LAS VEGAS — Healthcare has spent the better part of three decades laying the groundwork for AI adoption — shifting from paper records to a deeply digitized ecosystem, embracing digital health and weathering waves of technological innovation.
Still, even with that foundation, it hasn’t meant the industry has figured out how to use AI at scale, according to Ritu Agarwal, the Wm. Polk Carey distinguished professor of information systems and health at the Johns Hopkins Carey Business School.
Scaling AI means organizations would embed the tech into clinical and administrative workflows throughout an organization — think standardizing tools across multiple hospitals, unifying governance strategies and integrating AI workflows into electronic health records systems. Right now, many healthcare organizations are embracing AI but struggling to move past small pilot cohorts.
“If you look at all of the deficiencies that have been documented in healthcare practice and delivery from poor quality to high costs to limited access to inequities in healthcare delivery, I think AI has the potential to address all of that,” she said. “But there’s a lot of heavy lifting and hard work that needs to be done before we can realize the potential.”
A lot of this dynamic can be attributed to the fact that AI in healthcare is still in its infancy, she said. There are unresolved questions about clinical workflow integration, governance challenges, physician trust and return on investment that will take time and research to answer.
Healthcare Dive sat down with Agarwal at the HLTH 2026 conference to talk through what's standing in the way of broader AI adoption, how hospital executives should think about measuring AI ROI and where she expects the industry to be heading in the future.
Editor’s note: This interview has been edited for clarity and brevity.
HEALTHCARE DIVE: What are some of the biggest gaps between how a model performs in the pilot stage versus actually getting it into a health system?
RITU AGARWAL: I don’t think we have made that transition from pilot to true operational deployment. And that's because healthcare is complex, human beings are complex and medical conditions are complex.
When we develop AI models, we always face the challenge of having data that is robust enough to be able to guarantee or at least give us some assurance of adequate model performance under all conditions. And it's the ‘under all conditions’ that becomes kind of tricky.
A model that’s trained with data from Stanford Medical School is probably using a very different patient population at its underlying foundation than a rural hospital in Kentucky. So we have this concept of model drift, which happens when you take a model that was built in one context and try to apply it in another.
The short answer is that it will take a lot of additional development, testing and more representative data sets to be able to have models that we feel comfortable deploying in everyday clinical practice.
Many organizations are still a ways away from enterprise-wide adoption. I imagine clinician trust of AI plays a role in accelerating that. What factors determine whether a new AI tool gets embraced by clinicians versus something that dies at the pilot stage?
We haven’t quite cracked where and how AI should be integrated within the clinical workflow. So with every other technology, you have to think about what is the process redesign that you need to do. You don't want to bolt on AI to a bad process or an inefficient process. At the same time, you don’t want to take away the clinician’s discretion and autonomy.
So the really tricky part is figuring out how and where in the clinical decision making process AI should play a role. When should it recommend? When should it provide information? When should it synthesize? When should it question the doctor’s decision? I think physicians will resist AI if it’s just bolted onto their processes.
Also, physician involvement in the ideation and conceptualization and development of AI applications is important. I see a lot of vendors who have a [chief medical officer] and they have a handful of clinicians as part of their development team, but they, in my opinion, are letting the technologists call the shots. So you really need the clinicians to be not just front and center, but perhaps leading the development of AI applications.
Let’s touch on ROI. If a hospital executive asked you to actually measure whether AI investment is “worth it,” what metrics would you tell them to track?
We have done outcome evaluation studies with healthcare technologies for a while, but we are in the incipient startup phase of AI, so we don't have the outcome data just yet. But what we have to put in place is a very rigorous study design, or evaluation process where we actually measure baseline outcomes.
If I were to advise a health system leader right now, I would say put a measurement baseline system into place and clearly identify what outcomes you are expecting from a given AI — whether it’s efficiency in terms of time spent on certain activities, or effectiveness in terms of number of patients served or actual health outcomes.
And don’t try to boil the ocean all at once. Pick specific disease conditions for which you want to evaluate AI, and then go about it systematically. You’re not going to see bottom line or top line benefits right away, but the emerging evidence seems to suggest that you will eventually.
In other industries we see that happening. So there's no reason in healthcare we can't. I think the upside is potentially large. We have to wait and see if it's realized.
How do you see the health AI vendor market evolving as AI advances? We’re seeing a lot of big tech companies get into healthcare and releasing industry-specific tools. What will happen to the companies offering highly specialized point solutions?
I think there's going to be a lot of consolidation in the industry. The winners are the ones who have the most extensive set of relationships with the health systems and the most extensive sources of data.
So Google and Microsoft have leadership in the market just by virtue of size and the number of customers they have. So if they can construct these massive health data sets, they come out ahead. All the electronic health record companies have these massive data sets that they’ve curated over time, but they’re less nimble and flexible than the tech companies in terms of building out innovative solutions and technologies.
So I think we’ll see competition, we’ll see some consolidation, but we’ll see the winners will likely get bigger and the ones that are coming out with point solutions are going to get absorbed, bought out. This old business of platformization where you have these large platforms, I think that’s going to be a pretty powerful business model in the future for health tech.