AI transcript
0:00:15 Hello, and welcome to the NVIDIA AI podcast. I’m your host, Noah Kravitz.
0:00:20 Agentec AI is reshaping the pharmaceutical landscape from streamlining clinical trials
0:00:26 to enhancing patient engagement. Global healthcare intelligence company IQVIA processes data from
0:00:33 over 1 billion non-identified patient records across more than 100 countries, making our guests
0:00:39 uniquely positioned to discuss how intelligent automation can transform healthcare outcomes at
0:00:45 scale. Raja Shankar serves as vice president of machine learning at IQVIA, where he spearheads
0:00:50 the application of artificial intelligence to transform research and development workflows in
0:00:55 the life sciences industry. His expertise lies in developing AI solutions that accelerate clinical
0:01:01 research and drug development processes. Avinar Broy is vice president of commercial analytics
0:01:06 solutions at IQVIA, focusing on how AI can revolutionize pharmaceutical commercialization
0:01:12 strategies. He brings extensive experience in leveraging advanced analytics and machine learning
0:01:18 to optimize brand outreach and market access in healthcare. Gentlemen, welcome to the NVIDIA AI
0:01:21 podcast, and thank you so much for taking the time to join.
0:01:23 Thank you for having us. Thanks for having us.
0:01:27 Absolutely. I really appreciate it. I know you’re calling in from different time zones,
0:01:32 had to kind of move some things to get this to work. So appreciate you guys rolling with us to get the
0:01:38 podcast out. Let’s start with the basics. And maybe Avinar, you can start and just tell us a little bit
0:01:44 about what IQVIA is for listeners who might not know, and then a bit about your own role. And then,
0:01:46 Raja, maybe you can talk about your role as well.
0:01:52 Definitely, Noah. It’s a pleasure to be here. So IQVIA is a leader in using data, tech, and analytics
0:01:58 to accelerate innovation from clinical to commercial to drive life sciences pipeline, right? So if you think
0:02:04 about this, the main objective is how do we bring the drugs that are crucial for patients faster,
0:02:10 quicker, and when they need it, right? And IQVIA is the engine that is working with every life sciences
0:02:14 company and healthcare organization to make that dream come true, right? With the power of the data
0:02:20 and analytics, which is now the way to approach this problem statement. In my role, I’ve primarily
0:02:25 focused on the commercial business, which is basically once the drug is approved, now you have to actually
0:02:31 take this drug and make it available for the patients. So in that, you need to understand where
0:02:35 are your patients? How do you reach them? What’s the messaging? How do you make sure that every
0:02:39 patient in every direction get those drugs at the right time for their disease landscape? And my job is
0:02:44 to bring the tech and analytics together with the power of the data to make that happen.
0:02:46 And Raja, on your side of the house?
0:02:54 So I’m the counterpart to Avina more on the R&D side. So just as he mentioned, we start early from
0:02:58 a clinical development and a clinical trial perspective. We run clinical trials to get the
0:03:07 drug to market and then Avina takes over. In my role, what I’m leading is all of our AI and agentic AI
0:03:13 program to transform how clinical development is done and how clinical trials are run. So how can we
0:03:20 improve the quality of clinical operations and clinical trials? And how do we speed the time it takes for a
0:03:23 drug to get to market so that patients can have it quicker?
0:03:30 – Healthcare is one of the industries that is being most impacted and advanced and kind of evolved
0:03:35 by AI, machine learning, all the things we talk about. We’ve talked a lot about agentic AI this year
0:03:42 in general. Could one of you maybe, before we dig in, just kind of give an overview of what agentic AI
0:03:50 means in IQVIA and in the healthcare landscape, as opposed to other forms of AI, machine learning,
0:03:56 that have been in play up until recently. – I mean, IQVIA is like many organizations. We have been
0:04:02 doing machine learning for a long time, maybe more than about 15 years or so, right? And we have built
0:04:08 machine learning models or supervised machine learning models to do diagnosis, prognosis, patient finding,
0:04:16 and so on. Then of course with chat GPT, generative AI came in and everybody is now aware of the chat GPT
0:04:23 function, how we can generate documents, summarize, have a conversation, and so on. The key thing with
0:04:29 agentic AI is because you can interact with the AI in natural language and it can output a natural language,
0:04:36 and it can also call upon tools, look at data, and do different tasks. This allows us to build more
0:04:42 complex systems with multiple agents where different agents do different tasks and they talk to each
0:04:47 other. The output of an agent can be the input to another agent. And specifically within the context of
0:04:55 R and D in clinical trials, we have very complex workflow. Today, they are very manually driven
0:05:02 workflows. And what we are able to do this with agents is look at entire processes and say which of these
0:05:08 aspects of these processes could be done by an agent and could be done faster by an agent. It could be things
0:05:16 like document generation. It could be things like how can we track what is happening at a site level
0:05:22 and ensure that it is ready to enroll patients faster or everything is going fine at a site level. We can
0:05:27 talk a little bit more in detail about the agents in the future, but essentially take complex workflows
0:05:34 and have agents execute them faster and better. Right. Yeah. I mean, very simplistic and very nicely
0:05:40 articulated. Right. So the way I picture this one is the ML model when we were doing the statistical
0:05:45 model, AI ML, it’s like showing an image of the car and asking, is it a car? Then we went into a
0:05:50 generative AI where it says, show me a car. And it basically shows you an image of the car by predicting
0:05:57 it. And now you’re asking, drive me to the work and the car takes you. There’s the next era of it. Right. So I think
0:06:04 every aspects of the workflows within life sciences is trying to be identified. Right. Wherever there’s a
0:06:09 task that’s repetitive with human driven, for example, you are about to launch a product and
0:06:13 you’re trying to understand the disease landscape, patient burden, access control, so that you can
0:06:17 launch your product with the right success. Right. So that it can reach the patient. Think about
0:06:22 commercial operations. You build a strategy. You have a great plan. You know who to reach out,
0:06:27 but operational breaks down. Right. How do you actually reach out to the patient? Who are the HCPs that you
0:06:33 need to call upon? How do you pass on your message to your KOLs? And it happens today so rigorously with a lot of
0:06:39 manual efforts, a lot of Excel, SQL agents coming in and being a companion, helping the commercial ops to
0:06:43 get there faster. Right. And then end of the day, the engagements. Right. How do you make every
0:06:49 engagement count and make impactful? So across the value chain from the commercial space, agents are going
0:06:54 to make every process squeeze the time. And at the end, you’re getting the drugs to the right patient at
0:06:58 the right time. Right. So that’s where we see agentic. Yeah. What’s the current state of
0:07:05 agentic AI adoption across pharma? Is it, are we super early stages? Are we a little ahead of
0:07:11 other industries? What’s the state like? I think my, my point of view on this one is we are breaking that
0:07:18 pilot barrier and try to enter into a full-blown adoption and scale. But with that being said,
0:07:22 you know, there’s a lot of stats out there that it’s not going very well. You know,
0:07:27 the pilots are roaring success, but when it comes to large scale adoption, large scale
0:07:33 operationalization, companies are struggling. Right. And like any change, it’s a tough thing,
0:07:38 right? But here, one of the greatest things that has happened is this technology has been so easy to
0:07:42 adopt and so reachable. And it’s, it’s really, that’s why we all are talking about this. Right.
0:07:48 Sure. But the flip part of it is, it is doing what human does. Right. And try to be part of your
0:07:53 equation. So you have a new member coming and joining your house and you know, you need to adjust with
0:07:58 them. Right. It will take strategy. It will take an organization. So it will take the whole village to
0:08:03 come together to, to bring the new digital persona into your home. I, I, as you guys were talking,
0:08:07 I was thinking about how I actually need to make a doctor’s appointment. And I know this isn’t quite
0:08:12 life sciences, but I was thinking, man, I need an agent that can like make the appointment for me
0:08:16 and get me to the office. That’s what I need on my end. Right. Yeah. And then there’s the other side
0:08:20 of the equation, right. Which is the compliance governance regulations, right. To your point.
0:08:24 Yeah. How do you make sure that, that it becomes, it suggests the right doctor with the right
0:08:28 specialty to treat your specific one. Right. Yeah. This is not one of the area where you take
0:08:33 few shots and one of them hits at your success. Right. You need to make sure every shot counts. Right.
0:08:38 Right. But those, I think that the barriers that, that yet to be uncovered in a full potential
0:08:45 realization. Raja, on the R and D side, where do you see the, the biggest opportunity for agentic AI
0:08:51 to enhance life science workflows? I think there are two big types of transformation that are likely to
0:08:58 happen with AI and agentic AI. When I look at AI, it’s also this kind of generative models that have a lot of
0:09:04 power. And these two opportunities are going to be clinical development simulation, which means,
0:09:10 can we simulate what is going to happen in clinical development and clinical trials before we spend
0:09:17 tens or hundreds of millions of dollars on it so that we are more confident of success of how we design and
0:09:23 execute the trials, both from a trial design perspective in terms of efficacy, safety, inclusion,
0:09:29 exclusion criteria, as well as a trial strategy perspective into which countries and which sites
0:09:35 do we do trials. That’s one thing. Can we increase the chance of success? And then the second big
0:09:41 opportunity for me is clinical trial automation with agents. So every clinical trial has a lot of
0:09:47 processes. There are hundreds and hundreds of processes and they’re very manual and time consuming and
0:09:56 repetitive and sometimes could be mind-numbingly boring things that people have to do, like look at
0:10:03 documents and insert that in another file after looking there. So this is completely ripe for
0:10:09 agentification. And if you’re able to do that, we could reduce the time of clinical trials, which is a big win
0:10:16 to sponsors because every six months they get to market faster than the NPV, extra NPVs, and the hundreds of
0:10:18 millions of dollars for them.
0:10:23 And Avanab, on your side, on the commercial side, where do you see the biggest opportunities
0:10:25 for wins with the Gentic workflows?
0:10:31 I think it’s the same chain with the links, right? So if Raja can make the trials go faster and
0:10:37 the product comes out quicker, our journey on the commercial side starts from there, right? How do we make
0:10:43 sure that works reach the right patients, right? The one thing on the commercial side is commercial
0:10:48 side is very rich on data, all right? Fundamentally, if you think about this, behavioral data, consumer
0:10:53 data, execution data, clinical data, and you can keep on naming them, right? The challenge is too much
0:10:59 data leads to chaos, right? We have data all over the places. People are trying to stitch this data
0:11:04 together before they can strike the insight and your launch is torn because one single decision on those
0:11:09 data points can lead to a multi-million dollar or even billions of dollars lost with your launch,
0:11:15 right? So with all the launches that are planned over the next, I would say, five years across all
0:11:21 formal life sciences are very, very critical, right? So the way I see agentic ecosystem and AI ecosystem
0:11:25 evolving is on the three parts of the equation, right, that I mentioned earlier. One, as the launch
0:11:31 happens, understanding where to launch, how to launch, and how do you make your strategy impactful,
0:11:35 right? How do you create baseline forecast and then reiterate them very, very quickly to create
0:11:40 scenarios? How do you identify where is the unmet need of your patients? How do you understand what
0:11:46 is the payers are looking for in your differentiation messaging of your drug, right? So building the upfront
0:11:51 strategy, can agent bring this all structured, unstructured, heterogeneous data at a global scale
0:11:55 and enable you to give that insights when you need it? The second part of the equation is,
0:11:59 good, you have a strategy, you have to operationalize it. How do you create your, you know,
0:12:03 your standard call plan, segmentation, territory design, right? Who are the HCPs that you need to,
0:12:09 you know, enable to have those conversations, right? So your operational plan. And then the last is your
0:12:13 engagement. Good that you have a strategy, you made a plan, now you’re operationalizing it, but,
0:12:18 you know, at the end of the day, your HCPs need to be aware of your drug, your patient need to be
0:12:23 onboarded, they need to be compliant, adherent, right? So who are those people? They are your front-facing,
0:12:28 you know, communication. So how do you make your marketing channels much more automated? How do you
0:12:33 get mud merge, HCP-driven engagement, which are faster, quicker, right? So when you combine all these
0:12:37 three parts of the strategy, we see agent-to-KI transforming every aspect of insight generation,
0:12:40 task automations, and workflow execution.
0:12:47 To your first point, Avinab, about the data, we recently did an episode with another health tech
0:12:53 company, healthcare AI company called Cytoreason. And they were talking about the guest Shai was talking
0:12:59 about what he called the data insight gap, and talking about how each year, I mean, each moment,
0:13:08 really, right? There’s so much more new data being produced and generated that the chances of finding
0:13:14 insight with kind of the ratio, you know, of like overall data to insight keeps growing. And that
0:13:20 those insights are, if not harder to find, they’re just buried under more and more data. And I always
0:13:25 find it interesting because there’s talk kind of in the broader community, or there has been talk this
0:13:31 year about, you know, AI running out of data to train, and you talk about these giant foundational
0:13:36 models and where are we going to get more data and synthetic data and all of that. In the healthcare
0:13:41 space, is there, I don’t want to say too much data, but is the problem sort of inverse that, and especially
0:13:49 with a company like IQVIA, you process so much data every day globally. Is the challenge in that sense
0:13:55 finding ways to keep up with the data and find insights? Or how do you look at sort of the current
0:14:02 state of all the healthcare data available and how to turn that into, you know, actionable insights,
0:14:03 to use that term?
0:14:08 Yeah, I mean, that’s a very good question and very timely. We did a market research with
0:14:15 roughly 107 odd life sciences executives, right? And we asked this question, are you missing data,
0:14:20 or do you think data is going to be a critical factor of your AI success, right? And the response was,
0:14:24 is really interesting and very in line with what we are seeing in the industry is,
0:14:29 people are saying they are not missing data, right? Yes, there are some pockets of areas where you have
0:14:33 a panel gap, you have a channel gap, right? You’re not able to collect the data. But most of the
0:14:40 executor says the gap is connecting the data, right? Bringing the data together for the right use case to drive the
0:14:46 right insight. So it’s that inverse curve, right? 80% of the time is just stitching the data and finding where it is.
0:14:51 And then 20% of the data is making sense out of it. So as a problem, I don’t think so that,
0:14:55 obviously the data will evolve. We need more data always, right? So that we understand every aspects
0:15:01 of it. But right now for AI, the problem is how do you bring the right data into the system so that
0:15:05 you can generate the insight that the business, right? And don’t wait for a year to create, you know,
0:15:10 a multi-mega data lakes and data system that has always been a journey that we always take, right?
0:15:10 Right.
0:15:11 Every transformation.
0:15:19 So on this question, Noah, yes, there is some issue when companies like OpenAI or Anthropic
0:15:25 or Google are training on the internet data and maybe, I don’t know if data is running out,
0:15:30 but there’s a lot of the new data that is being created. It’s the same distribution of the old data.
0:15:36 So there is no new insight likely to come because it’s AI creating data, but it’s a similar kind of thing.
0:15:42 But there is a lot of data that sits within enterprises like Equivia or in health systems,
0:15:49 et cetera, which dwarf the public data in some sense. And the AI has not been trained on this
0:15:57 data. So there is information that is available in this data that is not there in these public models.
0:16:03 Now, one way that people are approaching this, including ourselves is, okay, let’s take this public
0:16:10 models and built agents that can scour our data and generate insight from it and help you make
0:16:16 decisions. And here we are addressing the challenges that Avina was talking about is how do you get the
0:16:20 agent to go to the right data source? How do you connect across multiple data sources? How do you have
0:16:25 multiple agents, data agents, et cetera, that bring insights from multiple data sources? I think that
0:16:32 that will take us quite far in generating insights from this data. But I also think we need to be
0:16:41 thinking about domain-specific foundation models, like say, Alpha Ford from Google, where we train models
0:16:48 on this domain-specific life sciences or healthcare data. And some of this has started to happen, but not at
0:16:55 scale, which allows us to understand the underlying distribution in this data and compress this data to
0:17:00 get information from this data that we can’t get from agents. Because agents, by definition, will only
0:17:07 be able to look at a sub-part of the data distribution. For every task, they cannot look at the entirety of the
0:17:13 data. So I think there are these two things that we need to do: the agents to generate insights from
0:17:20 the data as it is and better connect the data, but also think about domain-specific models that we build.
0:17:25 And we need to do this relatively quickly. And to Avina’s point, a lot of companies have spent time,
0:17:31 let’s build this beautiful data lake, let’s get all this data together, connect them, and so on. To me,
0:17:38 that’s like a fool’s errand to try to do that. I think let’s get going with the data that we have, and
0:17:42 connect, and prepare, and do this, do this fast.
0:17:49 I’m speaking with Raja Shankar and Avanab Roy. Raja is Vice President of Machine Learning at IQVIA,
0:17:55 and Avanab is VP of Commercial Analytics Solutions at IQVIA. And we’ve been talking about the healthcare
0:18:03 space as it relates to Agentec AI specifically, and everything that comes with working with processing,
0:18:08 mining data for insights. As Raja was just saying, certainly no shortage of data, but one of the big
0:18:15 challenges now is figuring out how to generate the right insights from that data. And training specialized
0:18:21 models, as we were talking about, is one of the paths right now that lots of folks are focused on.
0:18:26 I wanted to ask you, you know, the whole point of healthcare, right, is serving the patients,
0:18:33 patient outcomes, patient impact. So knowing that that’s the end goal here for what IQVIA is doing,
0:18:38 what your partners and customers are doing, how will actioning these opportunities, how will
0:18:43 developing Agentec AI systems really in the end support patient outcomes?
0:18:51 Think about it this way, Noah. Today, there are so many treatments available for patients, but there’s
0:19:00 a very heterogeneity of how physicians prescribe these patients today. We have enough data in the EMR
0:19:05 systems and others to determine, even with the treatments available today, to optimize the treatment
0:19:10 regimen and the treatment sequencing for each patient to optimize their outcome. Like for this cancer
0:19:16 patient, should I start on an immunotherapy or a combination immunoplasm chemotherapy and then go
0:19:24 on to something else? That information exists today. Similarly, we have information in the thousands and
0:19:29 thousands of half a million or more clinical trials that have been conducted today as to what works and
0:19:36 what doesn’t work. But unfortunately, this data is siloed. Every sponsor has their own clinical trial data.
0:19:44 If they were able to pull this data together and run AI on this, we would design much better clinical trials
0:19:49 than we are able to today. We would reduce the rate of failure of clinical trials and we would make sure that
0:19:54 the right patient gets the right treatment. So the AI has a lot of potential here to completely change this.
0:20:01 The challenge is not the technology. The technology exists today. The data exists today. The challenge
0:20:05 is partly organizational and partly the silos that exist in the industry.
0:20:09 I couldn’t agree more with this one. And the similar thing goes on the commercialization, right?
0:20:15 I think how do we serve the patient? End of the day, everything that we are doing is not about
0:20:20 quicker volume, higher volume of revenue. It’s more about serving the right patients and
0:20:25 at the right time, right? So if you think about as identifying where the unserved patients are,
0:20:31 where the underserved populations are, and triaging your drug launches, your drug messaging towards
0:20:35 those patient population. So your launch is not just you have a successful launch in terms of volume,
0:20:40 you have a successful launch because you’re reaching the right patients, right? Personalized engagement,
0:20:44 making sure that the HCPs who are actually serving those patients understand your drugs,
0:20:49 right? Bring them to the efficacy and the compliance, right? And adherence, right? So
0:20:54 that the patient just don’t take the drugs once and then fall off. They understand and they get educated,
0:20:59 right? So and then the continuous feedback loop because these agents are learning from the real
0:21:05 world experience of what worked, what didn’t work, right? So as you give more data and you give more
0:21:10 domain expertise to it, it’s taking, it’s getting much more smarter and giving much more better insights and
0:21:16 much more better workflow orchestration. And no, just on that, one other thing I wanted to
0:21:22 raise is often people talk about like, we talked about the silos, that’s one piece, but there’s also
0:21:29 data that exists in national health systems like NHS in the UK and other countries. But there’s a lot of
0:21:37 reluctance sometimes justified to apply AI on this data because there are fear of data privacy,
0:21:44 data leakage, security, all of these issues. And these are legitimate concerns. But we really need
0:21:51 to overcome these concerns because by not applying AI to these data sets today, it’s costing lives.
0:21:57 If we actually generate this insights, say for cancer patients or diabetes or cardiovascular,
0:22:02 we could change the treatment paradigms. We could see more patients quicker. And this has an impact on
0:22:08 actually quality of life as well as the length of life that patient has. So this is the balance,
0:22:15 right? How do we balance the data privacy and security versus the actual clinical outcomes for
0:22:20 patients? We haven’t got that balance right today. The technology exists and we need to figure it out.
0:22:20 Yeah.
0:22:26 Along that theme of balance, but from a different perspective, companies that are thinking of
0:22:34 adopting new technologies, adopting agentic AI, maybe adopting very specific agentic AI workflows for
0:22:39 clinical trials or for any of the things that we’ve been discussing. How do the companies have
0:22:46 confidence in new solutions, in leveraging, deploying an agentic AI workflow?
0:22:53 How does a company… We’ve talked a little bit about the data side and other sides about how we’re
0:22:59 making progress, but maybe not quite 100% with some of these different ways of doing things we’ve been
0:23:05 talking about. What advice do you give to an end user, a customer, a company when they’re thinking
0:23:12 about deploying an agentic AI solution and want to figure out, well, how do we measure ROI?
0:23:18 How do we think about success, you know, short term, longer term? What advice would you have for,
0:23:24 you know, an IT leader, a company thinking about embarking on a journey like this?
0:23:30 I think this question has very often been discussed with me and Rajya. We’re having multiple
0:23:35 discussions with our customers and colleagues and all. I can only imagine, yeah. But I think the answer to
0:23:41 this one is very simple, right? Which is, start with a clear business problem. Don’t get excited
0:23:45 because it’s a new AI. Where can I fit it in, right? It’s like I have a hammer looking for a nail.
0:23:51 First, understand what’s the problem, how the AI use case that you’re trying to go after fits with
0:23:56 your strategic goal. And strategic goal KPIs can be, how can we be launching our product time to market
0:24:02 quicker? Or can I increase my engagement lift with my HCPs? Or how can I have per cost acquisition of a
0:24:07 marketing campaign be lower, right? So you need to have a clear definition of a KPI, right? So your
0:24:12 problem doesn’t disappear just because you have an AI. You start with the problem. Second question is,
0:24:18 fail quickly. It’s good to have the holistic picture of the use case. Run some quick pilot,
0:24:23 right? With a clear goal that how you’re going to take decision of gate review those pilots
0:24:28 and take quick decisions, right? People struggle because they keep on, you know, extending the pilot,
0:24:33 pilot to POC, POC to something else before they can call it a full-blown operation, right?
0:24:40 Third one is, I would say, ensure your data readiness, right? Yeah, I have the AI. I got the tool from
0:24:42 OpenAI. I have a business case. Let’s run. We don’t have the data.
0:24:48 Or a data ecosystem. But when I say we don’t have the data, I don’t mean then let’s invest
0:24:53 five years of creating data lake. That’s not what I mean, right? I mean, like, in order to drive this,
0:24:57 do we have a compliant data sources? Do we have access to those data? Do we have enough metadata to
0:25:02 train the models, right? Do we have enough documentation to talk about the process workflow
0:25:06 that you’re trying to organize with the agents, right? That’s where I feel like many of the use cases
0:25:12 struggle. Because if you can’t tell the agent what to do, like a human will instruct, right? Agent
0:25:15 can’t automatically do that stuff, right? Yeah, definitely.
0:25:20 Last but not least, I would say design to scale. Don’t think from a POC. Like one of the key things
0:25:26 that we learned is POCs are very simple, right? It creates a huge amount of motivation. Oh my God,
0:25:32 it can do a specific task. The important part is that last mile taking the agent to the operationalization.
0:25:38 That is the hardest part. So think about scale, not just from a tech standpoint, but also from your
0:25:43 organizational readiness standpoint. Are you ready to adopt this? What is your change management strategy
0:25:49 looks like? Are you hiring the people for the next generation who will be working behind or beside
0:25:53 the agents, right? What will that ecosystem look like? So it’s a transformational change that people
0:25:59 need to adapt to rather than an empirical change that you’re making within silos of an organization.
0:26:07 And there’s a couple of things from Arvina that I would like to emphasize. One is on the KPIs. Aside
0:26:12 from the KPIs of speed and productivity, often the question is, how do I know that this agent works
0:26:19 is this agent works as it’s supposed to do? And then related to that is then you say, okay, let’s find
0:26:29 some ground truth data and then see if we can compare the performance. But what the challenge
0:26:35 we have found is when we have done this manually, there is no gold standard of performance because we
0:26:40 don’t know what is good look like because different people writing the same document will write it
0:26:46 differently or creating different insights will create differently. There is no benchmark saying,
0:26:52 is this 90% accurate or 70%. We don’t have any benchmarks. So sometimes we actually have people
0:26:57 saying, don’t use the historical information as ground truth because we don’t know if it’s good
0:27:01 enough. With an agent, at least you can measure that performance. So one of the things that’s going to
0:27:08 change over time is we are starting, we will start to get metrics of what good looks like in many different
0:27:13 areas, which we didn’t have before. So that’s one thing I really wanted to emphasize as we
0:27:16 as we think about KPIs.
0:27:20 No, absolutely. It’s great to point out. As we start to wrap up our conversation here,
0:27:26 I want to ask you about the partnership, the relationship. NVIDIA, IQVIA have a great working
0:27:34 relationship together. Can you speak a little bit to that, but through the lens of what comes next?
0:27:41 And, you know, it’s about the partnership, but it’s more about the technology, where this is all going,
0:27:46 where the industry, healthcare is such a broad industry, so many ways to look at this. But,
0:27:54 you know, where healthcare is going, life sciences are going through the lens of being able to do more
0:28:00 and more with the data as hardware, as software, as the whole AI stack evolves. Where do the two of
0:28:05 you see, you know, maybe speak to the relationship just a little bit now, but where do you see it headed
0:28:06 over the next few years?
0:28:14 Yeah, I mean, I think this is a partnership made in heaven, I would say. Our goals are same,
0:28:19 our objectives are driven, which is how do we make life sciences go and do what they’re supposed to
0:28:24 do quicker, right, which is putting the drugs to the patients at the end, right, and serving the
0:28:30 broader population for health. And if you think about this, I mean, NVIDIA, like we all know,
0:28:35 right, they have great capability when it comes to AI, highly optimized to run on scale. So I’m not going
0:28:39 to talk about all the great capabilities that they bring to the table with their microservices, with
0:28:45 their DGX clouds, and their multiple NemoTron models that we are utilizing. I think for me,
0:28:51 the core thing when I step back is our vision is to bring agents to life, which are life sciences
0:28:55 specific, who are trained on life sciences, understand what it takes to drive insights and
0:29:00 mine data to serve broader population of the life sciences workflows, right, which is the digital
0:29:05 workflows. How do you make that happen? In order to make that happen, you need three things,
0:29:09 right? I think we need the data, you need the domain, and you need the tech. But how do you infuse
0:29:14 that, right? Everybody talks about this, but how do you infuse that? So I think the way we see it is
0:29:20 what I think NVIDIA calls the data fly week. You get the data, you train the model, you have our domain
0:29:26 experts use this day in, day out, you capture that knowledge, and then you keep refining that agent.
0:29:31 So it’s like, you know, you have a graduate graduating from a university with all the theoretical knowledge,
0:29:36 that’s the agent on day one, and then they go for 10 years of learning of the domain, right?
0:29:42 But that agent can learn it in 10 days, right, given our uses, right? So I think that’s where
0:29:47 this infusion of this data tech and domain really comes to life for us, right? And if we can accelerate
0:29:52 that journey, that’s the goal and that’s the motto I see that IQV and NVIDIA are unlocking together.
0:29:54 Fantastic. Raja?
0:30:01 So for me, everything that I’ve been upset, I agree with. I’m looking to the future. I think there is,
0:30:07 we need to think about these kinds of partnerships differently. So NVIDIA has one advantage, right?
0:30:12 It’s not a competitor to any industry. It’s an enabler of any industry because it’s providing the
0:30:23 infrastructure and support. Similarly, if we look at parallel to IQVIA, IQVIA, as Avinov’s boss actually
0:30:29 said this once, IQVIA does everything that a pharma company does, but without having the truck,
0:30:36 which means that we can provide support to every life sciences company without competing with them.
0:30:41 So their objectives and our objectives are 100% aligned. Same as NVIDIA’s objectives and its industry
0:30:48 partners’ objectives are 100% aligned. So in that kind of context, can we think of new models of
0:30:56 partnership where NVIDIA and IQVIA can come together to catalyze the building of agentic platforms that can
0:31:03 serve multiple life sciences customers together? Because we’re enablers. We’re not competitors to
0:31:09 life sciences. So thinking about seeding and even bringing sometimes different life sciences
0:31:14 companies together, because a company that has to build any kind of platform, they have to build a
0:31:20 regulatory platform or a commercial operations platform. The ROI for every company to build its
0:31:28 own platform is not that high, unless you have hundreds of molecules that you are selling. Whereas if they
0:31:32 work with us and NVIDIA to build this kind of common platforms that can
0:31:38 accelerate everyone and they only are paying for a part of it, they get a whole value of it,
0:31:42 I think there is new models of partnership we need to think about. And this is something we are working
0:31:49 with NVIDIA on, like, how can we jointly enable acceleration in life sciences for all of our customers?
0:31:56 That’s the perfect place to wrap. That’s the goal. I love it. Raja Avanab, this has been fantastic. Again,
0:32:00 thank you so much for taking the time. A lot of really interesting, you know, the technical stuff,
0:32:05 obviously, is what drives this all in a lot of ways, but a lot of interesting considerations. I think what
0:32:09 you’re talking about at the end, Raja, about new ways of thinking about partnerships,
0:32:15 it’s kind of an era that we’re entering into across industries and certainly in the life sciences.
0:32:19 There’s a lot on the line. I’ll leave it that way. It means a lot. For listeners who would like to learn
0:32:27 more, the IQVIA website, IQVIA.com. Also, you have a big presence on LinkedIn. Good place to go.
0:32:27 Yes.
0:32:33 Fantastic. Well, gentlemen, again, thank you. It’s been a great conversation and we appreciate it. And
0:32:39 obviously, as somebody who depends on the life sciences and healthcare myself for a lot of things,
0:32:43 we appreciate all the work you’re doing and best of luck going forward. Let’s do it again.
0:32:49 Thank you, Noah. Thank you so much, Noah.
0:33:34 Thank you.
0:00:20 Agentec AI is reshaping the pharmaceutical landscape from streamlining clinical trials
0:00:26 to enhancing patient engagement. Global healthcare intelligence company IQVIA processes data from
0:00:33 over 1 billion non-identified patient records across more than 100 countries, making our guests
0:00:39 uniquely positioned to discuss how intelligent automation can transform healthcare outcomes at
0:00:45 scale. Raja Shankar serves as vice president of machine learning at IQVIA, where he spearheads
0:00:50 the application of artificial intelligence to transform research and development workflows in
0:00:55 the life sciences industry. His expertise lies in developing AI solutions that accelerate clinical
0:01:01 research and drug development processes. Avinar Broy is vice president of commercial analytics
0:01:06 solutions at IQVIA, focusing on how AI can revolutionize pharmaceutical commercialization
0:01:12 strategies. He brings extensive experience in leveraging advanced analytics and machine learning
0:01:18 to optimize brand outreach and market access in healthcare. Gentlemen, welcome to the NVIDIA AI
0:01:21 podcast, and thank you so much for taking the time to join.
0:01:23 Thank you for having us. Thanks for having us.
0:01:27 Absolutely. I really appreciate it. I know you’re calling in from different time zones,
0:01:32 had to kind of move some things to get this to work. So appreciate you guys rolling with us to get the
0:01:38 podcast out. Let’s start with the basics. And maybe Avinar, you can start and just tell us a little bit
0:01:44 about what IQVIA is for listeners who might not know, and then a bit about your own role. And then,
0:01:46 Raja, maybe you can talk about your role as well.
0:01:52 Definitely, Noah. It’s a pleasure to be here. So IQVIA is a leader in using data, tech, and analytics
0:01:58 to accelerate innovation from clinical to commercial to drive life sciences pipeline, right? So if you think
0:02:04 about this, the main objective is how do we bring the drugs that are crucial for patients faster,
0:02:10 quicker, and when they need it, right? And IQVIA is the engine that is working with every life sciences
0:02:14 company and healthcare organization to make that dream come true, right? With the power of the data
0:02:20 and analytics, which is now the way to approach this problem statement. In my role, I’ve primarily
0:02:25 focused on the commercial business, which is basically once the drug is approved, now you have to actually
0:02:31 take this drug and make it available for the patients. So in that, you need to understand where
0:02:35 are your patients? How do you reach them? What’s the messaging? How do you make sure that every
0:02:39 patient in every direction get those drugs at the right time for their disease landscape? And my job is
0:02:44 to bring the tech and analytics together with the power of the data to make that happen.
0:02:46 And Raja, on your side of the house?
0:02:54 So I’m the counterpart to Avina more on the R&D side. So just as he mentioned, we start early from
0:02:58 a clinical development and a clinical trial perspective. We run clinical trials to get the
0:03:07 drug to market and then Avina takes over. In my role, what I’m leading is all of our AI and agentic AI
0:03:13 program to transform how clinical development is done and how clinical trials are run. So how can we
0:03:20 improve the quality of clinical operations and clinical trials? And how do we speed the time it takes for a
0:03:23 drug to get to market so that patients can have it quicker?
0:03:30 – Healthcare is one of the industries that is being most impacted and advanced and kind of evolved
0:03:35 by AI, machine learning, all the things we talk about. We’ve talked a lot about agentic AI this year
0:03:42 in general. Could one of you maybe, before we dig in, just kind of give an overview of what agentic AI
0:03:50 means in IQVIA and in the healthcare landscape, as opposed to other forms of AI, machine learning,
0:03:56 that have been in play up until recently. – I mean, IQVIA is like many organizations. We have been
0:04:02 doing machine learning for a long time, maybe more than about 15 years or so, right? And we have built
0:04:08 machine learning models or supervised machine learning models to do diagnosis, prognosis, patient finding,
0:04:16 and so on. Then of course with chat GPT, generative AI came in and everybody is now aware of the chat GPT
0:04:23 function, how we can generate documents, summarize, have a conversation, and so on. The key thing with
0:04:29 agentic AI is because you can interact with the AI in natural language and it can output a natural language,
0:04:36 and it can also call upon tools, look at data, and do different tasks. This allows us to build more
0:04:42 complex systems with multiple agents where different agents do different tasks and they talk to each
0:04:47 other. The output of an agent can be the input to another agent. And specifically within the context of
0:04:55 R and D in clinical trials, we have very complex workflow. Today, they are very manually driven
0:05:02 workflows. And what we are able to do this with agents is look at entire processes and say which of these
0:05:08 aspects of these processes could be done by an agent and could be done faster by an agent. It could be things
0:05:16 like document generation. It could be things like how can we track what is happening at a site level
0:05:22 and ensure that it is ready to enroll patients faster or everything is going fine at a site level. We can
0:05:27 talk a little bit more in detail about the agents in the future, but essentially take complex workflows
0:05:34 and have agents execute them faster and better. Right. Yeah. I mean, very simplistic and very nicely
0:05:40 articulated. Right. So the way I picture this one is the ML model when we were doing the statistical
0:05:45 model, AI ML, it’s like showing an image of the car and asking, is it a car? Then we went into a
0:05:50 generative AI where it says, show me a car. And it basically shows you an image of the car by predicting
0:05:57 it. And now you’re asking, drive me to the work and the car takes you. There’s the next era of it. Right. So I think
0:06:04 every aspects of the workflows within life sciences is trying to be identified. Right. Wherever there’s a
0:06:09 task that’s repetitive with human driven, for example, you are about to launch a product and
0:06:13 you’re trying to understand the disease landscape, patient burden, access control, so that you can
0:06:17 launch your product with the right success. Right. So that it can reach the patient. Think about
0:06:22 commercial operations. You build a strategy. You have a great plan. You know who to reach out,
0:06:27 but operational breaks down. Right. How do you actually reach out to the patient? Who are the HCPs that you
0:06:33 need to call upon? How do you pass on your message to your KOLs? And it happens today so rigorously with a lot of
0:06:39 manual efforts, a lot of Excel, SQL agents coming in and being a companion, helping the commercial ops to
0:06:43 get there faster. Right. And then end of the day, the engagements. Right. How do you make every
0:06:49 engagement count and make impactful? So across the value chain from the commercial space, agents are going
0:06:54 to make every process squeeze the time. And at the end, you’re getting the drugs to the right patient at
0:06:58 the right time. Right. So that’s where we see agentic. Yeah. What’s the current state of
0:07:05 agentic AI adoption across pharma? Is it, are we super early stages? Are we a little ahead of
0:07:11 other industries? What’s the state like? I think my, my point of view on this one is we are breaking that
0:07:18 pilot barrier and try to enter into a full-blown adoption and scale. But with that being said,
0:07:22 you know, there’s a lot of stats out there that it’s not going very well. You know,
0:07:27 the pilots are roaring success, but when it comes to large scale adoption, large scale
0:07:33 operationalization, companies are struggling. Right. And like any change, it’s a tough thing,
0:07:38 right? But here, one of the greatest things that has happened is this technology has been so easy to
0:07:42 adopt and so reachable. And it’s, it’s really, that’s why we all are talking about this. Right.
0:07:48 Sure. But the flip part of it is, it is doing what human does. Right. And try to be part of your
0:07:53 equation. So you have a new member coming and joining your house and you know, you need to adjust with
0:07:58 them. Right. It will take strategy. It will take an organization. So it will take the whole village to
0:08:03 come together to, to bring the new digital persona into your home. I, I, as you guys were talking,
0:08:07 I was thinking about how I actually need to make a doctor’s appointment. And I know this isn’t quite
0:08:12 life sciences, but I was thinking, man, I need an agent that can like make the appointment for me
0:08:16 and get me to the office. That’s what I need on my end. Right. Yeah. And then there’s the other side
0:08:20 of the equation, right. Which is the compliance governance regulations, right. To your point.
0:08:24 Yeah. How do you make sure that, that it becomes, it suggests the right doctor with the right
0:08:28 specialty to treat your specific one. Right. Yeah. This is not one of the area where you take
0:08:33 few shots and one of them hits at your success. Right. You need to make sure every shot counts. Right.
0:08:38 Right. But those, I think that the barriers that, that yet to be uncovered in a full potential
0:08:45 realization. Raja, on the R and D side, where do you see the, the biggest opportunity for agentic AI
0:08:51 to enhance life science workflows? I think there are two big types of transformation that are likely to
0:08:58 happen with AI and agentic AI. When I look at AI, it’s also this kind of generative models that have a lot of
0:09:04 power. And these two opportunities are going to be clinical development simulation, which means,
0:09:10 can we simulate what is going to happen in clinical development and clinical trials before we spend
0:09:17 tens or hundreds of millions of dollars on it so that we are more confident of success of how we design and
0:09:23 execute the trials, both from a trial design perspective in terms of efficacy, safety, inclusion,
0:09:29 exclusion criteria, as well as a trial strategy perspective into which countries and which sites
0:09:35 do we do trials. That’s one thing. Can we increase the chance of success? And then the second big
0:09:41 opportunity for me is clinical trial automation with agents. So every clinical trial has a lot of
0:09:47 processes. There are hundreds and hundreds of processes and they’re very manual and time consuming and
0:09:56 repetitive and sometimes could be mind-numbingly boring things that people have to do, like look at
0:10:03 documents and insert that in another file after looking there. So this is completely ripe for
0:10:09 agentification. And if you’re able to do that, we could reduce the time of clinical trials, which is a big win
0:10:16 to sponsors because every six months they get to market faster than the NPV, extra NPVs, and the hundreds of
0:10:18 millions of dollars for them.
0:10:23 And Avanab, on your side, on the commercial side, where do you see the biggest opportunities
0:10:25 for wins with the Gentic workflows?
0:10:31 I think it’s the same chain with the links, right? So if Raja can make the trials go faster and
0:10:37 the product comes out quicker, our journey on the commercial side starts from there, right? How do we make
0:10:43 sure that works reach the right patients, right? The one thing on the commercial side is commercial
0:10:48 side is very rich on data, all right? Fundamentally, if you think about this, behavioral data, consumer
0:10:53 data, execution data, clinical data, and you can keep on naming them, right? The challenge is too much
0:10:59 data leads to chaos, right? We have data all over the places. People are trying to stitch this data
0:11:04 together before they can strike the insight and your launch is torn because one single decision on those
0:11:09 data points can lead to a multi-million dollar or even billions of dollars lost with your launch,
0:11:15 right? So with all the launches that are planned over the next, I would say, five years across all
0:11:21 formal life sciences are very, very critical, right? So the way I see agentic ecosystem and AI ecosystem
0:11:25 evolving is on the three parts of the equation, right, that I mentioned earlier. One, as the launch
0:11:31 happens, understanding where to launch, how to launch, and how do you make your strategy impactful,
0:11:35 right? How do you create baseline forecast and then reiterate them very, very quickly to create
0:11:40 scenarios? How do you identify where is the unmet need of your patients? How do you understand what
0:11:46 is the payers are looking for in your differentiation messaging of your drug, right? So building the upfront
0:11:51 strategy, can agent bring this all structured, unstructured, heterogeneous data at a global scale
0:11:55 and enable you to give that insights when you need it? The second part of the equation is,
0:11:59 good, you have a strategy, you have to operationalize it. How do you create your, you know,
0:12:03 your standard call plan, segmentation, territory design, right? Who are the HCPs that you need to,
0:12:09 you know, enable to have those conversations, right? So your operational plan. And then the last is your
0:12:13 engagement. Good that you have a strategy, you made a plan, now you’re operationalizing it, but,
0:12:18 you know, at the end of the day, your HCPs need to be aware of your drug, your patient need to be
0:12:23 onboarded, they need to be compliant, adherent, right? So who are those people? They are your front-facing,
0:12:28 you know, communication. So how do you make your marketing channels much more automated? How do you
0:12:33 get mud merge, HCP-driven engagement, which are faster, quicker, right? So when you combine all these
0:12:37 three parts of the strategy, we see agent-to-KI transforming every aspect of insight generation,
0:12:40 task automations, and workflow execution.
0:12:47 To your first point, Avinab, about the data, we recently did an episode with another health tech
0:12:53 company, healthcare AI company called Cytoreason. And they were talking about the guest Shai was talking
0:12:59 about what he called the data insight gap, and talking about how each year, I mean, each moment,
0:13:08 really, right? There’s so much more new data being produced and generated that the chances of finding
0:13:14 insight with kind of the ratio, you know, of like overall data to insight keeps growing. And that
0:13:20 those insights are, if not harder to find, they’re just buried under more and more data. And I always
0:13:25 find it interesting because there’s talk kind of in the broader community, or there has been talk this
0:13:31 year about, you know, AI running out of data to train, and you talk about these giant foundational
0:13:36 models and where are we going to get more data and synthetic data and all of that. In the healthcare
0:13:41 space, is there, I don’t want to say too much data, but is the problem sort of inverse that, and especially
0:13:49 with a company like IQVIA, you process so much data every day globally. Is the challenge in that sense
0:13:55 finding ways to keep up with the data and find insights? Or how do you look at sort of the current
0:14:02 state of all the healthcare data available and how to turn that into, you know, actionable insights,
0:14:03 to use that term?
0:14:08 Yeah, I mean, that’s a very good question and very timely. We did a market research with
0:14:15 roughly 107 odd life sciences executives, right? And we asked this question, are you missing data,
0:14:20 or do you think data is going to be a critical factor of your AI success, right? And the response was,
0:14:24 is really interesting and very in line with what we are seeing in the industry is,
0:14:29 people are saying they are not missing data, right? Yes, there are some pockets of areas where you have
0:14:33 a panel gap, you have a channel gap, right? You’re not able to collect the data. But most of the
0:14:40 executor says the gap is connecting the data, right? Bringing the data together for the right use case to drive the
0:14:46 right insight. So it’s that inverse curve, right? 80% of the time is just stitching the data and finding where it is.
0:14:51 And then 20% of the data is making sense out of it. So as a problem, I don’t think so that,
0:14:55 obviously the data will evolve. We need more data always, right? So that we understand every aspects
0:15:01 of it. But right now for AI, the problem is how do you bring the right data into the system so that
0:15:05 you can generate the insight that the business, right? And don’t wait for a year to create, you know,
0:15:10 a multi-mega data lakes and data system that has always been a journey that we always take, right?
0:15:10 Right.
0:15:11 Every transformation.
0:15:19 So on this question, Noah, yes, there is some issue when companies like OpenAI or Anthropic
0:15:25 or Google are training on the internet data and maybe, I don’t know if data is running out,
0:15:30 but there’s a lot of the new data that is being created. It’s the same distribution of the old data.
0:15:36 So there is no new insight likely to come because it’s AI creating data, but it’s a similar kind of thing.
0:15:42 But there is a lot of data that sits within enterprises like Equivia or in health systems,
0:15:49 et cetera, which dwarf the public data in some sense. And the AI has not been trained on this
0:15:57 data. So there is information that is available in this data that is not there in these public models.
0:16:03 Now, one way that people are approaching this, including ourselves is, okay, let’s take this public
0:16:10 models and built agents that can scour our data and generate insight from it and help you make
0:16:16 decisions. And here we are addressing the challenges that Avina was talking about is how do you get the
0:16:20 agent to go to the right data source? How do you connect across multiple data sources? How do you have
0:16:25 multiple agents, data agents, et cetera, that bring insights from multiple data sources? I think that
0:16:32 that will take us quite far in generating insights from this data. But I also think we need to be
0:16:41 thinking about domain-specific foundation models, like say, Alpha Ford from Google, where we train models
0:16:48 on this domain-specific life sciences or healthcare data. And some of this has started to happen, but not at
0:16:55 scale, which allows us to understand the underlying distribution in this data and compress this data to
0:17:00 get information from this data that we can’t get from agents. Because agents, by definition, will only
0:17:07 be able to look at a sub-part of the data distribution. For every task, they cannot look at the entirety of the
0:17:13 data. So I think there are these two things that we need to do: the agents to generate insights from
0:17:20 the data as it is and better connect the data, but also think about domain-specific models that we build.
0:17:25 And we need to do this relatively quickly. And to Avina’s point, a lot of companies have spent time,
0:17:31 let’s build this beautiful data lake, let’s get all this data together, connect them, and so on. To me,
0:17:38 that’s like a fool’s errand to try to do that. I think let’s get going with the data that we have, and
0:17:42 connect, and prepare, and do this, do this fast.
0:17:49 I’m speaking with Raja Shankar and Avanab Roy. Raja is Vice President of Machine Learning at IQVIA,
0:17:55 and Avanab is VP of Commercial Analytics Solutions at IQVIA. And we’ve been talking about the healthcare
0:18:03 space as it relates to Agentec AI specifically, and everything that comes with working with processing,
0:18:08 mining data for insights. As Raja was just saying, certainly no shortage of data, but one of the big
0:18:15 challenges now is figuring out how to generate the right insights from that data. And training specialized
0:18:21 models, as we were talking about, is one of the paths right now that lots of folks are focused on.
0:18:26 I wanted to ask you, you know, the whole point of healthcare, right, is serving the patients,
0:18:33 patient outcomes, patient impact. So knowing that that’s the end goal here for what IQVIA is doing,
0:18:38 what your partners and customers are doing, how will actioning these opportunities, how will
0:18:43 developing Agentec AI systems really in the end support patient outcomes?
0:18:51 Think about it this way, Noah. Today, there are so many treatments available for patients, but there’s
0:19:00 a very heterogeneity of how physicians prescribe these patients today. We have enough data in the EMR
0:19:05 systems and others to determine, even with the treatments available today, to optimize the treatment
0:19:10 regimen and the treatment sequencing for each patient to optimize their outcome. Like for this cancer
0:19:16 patient, should I start on an immunotherapy or a combination immunoplasm chemotherapy and then go
0:19:24 on to something else? That information exists today. Similarly, we have information in the thousands and
0:19:29 thousands of half a million or more clinical trials that have been conducted today as to what works and
0:19:36 what doesn’t work. But unfortunately, this data is siloed. Every sponsor has their own clinical trial data.
0:19:44 If they were able to pull this data together and run AI on this, we would design much better clinical trials
0:19:49 than we are able to today. We would reduce the rate of failure of clinical trials and we would make sure that
0:19:54 the right patient gets the right treatment. So the AI has a lot of potential here to completely change this.
0:20:01 The challenge is not the technology. The technology exists today. The data exists today. The challenge
0:20:05 is partly organizational and partly the silos that exist in the industry.
0:20:09 I couldn’t agree more with this one. And the similar thing goes on the commercialization, right?
0:20:15 I think how do we serve the patient? End of the day, everything that we are doing is not about
0:20:20 quicker volume, higher volume of revenue. It’s more about serving the right patients and
0:20:25 at the right time, right? So if you think about as identifying where the unserved patients are,
0:20:31 where the underserved populations are, and triaging your drug launches, your drug messaging towards
0:20:35 those patient population. So your launch is not just you have a successful launch in terms of volume,
0:20:40 you have a successful launch because you’re reaching the right patients, right? Personalized engagement,
0:20:44 making sure that the HCPs who are actually serving those patients understand your drugs,
0:20:49 right? Bring them to the efficacy and the compliance, right? And adherence, right? So
0:20:54 that the patient just don’t take the drugs once and then fall off. They understand and they get educated,
0:20:59 right? So and then the continuous feedback loop because these agents are learning from the real
0:21:05 world experience of what worked, what didn’t work, right? So as you give more data and you give more
0:21:10 domain expertise to it, it’s taking, it’s getting much more smarter and giving much more better insights and
0:21:16 much more better workflow orchestration. And no, just on that, one other thing I wanted to
0:21:22 raise is often people talk about like, we talked about the silos, that’s one piece, but there’s also
0:21:29 data that exists in national health systems like NHS in the UK and other countries. But there’s a lot of
0:21:37 reluctance sometimes justified to apply AI on this data because there are fear of data privacy,
0:21:44 data leakage, security, all of these issues. And these are legitimate concerns. But we really need
0:21:51 to overcome these concerns because by not applying AI to these data sets today, it’s costing lives.
0:21:57 If we actually generate this insights, say for cancer patients or diabetes or cardiovascular,
0:22:02 we could change the treatment paradigms. We could see more patients quicker. And this has an impact on
0:22:08 actually quality of life as well as the length of life that patient has. So this is the balance,
0:22:15 right? How do we balance the data privacy and security versus the actual clinical outcomes for
0:22:20 patients? We haven’t got that balance right today. The technology exists and we need to figure it out.
0:22:20 Yeah.
0:22:26 Along that theme of balance, but from a different perspective, companies that are thinking of
0:22:34 adopting new technologies, adopting agentic AI, maybe adopting very specific agentic AI workflows for
0:22:39 clinical trials or for any of the things that we’ve been discussing. How do the companies have
0:22:46 confidence in new solutions, in leveraging, deploying an agentic AI workflow?
0:22:53 How does a company… We’ve talked a little bit about the data side and other sides about how we’re
0:22:59 making progress, but maybe not quite 100% with some of these different ways of doing things we’ve been
0:23:05 talking about. What advice do you give to an end user, a customer, a company when they’re thinking
0:23:12 about deploying an agentic AI solution and want to figure out, well, how do we measure ROI?
0:23:18 How do we think about success, you know, short term, longer term? What advice would you have for,
0:23:24 you know, an IT leader, a company thinking about embarking on a journey like this?
0:23:30 I think this question has very often been discussed with me and Rajya. We’re having multiple
0:23:35 discussions with our customers and colleagues and all. I can only imagine, yeah. But I think the answer to
0:23:41 this one is very simple, right? Which is, start with a clear business problem. Don’t get excited
0:23:45 because it’s a new AI. Where can I fit it in, right? It’s like I have a hammer looking for a nail.
0:23:51 First, understand what’s the problem, how the AI use case that you’re trying to go after fits with
0:23:56 your strategic goal. And strategic goal KPIs can be, how can we be launching our product time to market
0:24:02 quicker? Or can I increase my engagement lift with my HCPs? Or how can I have per cost acquisition of a
0:24:07 marketing campaign be lower, right? So you need to have a clear definition of a KPI, right? So your
0:24:12 problem doesn’t disappear just because you have an AI. You start with the problem. Second question is,
0:24:18 fail quickly. It’s good to have the holistic picture of the use case. Run some quick pilot,
0:24:23 right? With a clear goal that how you’re going to take decision of gate review those pilots
0:24:28 and take quick decisions, right? People struggle because they keep on, you know, extending the pilot,
0:24:33 pilot to POC, POC to something else before they can call it a full-blown operation, right?
0:24:40 Third one is, I would say, ensure your data readiness, right? Yeah, I have the AI. I got the tool from
0:24:42 OpenAI. I have a business case. Let’s run. We don’t have the data.
0:24:48 Or a data ecosystem. But when I say we don’t have the data, I don’t mean then let’s invest
0:24:53 five years of creating data lake. That’s not what I mean, right? I mean, like, in order to drive this,
0:24:57 do we have a compliant data sources? Do we have access to those data? Do we have enough metadata to
0:25:02 train the models, right? Do we have enough documentation to talk about the process workflow
0:25:06 that you’re trying to organize with the agents, right? That’s where I feel like many of the use cases
0:25:12 struggle. Because if you can’t tell the agent what to do, like a human will instruct, right? Agent
0:25:15 can’t automatically do that stuff, right? Yeah, definitely.
0:25:20 Last but not least, I would say design to scale. Don’t think from a POC. Like one of the key things
0:25:26 that we learned is POCs are very simple, right? It creates a huge amount of motivation. Oh my God,
0:25:32 it can do a specific task. The important part is that last mile taking the agent to the operationalization.
0:25:38 That is the hardest part. So think about scale, not just from a tech standpoint, but also from your
0:25:43 organizational readiness standpoint. Are you ready to adopt this? What is your change management strategy
0:25:49 looks like? Are you hiring the people for the next generation who will be working behind or beside
0:25:53 the agents, right? What will that ecosystem look like? So it’s a transformational change that people
0:25:59 need to adapt to rather than an empirical change that you’re making within silos of an organization.
0:26:07 And there’s a couple of things from Arvina that I would like to emphasize. One is on the KPIs. Aside
0:26:12 from the KPIs of speed and productivity, often the question is, how do I know that this agent works
0:26:19 is this agent works as it’s supposed to do? And then related to that is then you say, okay, let’s find
0:26:29 some ground truth data and then see if we can compare the performance. But what the challenge
0:26:35 we have found is when we have done this manually, there is no gold standard of performance because we
0:26:40 don’t know what is good look like because different people writing the same document will write it
0:26:46 differently or creating different insights will create differently. There is no benchmark saying,
0:26:52 is this 90% accurate or 70%. We don’t have any benchmarks. So sometimes we actually have people
0:26:57 saying, don’t use the historical information as ground truth because we don’t know if it’s good
0:27:01 enough. With an agent, at least you can measure that performance. So one of the things that’s going to
0:27:08 change over time is we are starting, we will start to get metrics of what good looks like in many different
0:27:13 areas, which we didn’t have before. So that’s one thing I really wanted to emphasize as we
0:27:16 as we think about KPIs.
0:27:20 No, absolutely. It’s great to point out. As we start to wrap up our conversation here,
0:27:26 I want to ask you about the partnership, the relationship. NVIDIA, IQVIA have a great working
0:27:34 relationship together. Can you speak a little bit to that, but through the lens of what comes next?
0:27:41 And, you know, it’s about the partnership, but it’s more about the technology, where this is all going,
0:27:46 where the industry, healthcare is such a broad industry, so many ways to look at this. But,
0:27:54 you know, where healthcare is going, life sciences are going through the lens of being able to do more
0:28:00 and more with the data as hardware, as software, as the whole AI stack evolves. Where do the two of
0:28:05 you see, you know, maybe speak to the relationship just a little bit now, but where do you see it headed
0:28:06 over the next few years?
0:28:14 Yeah, I mean, I think this is a partnership made in heaven, I would say. Our goals are same,
0:28:19 our objectives are driven, which is how do we make life sciences go and do what they’re supposed to
0:28:24 do quicker, right, which is putting the drugs to the patients at the end, right, and serving the
0:28:30 broader population for health. And if you think about this, I mean, NVIDIA, like we all know,
0:28:35 right, they have great capability when it comes to AI, highly optimized to run on scale. So I’m not going
0:28:39 to talk about all the great capabilities that they bring to the table with their microservices, with
0:28:45 their DGX clouds, and their multiple NemoTron models that we are utilizing. I think for me,
0:28:51 the core thing when I step back is our vision is to bring agents to life, which are life sciences
0:28:55 specific, who are trained on life sciences, understand what it takes to drive insights and
0:29:00 mine data to serve broader population of the life sciences workflows, right, which is the digital
0:29:05 workflows. How do you make that happen? In order to make that happen, you need three things,
0:29:09 right? I think we need the data, you need the domain, and you need the tech. But how do you infuse
0:29:14 that, right? Everybody talks about this, but how do you infuse that? So I think the way we see it is
0:29:20 what I think NVIDIA calls the data fly week. You get the data, you train the model, you have our domain
0:29:26 experts use this day in, day out, you capture that knowledge, and then you keep refining that agent.
0:29:31 So it’s like, you know, you have a graduate graduating from a university with all the theoretical knowledge,
0:29:36 that’s the agent on day one, and then they go for 10 years of learning of the domain, right?
0:29:42 But that agent can learn it in 10 days, right, given our uses, right? So I think that’s where
0:29:47 this infusion of this data tech and domain really comes to life for us, right? And if we can accelerate
0:29:52 that journey, that’s the goal and that’s the motto I see that IQV and NVIDIA are unlocking together.
0:29:54 Fantastic. Raja?
0:30:01 So for me, everything that I’ve been upset, I agree with. I’m looking to the future. I think there is,
0:30:07 we need to think about these kinds of partnerships differently. So NVIDIA has one advantage, right?
0:30:12 It’s not a competitor to any industry. It’s an enabler of any industry because it’s providing the
0:30:23 infrastructure and support. Similarly, if we look at parallel to IQVIA, IQVIA, as Avinov’s boss actually
0:30:29 said this once, IQVIA does everything that a pharma company does, but without having the truck,
0:30:36 which means that we can provide support to every life sciences company without competing with them.
0:30:41 So their objectives and our objectives are 100% aligned. Same as NVIDIA’s objectives and its industry
0:30:48 partners’ objectives are 100% aligned. So in that kind of context, can we think of new models of
0:30:56 partnership where NVIDIA and IQVIA can come together to catalyze the building of agentic platforms that can
0:31:03 serve multiple life sciences customers together? Because we’re enablers. We’re not competitors to
0:31:09 life sciences. So thinking about seeding and even bringing sometimes different life sciences
0:31:14 companies together, because a company that has to build any kind of platform, they have to build a
0:31:20 regulatory platform or a commercial operations platform. The ROI for every company to build its
0:31:28 own platform is not that high, unless you have hundreds of molecules that you are selling. Whereas if they
0:31:32 work with us and NVIDIA to build this kind of common platforms that can
0:31:38 accelerate everyone and they only are paying for a part of it, they get a whole value of it,
0:31:42 I think there is new models of partnership we need to think about. And this is something we are working
0:31:49 with NVIDIA on, like, how can we jointly enable acceleration in life sciences for all of our customers?
0:31:56 That’s the perfect place to wrap. That’s the goal. I love it. Raja Avanab, this has been fantastic. Again,
0:32:00 thank you so much for taking the time. A lot of really interesting, you know, the technical stuff,
0:32:05 obviously, is what drives this all in a lot of ways, but a lot of interesting considerations. I think what
0:32:09 you’re talking about at the end, Raja, about new ways of thinking about partnerships,
0:32:15 it’s kind of an era that we’re entering into across industries and certainly in the life sciences.
0:32:19 There’s a lot on the line. I’ll leave it that way. It means a lot. For listeners who would like to learn
0:32:27 more, the IQVIA website, IQVIA.com. Also, you have a big presence on LinkedIn. Good place to go.
0:32:27 Yes.
0:32:33 Fantastic. Well, gentlemen, again, thank you. It’s been a great conversation and we appreciate it. And
0:32:39 obviously, as somebody who depends on the life sciences and healthcare myself for a lot of things,
0:32:43 we appreciate all the work you’re doing and best of luck going forward. Let’s do it again.
0:32:49 Thank you, Noah. Thank you so much, Noah.
0:33:34 Thank you.
Learn how agentic AI is transforming pharma. IQVIA’s Raja Shankar and Avinob Roy share how intelligent automation accelerates clinical trials, improves commercial strategy, and powers better patient care — turning healthcare data into outcomes at scale.
Listen to the full archive: ai-podcast.nvidia.com
Leave a Reply
You must be logged in to post a comment.