that nobody can explain, on a patient nobody anticipated, in a situation nobody planned for. By then, the roadmap won’t save them, because nobody in the room will understand what the algorithm is doing, or why.
The treadmill isn’t the answer. Get off it and build something you own, understand and can defend.
Dr. Khalpey is chief medical AI officer at Atari AI, chair of applied clinical AI at the Atari AI Foundation and director of Khalpey AI Lab.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.