Two-thirds of hospitalists are already using artificial intelligence in clinical practice—but they're doing it largely on their own, without the training, governance, or organizational support that could determine whether the technology actually improves care. A new perspective published in the Journal of Medical Internet Research by researchers at the University of Colorado School of Medicine sounds an urgent alarm: adoption rates alone tell us nothing about whether AI will make medicine better or simply busier for the doctors leading this charge.
Doctors have adopted AI tools at breathtaking speed, with most gravitating toward large language model platforms that help generate differential diagnoses and suggest management options. The pull makes intuitive sense—these tools can accelerate clinical thinking and unlock possibilities in a patient conversation. Yet as medical institutions have learned painfully before, the speed of adoption can easily outpace the thoughtfulness of implementation. When electronic health records rolled out across hospitals in the 2000s and 2010s, they brought genuine safety improvements and standardization. They also brought something else: clinician burnout, cognitive overload, and—in some cases—worse patient outcomes than the paper systems they replaced. The difference between success and failure had almost nothing to do with the technology's inherent capabilities and everything to do with how it was rolled out.
Hospitalists at the University of Colorado and elsewhere are now facing a critical crossroads. The authors—Dr. Anna Maw, Dr. Aakriti Pandita, and Marisha Burden—argue that AI's ultimate impact will be determined not by use rates, but by implementation quality and fit. Poorly integrated tools increase clinician workload and burnout despite their intended benefits. Early evidence on diagnostic AI powered by large language models reinforces this risk: when integration, training, and workflow design are inadequate, the systems that clinicians rely on to support decision-making can actually make clinical reasoning worse.
The picture is not hopeless—but it demands action now, before AI becomes as entrenched and as poorly supported as EHRs once were. The researchers identify three urgent priorities. First, clinicians need genuine, structured training on how to prompt AI systems and—crucially—how to interpret what comes back. Second, health systems need to apply implementation science frameworks, the proven methodologies that helped organizations learn from EHR rollouts. These frameworks assess workflow integration, training infrastructure, and unintended consequences, guiding teams to adapt their strategies based on real conditions. Third, hospitals must build continuous monitoring systems using routine clinical and workflow data, allowing them to track whether AI actually delivers on its promises of better outcomes, greater equity, reduced costs, and improved clinician and patient experience.
The authors offer a measured but firm conclusion: AI adoption in hospital medicine is likely inevitable. But inevitability is not destiny. Whether these tools advance medicine—whether they truly optimize patient outcomes, health equity, costs, and clinician wellbeing—hinges entirely on the quality of the work that happens after the first AI tool gets downloaded. The time to build that infrastructure is not in five years, when problems have already accumulated across hundreds of hospitals. It is now, while two-thirds of hospitalists are still finding their footing with these powerful and unpredictable tools.