Aashima Gupta (Google Cloud): "Healthcare doesn't lack vision. It lacks courage."

[00:00:00] We believe agents will bridge that gap between the chore and the purpose. I believe it's not agent replacing clinicians, it's agent making it possible from having an episodic to continuous. Really? Where humans still provide the high touch, empathetic clinical judgment engagement, but that 24 by 7, there's not enough clinical or human capacity. So that 24 by 7, always on agent, I believe we all need that.

[00:00:30] Dear listeners, welcome to Faces of Digital Health, a podcast about digital health and how healthcare systems around the world adopt technology with me, Tjasa Zajc. Most conversations about agentic AI in healthcare get stuck on capability. The discussion you're about to hear focuses on the gap between capability and deployment and what closes it.

[00:01:00] At HIMSS Europe, I spoke with Ashima Gupta, Global Director of Healthcare Strategy and Solutions at Google Cloud, who argues that healthcare's bottleneck isn't vision, it's courage. The processes are documented poorly or not at all. AI fluency programs reach a fraction of employees who want them.

[00:01:22] And most enterprises are running agents without the harnesses, which refers to grounding, evaluation and red teaming, which is what production deployment actually requires. On the other hand, patients are embracing AI fully. So in this discussion, we discussed a little bit how Ashima sees the European versus the US market.

[00:01:48] We talked about how do we navigate and orchestrate agents so we don't run into the next level interoperability mess. What can agents do and what we should be mindful of when using them. So enjoy the show. And if you haven't yet, make sure to subscribe to the podcast wherever you listen to your shows. Check out the video versions on our YouTube channel.

[00:02:14] And also check out our special series on the use of AI from the patient perspective, the agentic patient. In that series, I speak with patients about concrete use of AI, which prompts they use, which tools, what works, what doesn't, what guardrails they use and more. So check out that as well.

[00:02:39] And also subscribe to our newsletter, which you can find at FODH.substack.com. That's FODH.substack.com. I will add all the links into show notes. And now let's see what Ashima has to say.

[00:03:14] Yeah, we're gonna talk a little bit about healthcare and Google Cloud and everything that you're doing. We're in Europe. How like how do you see this market compared to the US? We had a discussion at HIMSS Global and we talked a lot about the agentic AI. And the US is much more advanced in that sense, in terms of what the market already offers to clinicians and healthcare.

[00:03:43] So how, what kind of, what does the market in Europe mean to you? Do you see that the challenges are different? From Europe perspective, I'm seeing Europe taking much more cautious, responsible approach, including the data sovereignty. Data not leaving the country is super important. The other part is, I believe the AI regulation framework is slightly different, but the use cases are exactly the same. Healthcare, no matter where you go on the globe, we have common challenges.

[00:04:13] Workforce shortages, people feeling burnt out, access is a big issue. Now, with the, with the, Europe is much more kind of public health ecosystem. And the US is much more kind of private and you get into public health, but more on the Medicare and Medicaid side. So that, I really like about Europe, the public health system, the universal health access. But when you peel down the onion, the challenges remain the same.

[00:04:43] And I believe that's the moment across the globe we are all in to have AI and agentic AI specifically to help with the different challenges. We're seeing similar trends across the globe. I believe Europe is not an exception to that. We spoke a lot about the agentic AI and that's also something that Google Cloud is massively focused on.

[00:05:12] So can you talk a little bit about that? So how do you see the whole orchestration of AI agents in healthcare? How many agents do you use, by the way? I use the three big agents that I use for my personal use. And I'm seeing from the healthcare perspective, I go back, I'm seeing like five trends. And those are like agent for every employee.

[00:05:37] Like I use agents for myself, for my calendar management, for my email searching, for my shopping. But if you think of employees, we believe in the vision agent for every employee. What that means is no matter which function you're on, not just clinician or nurses, maybe in revenue cycle claims, you could be in case management. There's always a chore of the job and there's a purpose of the job.

[00:06:07] So we believe agents will bridge that gap between the chore and the purpose. So that's first trend we see. The second trend I see is agents for patients. We all have been there. Patients are trying to navigate healthcare in one of the most vulnerable moments. It could be their family member, it could be them themselves. And that today, when we talk about episodic, I believe it's not agent replacing clinicians.

[00:06:36] It's agent become making it possible from having an episodic to continuous. Really? Where humans still provide the high touch, empathetic clinical judgment engagement. But that 24 by 7, there's not enough clinical or human capacity. So that 24 by 7, always on Asian. I believe we all need that. I may have a dumb question about my health and I want to talk to my doctor about it. Exactly.

[00:07:02] I just actually started a special series called the agentic patient, where I talk to patients about how they use AI, which prompts very specifically, which tools, which prompts. Exactly because that those in between moments when you don't have the clinical personnel with you, but AI can be massively helpful. However, one thing that I'm kind of wondering here about is the whole trust issue.

[00:07:29] So, you know, when I think about creating agents for maybe just, you know, promotion and social media posts for them to do that instead of me, it's a straightforward use case. But like from the cybersecurity perspective, I don't think I really want to give passwords and everything required in order to run that to AI agents.

[00:07:52] And when you translate that into healthcare, how do you see that we can mitigate the whole challenge with the trust and like just something going wrong with the AI agents? Hannah Fry just did a great series called AI agents. Is this the best thing or the worst thing we've invented? I do believe it's going to be the best thing, but healthcare will and should move at the speed of trust.

[00:08:20] And we are seeing a lot of AI agents and a lot of AI hype. But there's a reality. There is this new level of engineering that is required to ensure that your agents are not hallucinating. They're connected to the enterprise ground truth. We call it grounding. And there are techniques like retrieval augmented generation and a few others where you are ensuring that agents are connecting to the data. The second is, what are your evaluation framework?

[00:08:50] How are you evaluating what agents are producing? Are they are of the standard? So I call it that evaluation framework are important. Where our customers tell us where they are bringing the clinician and the nurses to ensure that the responses the agents are seeing are the right ones. And you have to build those harnesses, those genetic harnesses.

[00:09:15] One of the big questions with AI is also the energy consumption. And at the same time, what the industry is kind of doing to tackle that is to create these basically small, large language models, very specific for use cases and the models that run on local devices.

[00:09:37] Can you talk a little bit about that just so we can understand a little bit more from someone who is following this on a daily basis on the front line? That's a great question. It is front and center for how we are developing from Google Cloud perspective, our AI hyper computer, what I call it. And it is important to understand the nuance. It is a full stack, meaning it starts from just not just the data center of the energies.

[00:10:06] It's the chips. We have our TPU, Ironwood. We then have our own model like Gemini. So it's you're building your model for the chip and chip, which is more efficient and performance to the model. And then, of course, we have from infrastructure to data center to the chip to the model and the innovations on top. And to me, that integrated stack is where you will see the performance improvement. We're already seeing it.

[00:10:34] And that layer, we will see much more advancements coming out in the future. To your point, not every use case requires a robust model. So how do you match your use cases to the model sizes? And that I go back and say this is a new level of kind of AI engineering. How is your CIO thinking about or your, you know, your infrastructure team? Are you involving them?

[00:11:01] Because you don't want to get surprised with even your own builds from the token perspective. So you need to be very judicious and you need to be very thoughtful on how you build that your AI stack. We have, as I mentioned, AI hyper computer from Google Cloud. So we, it's in our DNA, we've been building TPUs not today.

[00:11:23] It's been like more than a decade when we saw the TPUs because we were seeing the large models require the type of compute, type of efficiencies. That's what we are building. And I'll give you another example. A few years ago, DeepMind actually looked into reducing our electricity bill and data center using AI. And we were able to reduce by 40%.

[00:11:46] So we need to also talk about how can we build new level of efficiency in even reducing our data center footprint, our data center electricity usage. This is an example from Google. So we will see as we move forward innovations like this. Energy is top of mind for everyone. And we will continue to see that. Yeah.

[00:12:09] I'm glad you mentioned the tokens and, you know, the use when you're building stuff because that's one of, that's another concern that I have when I think about AI agents. You know, you create an agent, you give it a lot of tasks to do, you go to bed and in the morning you realize that, you know, you're going to get a huge bill for all the tokens that the agents use.

[00:12:36] So I guess the question here is how do you build trust among clinicians, among the teams that are building these things on larger scales to prevent mishaps like that happening? Yeah, there's no building an agent and going to the bed. I would recommend that. I think that that's where, again, very thoughtful execution that is required. A, agents have agency.

[00:13:03] They are able to execute certain steps, but in certain cases you don't want agent to have the agency. So that combination of agency with very rule-based deterministic AI and bringing that together where you want agent to have the agency and where you don't want them to have agency. And to me, that's the construction of the agentic architecture of combination of those two.

[00:13:26] And I believe where we've seen the progress, I think we don't talk about that in healthcare, which I hope that we do need to about AI fluency and literacy across the board. People are going to be using this. How are they evaluating? How are they using in their day-to-day? You ask me what agents do you use? If you ask in the enterprise setting, how many agents is your average employee using? The hands won't go up.

[00:13:56] So to me, that AI fluency, first you need to understand what you're going to use before you actually scale it or deploy it. And that needs to happen not just at the CEO or the board level. It needs to happen at the employee level. And where I see traction in my role in the customers, they are investing in the future workforce that understands its nuances.

[00:14:20] They know how to harden the infrastructure so that mishaps don't happen like this. But it's not like spray and pray approach, build agents, like average enterprise CIO. If you allow, there's hundreds of agents. How are you going to manage it? We, from cloud perspective, are Gemini enterprise. It's a single pane of glass. You really control how many agents you've launched, the auditability, what actions they took. You want to know if an agent worked on my behalf. What is the auditability?

[00:14:50] All that is a platform play. And this is where we pride ourselves in providing that through Gemini enterprise to be able to have single pane of glass in a given enterprise, hundreds of agents. Because what you don't want is each department having their own agent, their own way of looking into it. Then you will have a agentic proliferation, which is very hard to manage. And you don't want that.

[00:15:17] So that's where a platform approach, a single pane of glass approach, your platform components, auditability, transparency, grounding, evaluation, that is common. No matter if you're building an agent in one side of the enterprise versus the other. And to me, that holistic approach towards the architecture is what differentiates from people who are laggard or worse is who are actually deploying that in production.

[00:15:46] Mm-hmm. When you talk about digital literacy, how do you see that clients are doing that? Do they invest in the existing workforce? Like, do they've got courses that they just oblige their employees to take? You know, technology is developing so rapidly that, for example, a friend of mine recently said,

[00:16:14] with the model today, I do something in an hour that I did in a month, three months ago. So, like, the question is, should I even bother building something today if I know that it's going to take me much less time to do that in two months or something? So it's just, it's, people are already kind of burnt out by the speed of change and the demand also on productivity.

[00:16:41] So what kind of approaches in terms of digital literacy do you see clients are taking in healthcare? First of all, there's a gap. So we did a survey, I believe, earlier in January where there's 84% or some employees want it and only 24% have the enterprise level AI fluency and literacy programs that are available to them.

[00:17:05] So that's a systemic gap I see. And this gap is going to be profound, especially in healthcare. You're talking about patient care. You're talking about trust. First, where I've seen success is training and fluency always goes with kind of experimentation. You need to be in the arena to be able to see where guardrails are needed.

[00:17:28] I'll give you an example. We work with Highmark. They had, I believe, 1 million prompts that they've created. Like, there's a shared knowledge. It's not just all by myself. So where do you create the culture where people are sharing prompts for common workflows? You think of any workflow in healthcare. There's a, these are interdependent, brittle, there's a handoff, there's dependencies. And then when you want the agent to do the task, how many have actually even documented that process?

[00:17:58] So to me, it's not just AI literacy. It's a process literacy. Do you understand the process? Because some of these systems have been existing for decades. Yeah. And the process just exists. And again, where I see progress where people are actually documenting, breaking down steps and tasks, that this is what this process is like, 27 steps. And then I can bring an agent and then say what part I will do to determine as take, which step agents will be allowed to have the agency.

[00:18:29] I don't believe most enterprises have done their process mapping yet. And to me, that's also part of education. That's also part of, it's not just AI for AI sake. It's really re-imagining. You're re-imagining a process where redundant work can be taken care of by this tool that I have. But you can be consumed by it. It's not AI for AI sake.

[00:18:54] It's AI to re-imagine a process where you are bringing that joy back into our day-to-day job. Like in my job, I'm sure I have a lot of tools. So can we bring that? And to me, that level of literacy, that level of training program needs to work. And then sharing. Sharing that one million prompt example from Highmark, but also sharing within healthcare. These processes are not unique.

[00:19:22] Maybe there's new ones for different, but I don't see the level of culture of share here. I have documented these processes now. What do you see? And to me, we will all, we will lift all boats if that knowledge is shared widely, that knowledge is open for public to consume. And then you apply technology. You can't be running after a shiny object. And I'm from tech and I'm seeing that.

[00:19:51] Technology is incredible. We are seeing a tremendous amount of progress. But again, it needs to be like re-imagination. What we do. Do you think we lack visionaries in general? Because like basically what you're talking about is the visionary mindset and the same problem

[00:20:13] that we had at the beginning of digital transformation, when, where bad IT solutions would just digitize, not digitalize, not rethink the processes. And like with AI, I guess it seems that's kind of on steroids, but like you really need to have a vision. And as people, we don't like change. So tough luck.

[00:20:36] I would say vision, but also what we lack sometimes in healthcare is courage. Because this is our regulated system and it's a risk averse industry and rightfully so. There's a patient at the end of it, back in my, if I drop on my Kaiser experience, we used to call them life critical systems. But that should not be an excuse to not be courageous to re-imagine a process.

[00:21:04] Like it requires that sponsorship, that championship to enable our vision. Like I will be bold enough, will have courage because it will yield to better outcomes for my patient, better day-to-day work-life balance for my workforce. It will get the patient, maybe get the prior authorization faster. So it's a very mission-driven industry and proud to be part of it.

[00:21:33] I believe we had more courage to lean in. And I also see the other side. Like we have in the past, we built complexity with each innovation, EHR, automation here in this workflow or a point solution there. Each step added complexity, added more screens, more clicks, more fragmentation. Now what's different is AI within the flow of work.

[00:22:02] I think that's where agents are so incredibly useful. They are within the flow of work. You can call and lean upon that agent. They can reason, they can act. These are not a piece of software sitting on the side. They can adapt. They can reason. They can act. And to me, that's the moment we are all in. I believe this is a time where if we have the courage to reimagine our processes, take a hard look and say, where's the efficiency gains?

[00:22:31] Can you quantify that? And then launching it. Yeah. Experimenting with it because you will make mistakes. But doing with the right guardrails, the right evaluation before you scale, I think we are all, I do see next three to five years industry leaning in. What do you expect to see in three to five years? Maybe also just from the lens of the impact that patients might have in their care journey.

[00:22:59] I, you know, everything you said, I kept thinking, yeah, but like clinicians are risk averse for a reason because they're the ones that are going to be liable if any wrong decisions are made. However, patients on the other hand, it feels like that's at least my impression when I'm talking to them, that there's a whole, like, it really is a different world with AI and how much you can learn about the care that you can expect or have ideas

[00:23:27] or be a informed collaborator with the clinician. So since you said three to five years, do you have any expectations? You are a visionary. I don't have a crystal ball, but I would say we will see a lot of agentic in the back office, in the operations. A, the bar is lower than patient safety when you're going directly to patients. And these are long running processes. We see a lot of work.

[00:23:55] We just announced with OptumReal as an example. We are working with HCA for nurse handoffs. So whenever there's a workflow and there are board leaders in the industry who are reimagining, we are seeing traction. I'll continue to see that. On the patient side, if you think this is what you just said, even from the patient's perspective, navigation is hard. So patients are not ready, not as a healthcare, on the diagnostic side,

[00:24:25] but help me navigate my appointment. Tell me what's my co-pay? Who's the specialist? Who takes my insurance? What's my... Do they speak my language? All that, when you are in a vulnerable moment, a patient navigating your own health, or a caregiver navigating health for someone you love, all these are burden are on the patient. So to me, the new experience, which is going to be real in three to five years,

[00:24:54] where agents feel that always on, maybe boring questions for healthcare nurses and clinicians to keep answering. But very important to patient, can they take their workload and then leave the clinical judgment at the high touch intelligence, care through clinician and nurses? Do they do a lot of this together? Can I take... So to me, that experience will see come to life.

[00:25:23] The always on component by agent, the high touch clinical judgment by clinician and nurses, and reimagining that. Case in point, CVS has launched Health 100. They work with us in Google Cloud. And they are looking into reimagining that. Can I take the friction away? You all have different insurance companies. You know, in the US alone, the clock speed, the three different clock speeds, your insurance changes every year.

[00:25:52] The insurance roster of which physicians, which hospitals they enroll, kind of changes every year. The Medicare prices keep changing. So now all this, but my health is in a different clock speed. We don't run year by year. So to me, that's the opportunity for AI agents to help bridge that gap that health doesn't happen the one year I'm enrolled with one insurance versus another. To me, it's my journey. And today, that's where the fragmentation happens

[00:26:21] and leaving a lot of these tasks for patients to figure out. Now you figure out. Even things like that. Can an agent do that? Or explain my diagnosis to me like a 10 year old. Mm-hmm. Tell me, should I be worried? Explain to my mom in a way that, in her language. So there are things like that. We will see where it will make it easier. You know, the best thing we can do with this technology is empower the patients

[00:26:51] and empower the providers. If we solve these two in three to five years, where they feel empowered, we won't have fully solved it. But empowerment from clinician nurses meaning I'm doing less paperwork. Empowerment for patients is I'm not doing this entire navigation. That's the trick. Can we empower these two for better health outcomes for all? Do you have a favorite prompt for health, for a large language model?

[00:27:20] I often, I, from the prompt perspective, I often find myself taking my test results. And then my prompt is, can you compare my journey from the past five years? How's my different biomarkers are progressing? So I do leverage that to see how my health at the metabolic level, which I would not ask my doctor for this 10 minutes. So that's a prompt I use. Yeah.

[00:27:49] Just one more question. I mentioned earlier, you know, how fast the models are developing and also the way that we manage hallucinations is changing. Like some of the advice, for example, is to ask AI, give me an assessment of how convinced you are in the quality of your suggestions.

[00:28:15] And then it can tell you how sure it is for each suggestions. Even if you just say, don't hallucinate, that kind of helps. But, you know, from the tech perspective, how do you see that the challenge of hallucinations is improving? And where are we currently with that? Oh, absolutely. From where we started to where we are, I would say hallucinations have come down tremendously.

[00:28:44] To your point about the architecture, you know, we have recently published our, my colleagues in DeepMind, AI Harness called AI Co-Clination. It is based on one is a reflective agent. Other is reading the medical guidelines. Third is checking the work. So that agent harness, where multiple agents work together to check each other's work, reflect, and then act.

[00:29:09] I think that agent harness, if you can apply that to multiple use cases, similar to your point. Like one agent is doing one, the other is checking the work. And I believe we will continue to see this agent harnesses. This is our work where we publish that. But you can imagine that. I'm seeing that even in security. So we, this agent in defense. So there's like three agents, red, green, and blue.

[00:29:37] Red is for red teaming, adversarial testing. It knows how to attack. The blue agent is about defending. How do I triage? And how do I detect? The green is, okay, we have detected, but how do I make sure that this continuous, this goes on and I make it like real. So those three agents, like this is our work that we have done.

[00:30:01] So you'll see that this combination of this agent harnesses kind of come to life to solve for that. Mm-hmm. You just gave me an idea because one of the people that I spoke with who did, who found an insurance plan for him with the help of AI, one of his advice was to, anytime he talks to AI, to also tell it to do the red team analysis.

[00:30:25] So now I guess I need to also ask and try what happens if you say do the blue team and green team. The blue team and the green team. I think that's where, and I believe that's such an important point. I believe no AI should go to production without red teaming, without adversarial testing. And I see leaders like HCA, like Mayo, they already have these practices, but they are not uniformly applicable.

[00:30:55] That's why we see a lot of pilots. That's the, I call it the O-gap, going from pilot to production. And this security testing, red teaming is going to be critical. But yeah, check out the red and blue and green. I will. So Ashima, thank you so much for this short discussion. And I look forward to seeing what Google Cloud is going to do next. Thank you for having me. Bye.

[00:31:20] You've been listening to Faces of Digital Health, a proud member of the Health Podcast Network. If you enjoyed the show, do leave a rating or a review, wherever you get your podcast, subscribe to the show or follow us on LinkedIn. Additionally, check out our newsletter. You can find it at fodh.substack.com. That's fodh.substack.com. Stay tuned.