Pod Insights a16z Podcast Can AI Solve Healthcare's Urgent Workforce Challenges? w/ Ankit Jain About this episode Ankit Jain, CEO and cofounder of Infinitus Systems, Inc, joins Julie Yoo, a16z Bio + Health general partner, to discuss Infinitus's work in solving one of healthcare's most pressing challenges: workforce shortages. Leveraging LLMs and AI voice agents, Infinitus automates repetitive tasks, such as benefits verification and prior authorization, freeing up human talent for higher-value roles. Ankit reflects on the company’s journey—from early proof-of-concept calls to scaling over five million patient-centric interactions—and shares their approach to mitigating the risks of AI errors through layered guardrails.
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Transcript 0:14 Speaker 1 Describe how your company is using LLMs in healthcare.
0:17 Speaker 0 So we are a company that wants to use AI and AI agents to really solve the workforce shortage problem that's there in healthcare. And we started with voice agents that take over so voice interactions take up so much time in healthcare. And so we've built a set of voice agents that are able to augment administrative and clinical teams in their aim to serve their patient populations. And LLMs are a key part of it. And we can dissect whether it's on the speech-to-text side, the natural language processing side, the speech generation side. There's pieces, every piece of our stack is full of large language models and small language models.
0:54 Speaker 1 And you are early to the voice trend. I mean, now it seems like everyone and their mother is trying to do something in voice. And you know, tell us about, I remember you talking at the early days. I mean, you guys were founded in 2019, so very much pre-the-current hype about voice AI. And you talked a lot about when you were making your first calls. You weren't sure like how humans would respond to them. So talk to us about that, and then also like has that changed. Like, do you actually see now that humans are just inherently more comfortable with uh talking to a robot, or do you still think that there's um a lot that needs to change for this to become much more mainstream?
1:26 Speaker 0 For the record, some dads try to uh create voice agents as well. Uh but but in 2019, uh there were a couple of things. We weren't sure whether people would feel comfortable talking to machines uh if they didn't know a machine was calling. Um the first set of calls that we did uh were to united healthcare on behalf of a set of providers that we were supporting to do benefit verifications for chronic condition patients uh for very expensive therapies. So it was our, but before we did it on behalf of a provider, just to prove the viability of our business. Before we incorporated Infinitas, um we built a demo that could call United Healthcare and say, hi, I'm calling to check benefits for patient Bruce Willis because we didn't want to use our own data. We just wanted to see whether they would even talk to us.
2:12 Speaker 0 Uh the first agent hung up and we're like, oh boy, maybe this is not that great an idea. So the second time we called, and again, we got a different agent, and we said we're calling on behalf of patient Bruce Willis. And then they said, okay, uh, before we can get started, what's uh your the what's the patient's date of birth? And we said July 1st, 1960 or some some made up uh number. Uh and they said, Ah, sorry, I cannot find the patient, and we said, ah, we're sorry, we'll call back when we have the right data. Uh but that was proof point number one that it is possible to automate a call between a machine and a human. Um, and that was the jumping off point to say, let's build a company around this to automate these processes that millions of people in the US are doing every single day.
2:57 Speaker 1 So give us a sense of the spectrum, like where do you draw the line around what use cases you are currently pursuing with your platform?
3:04 Speaker 0 So just to put this in scale, uh, we've done over five million phone calls, over a hundred million hours of audio of conversations between machines and humans. Being able to scale to that in a five-year period where the average phone call is over half an hour long, would just not have been possible in a world before LLMs. Um but I want to make sure that it's not just LLMs. LLMs are a great tool for understanding for being able to go into the long tail of knowledge and bringing it back, but we use a lot of small language models that go alongside those LLMs to make sure we can give safe, compliant, guardrailed conversations, which are so critical in the world of healthcare.
3:46 Speaker 1 Got it. So that's kind of how you mitigate risk against like hallucinations and things like that is to have that.
3:51 Speaker 0 There's many, many such guardrails. There's technical guardrails, there's human guardrails, uh, because the the risk of getting something wrong is is very real. Like it can mean the patient does not get on therapy, which can lead to downstream consequences. It can mean financial losses for patients or providers. Um, and so there's a lot of scrutiny on the speed of returning the data, the quality of data that we collect, uh and the other part is just change management. Uh people want perfection out of technology rather than comparing it to the human counterpart. So, like any other technology system, we're not perfect. We're significantly better than human counterparts.
4:33 Speaker 1 Yeah, so unpack a little bit more. You mentioned therapy, you mentioned benefits verification, like what's kind of the constellation of use cases that are um the main products for from your platform.
4:43 Speaker 0 Yeah, so so when we started, uh we started in the administrative back office of chronic conditions, so benefit verifications, prior authorization, status checks. Uh those were the places where we started, because that's where almost 50% of patients who get on therapy or who are supposed to get on therapy, a script is written, never get on that therapy because of the complexities and delays in those systems, which leads to a lot of avoidable medical cost that ends up in the in the ecosystem. One of our theses at Infinitas is that a set of different agents should come together to make healthcare more proactive instead of reactive. A lot of the problems that exist in healthcare today are because of their reactivity. We don't have enough staff to get ahead of problems.
5:29 Speaker 0 We wait for a patient to get super anxious and then call in to say, what's the status of my process? United Airlines tells me every time there's a gate change, dominoes tells me who's putting the toppings on my pizza, but the healthcare system can't tell the patient that their provider has got all the medical documentation and submitted it to the peer and is waiting for a response which is expected in the next 48 hours. Or when it gets denied, hey, we're working on it, don't worry, we just need to appeal it. Yeah. The system is working. We just need to be transparent about that, be proactive about it.
6:06 Speaker 1 Right. And in order to build this like pizza tracking for healthcare, I love that analogy. I think a lot of questions in this space are around do people actually like picking up the phone and spending the time actually talking in asynchronous fashion, or um is it going to be more efficient if we open up text messaging and emails and all these other things? How do you guys approach that?
6:22 Speaker 0 I I I think the the thing that I've learned as a first-time entrepreneur in healthcare is that you have to meet people where they are. You have to meet the health care systems where they are. You have to meet patients where they are. Some patients will want to talk in voice. Some people will want to be in text, some will want to be in chat. Security compliance and all the the bells and whistles are going to be needed. But for the system to actually be transformed, we have to meet people where they are. It can't just be one modality. The truth is a patient who's given a scary diagnosis is just taking it in. They're they're stepping back and going, what does this mean for my dreams? What does this mean for my family? What does it mean for my future? They're not thinking about the 10 clarifying questions on the medical side. That's what they think about at 10 p.m.
7:08 Speaker 0 when they're laying in bed and finally trying to accept it. And at that point, how can we provide the education? How can we provide the guardrailed resources so that they can get those answers and not get up the next morning and go, I know I had those questions, I forgot what they are. Because that's what happens to all of us today. We have the questions when we're about to go to bed, the next morning we don't remember one or two of them.
7:31 Speaker 1 What one of the other things I know you guys have recently talked about is co-pilot versus autonomous. And I believe you started fully autonomous and then recently actually added a co-pilot product, which is kind of the reverse of I think what we see with a lot of other companies. Talk to us about what that the thesis was and why offer both.
7:47 Speaker 0 We think about our platform as one that provides two pillars of offerings. One are AI agents. Some of them are voice agents, some of them are data collection agents that happen online through APIs through EDI, logging into portals. Those are all autonomous agents that can complete a task end-to-end. The second pillar are AI co-pilots that meet our customers' employees where they are, but drive more efficiency there. There's certain kinds of phone calls, for example, that we believe machines can and should do because they're wrote repetitive and technology is good enough to do them today. There's others where we think humans should continue to play a role. In some cases, humans are required to play a role, whether it's some way a certified pharmacist that has to get on the line or a provider who has to get on a peer-to-peer call.
8:36 Speaker 0 But that person shouldn't have to wait on hold for half an hour before they can talk to the certified person on the other side. The machine can wait on hold. So we have a co-pilot called Fast Track that goes through the IVR systems, waits on hold, and then drops in our customers' employees when the other person is ready to talk. So it just saves them a lot of time.
8:58 Speaker 1 When you think about the, you know, the full spectrum of use cases that you are covering today. Can you give us a sense of like what percentage of those do you feel are things that actually over time will most likely be done through automated APIs once those APIs are built, because we know that not all the APAs actually exist today in our system versus things that you genuinely feel always will require some sort of like synchronous phone call-based conversation?
9:21 Speaker 0 Um as a dreamer, I I I hope a lot of the back office communication is APIs. And I hope a lot of the front office communication, patient-facing is however that patient wants it. We meet them where they are. That probably won't be an API. In some cases, it might be if information is just available in an app or a website. Uh, but it might be over text, might be over phone calls. Um, and the back office communication that stays over phone calls reduces over time and it becomes more richer clearing houses of the next level of data than than what's available today.
9:57 Speaker 1 Have you seen examples already where you were doing phone calls because Because an API didn't exist, and then the comp that your customer or the partner or whoever it is said, okay, we're getting all these phone calls, this is ridiculous. Like, let's actually move it to an API. Do you have examples of that actually happening? So
10:09 Speaker 0 a couple of fun stories. You know, when we first started making phone calls, uh, one of the most common questions that's asked when or actually it's asked on every single phone call is what's a good callback number? And we've all faced this, right? We talk to a customer service uh agent on the other side before they go too far, they're like, what's a callback number in case we get disconnected? Um, for the first three years of the company, that callback number was my cell phone. Uh right, and we were doing hundreds of thousands of calls. So I was getting 50 to 60 calls a day uh for various reasons. The customer service agent gave the wrong information, we got disconnected for some reason and comes back to my cell phone. And I kept it that way because there was a lot to be learned through those interactions.
10:54 Speaker 0 And I could also ask, hey, how was your interaction talking to an AI agent? Right? That was real feedback from people in a non, you know, not while talking to an AI agent. Um so we got great learnings and feedback there. But once in a while, we used to get a supervisor of someone in one of these large payer call centers calling me and saying, Hey, who are you and why is your bot calling us? And I'm like, uh, thank you, sir or ma'am. Uh it's an AI agent named Eva. She's very smart, uh, and we hope she can drive a lot of efficiency for you. But the reason she's calling is these providers are calling you anyway. And the next question uh that was often asked is how do we know that Eva is legitimate? Like, how do we know she's not a scammer? Like, well, when How do
11:36 Speaker 1 you know that they can use a scammer?
11:38 Speaker 0 Or when Julie calls in, how do you know they're not a scammer? Well, we asked them three HIPAA validation questions. Like, great, you should do the same with Eva. And if Eva can answer it, Eva is as good as anybody else. And by the way, we can have voices that sound just like humans, but we don't use those when we call you. Yeah, we use a sound that sounds like a computer because we're not trying to hide it. We call from the same phone number. That led us to have a lot of partnerships with the largest payers in the country, largest PBMs in the countries, where some of them have dedicated call centers just for EVA. That's interesting. Right, because those people know how to fly through these calls, they know how Eva interacts, and they're just saying, let's help as many patients as possible, get on therapies as soon as possible.
12:24 Speaker 0 And then the next thing they said is why are my humans talking to your machines? And for some of the big ones, they've created APIs in concert with our team so that some of the information we can get digitally. They would love to have it all come to us digitally because then you get rid of the phone calls. But the truth is for chronic conditions, specialty drugs, a lot of that data is deep within documents, and that payer call center, a PBM call center employee is flipping through four or five screens, multiple PDFs to get to that information. Now, one of the promises of LLMs is to say, give me all of your documents, I will extract it and create APIs. And so with some of our partners, we're working on those projects as well to create the APIs on the other side to eventually get rid of some of those phone calls.
13:10 Speaker 0 We've probably gotten rid of a few single-digit percentage of phone calls entirely. But in a pretty good percentage of calls, 30 to 40% of data we can now get digitally instead of needing on the phone.
13:21 Speaker 1 That's incredible. So you're also um referring to a bit of a network effect, if I'm hearing you correctly, where by virtue of calling someone who is not necessarily a dedicated and infinitist customer, you are getting their attention, they're changing their behaviors and opening up new um capabilities for you. Um the killer question that everyone always asks about Infinitas is do you have do you have instances where there is bot to bot conversation happening and what what occurs in those instances? There
13:46 Speaker 0 are. Um I I will say it's unfortunate that it's uh it's it's happening because you know, my my thesis has always been why should you have two machines talk to each other in English? They should talk to each other in bits and bytes. Uh, but the reality is for some large enterprises, it's easier to update their IVR systems than it is to create an API, go through info security and expose an API to an external party because they've never done that. Um change management steps with their teams, but they wanted to have the impact sooner rather than later. And so they have enabled their IVR systems to give us when we call from our phone number more data than they typically do.
14:29 Speaker 1 Fascinating. Let's actually zoom out the lens a little bit. I mean, all the uh amazing innovation that you're talking about is obviously predicated on the overall innovation that's happening in the broader LLM foundation model space. And in some ways it's amazing. Like every day I wake up and there's some new um innovation that's happened, some new um, you know, form of progress. And yet I'm guessing for an entrepreneur who has to build products on top of these things, it can be quite overwhelming. So talk to us about how that's been in the last couple of years in particular. Um, what's been your uh approach, and feel free to get a little technical if you want um about how you are able to both take advantage of the innovation that's happening, but also mitigate your risk of either tethering yourself to the wrong platform or missing out on some new progress that's happening out there.
15:09 Speaker 0 When we started the company in 2019, this was for for those in the ecosystem post-LLM for the rest of the world, pre-LLM. Sure, sure. Uh right, we were using the T5 models, we were you know fine-tuning them back in the day. And the truth was they were hallucinating a lot. They were saying things that were obviously wrong. Um so from day zero, we've been building this guardrail system to what we like to call a discrete action space. We use the underlying LLMs to understand the intents and map it on to a very safe space that we have blessed in a given conversational area. Uh, it's a very differentiated way than in the way most people use LLMs today. It worked well back then. As the LLMs have gotten better, as we've unlocked more knowledge, we found that the kinds of errors they make are much more nuanced.
15:59 Speaker 0 So having those guardrails continues to give us uh a leg up from a safety and compliance story perspective and really being able to deliver value uh to our customers in a safe and compliant way, but at the same time, the underlying infrastructure we have is focused on being able to rip and replace any model, the best one that's out there, and sometimes use multiple models from different vendors after fine-tuning them with our hundreds of millions of utterances that are labeled to our use cases. So uh it it we have the ability to move in a very nimble way to to the best models to a concert of models. Um we're thankful for all the progress that's happening. It's really helping us. Our customers don't care what models we use.
16:45 Speaker 0 They want to make sure they can deliver a better, faster and more proactive experience to their patient populations. That's what they care about. Um if we can do that, that's why they trust us.
16:57 Speaker 1 Yeah. Talk to us a little bit about the talent side of your team and how you've approached that. How it's how has it evolved? Like, is it incredibly hard to find talent to build these AI models because there's such a huge demand for them, or have you found pockets of folks that are appropriate for kind of the intersection of healthcare and technology?
17:14 Speaker 0 What's kind of the mix? So, you know, the thing that we scan for first, before anything, before technical ability is whether people have a curiosity for and a passion for solving healthcare problems. And it turns out Silicon Valley is getting an injection of people that have spent their careers working on ads and games that as they go through their careers go, I want to do something that matters to me. Like the most exciting thing to me is that when I built for a healthcare, I'm building for myself and for my family.
17:43 Speaker 1 Yeah.
17:44 Speaker 0 So that's kind of bar number one. That said, uh getting the best talent is important. I think it's getting easier over time. I think it's because every major company building in AI, the foundational models is making it easier to fine-tune, making it easier to bring your data, which becomes a moat to be able to build better experiences. So we see our strengths in building models where there are no third-party models, fine-tuning any of the available models with our proprietary data, and then building the last mile integrations to our customers and meeting them where they are.
18:20 Speaker 1 And you're you're alluding to the fact that there is like excitingly a bit of democratization that's happening across the industry with respect to healthcare people being able to build amazing AI technology and vice versa. And you know, let me ask the annoying investor question that I'm sure you always get. Why couldn't Google do this? Or why couldn't Amazon do this? Like, what's your answer to the question of why couldn't an open AI or an anthropic or a mistral do what you are doing today within healthcare?
18:43 Speaker 0 This goes back to how a lot of healthcare institutions were built. You think about most hospitals, like Stanford Hospital down in in Palo Alto over here, it was one building a hundred something years ago, and over time they add new departments, they grow, and now we've got clinics that become hospitals that become health systems, but still every team does similar things slightly differently. And there's all kinds of standard operating procedures, there's all kinds of nuances that make one size fit all just not be a reality of today. And meeting the customers where they are is something that vendors like us, like many of the in recent portfolio companies have to do. We have to go build the the integration into the workflow so that we can make the change management easy. Building the core technology is going to be commoditized over time.
19:31 Speaker 0 But actually delivering the value by being part of the workflow is where there's a lot of challenges and being able to use proprietary data to tune it, that's where the differences are. Yeah. So
19:42 Speaker 1 it's more on the commercial end of the stick, is what you're saying. Yeah. So to that end, let's actually talk about what the go-to-market has been like because you've obviously been uh live with your products over a longer period of time than many of the AI companies that exist today, and I'm sure you've seen quite a bit of evolution on that front. So we did see the advent of these sort of like chief families. AI officers and they were like the new the next version of like chief innovation officers that were a little bit outside of the main chassis of the business, but were kind of the ones who were anointed with evaluating these newfangled things that probably the CIO or the CTO maybe didn't have as much exposure to. So is that what you experienced at the beginning? Has that changed, or who who are your actual buyers and how is that evolved over time?
20:19 Speaker 0 No, it it's it's it's been interesting. Five and a half years ago when we started selling this platform of AI agents, we call them digital assistants uh then. We were selling to the PL owners, like the operations leaders that had the people doing this work. And we're just like, you have a shortage of these folks, and we can come and augment them and we can help you scale and support more patients and give them uh the experience that you want to give them. And I they said what you're telling us is black magic. There's no way a machine can have a 30 minute long phone call. Yeah, there's no way that's possible. So we spent a lot of our time proving through pilots and all kinds of uh tests that what we were doing was real, that it wasn't just pixie dust that's you know, and how are
21:06 Speaker 1 those structures? Did you just do like a here's a three-month pilot, you can use it and just see it in real life? Or we used to say, let's
21:12 Speaker 0 first understand what the data interchanges. What are the inputs and the outputs? For any agent, it's important to understand the inputs and the outputs, the goal and what you have going into it, the context. So we established that. We said, let's do a set of a hundred calls, five hundred calls, two thousand calls, whatever your team has the ability to actually analyze. We will do them, your team will do them.
21:37 Speaker 1 Let's compare them. Got it. Like a bake off.
21:40 Speaker 0 It's a big like how would your QA team look at them? And what we found was the for benefit verification calls, just as an example, we were different 25% of the time.
21:49 Speaker 1 Oh, interesting. And what was the nature of those differences?
21:52 Speaker 0 Well, at first we didn't know, and they were like, ah, you're wrong 25% of the time. And we said, well, how about we do it the AI call twice? And we found that we were different 25% of the time. So you call the same payer twice for the same patient and you will get different answers 25% of the time. Because it's a human on the other side that may have opened the wrong
22:14 Speaker 1 PDF. Okay, so you unearthed just like inherent error in the in the system. In the
22:18 Speaker 0 system that we wouldn't have known. And so we built a knowledge graph uh based on all the calls that we do to know what right looks like so that now our AI agents, if they hear something that seems wrong, will push back and say, Can you check that again? And we find 70 to 80% of the time when we push back on the case. I feel like you should publish these stats. And this is the reality. And you talk to the payers, the PBMs, or any group, everyone knows this. That's why they make multiple calls. They uh don't always believe it, they really push, they ask for escalations. When you think about time savings, yeah, it's the fact that you don't have to do the same task multiple times. It's the fact that in real time you're doing QA, it's a win for everybody. So that's how we prove to them that what we were doing was was real and good, and then you get deployed.
23:07 Speaker 0 Come the world of chat GPT, everyone's like, that's all you can do. Can't you also do X, Y, and Z and you can, you know, build build the world on top of it. And now it's a race to how much more can you do faster? Because my team wants it. And then you know, that was 2022, 2023. As I think about this year, I think there's a lot of breaks being put because there's too many AI vendors, or everybody is an AI vendor and no large entity. I think we we support something like 44 or 45% of the Fortune 50. We support the largest entities, and all of them are being inundated by every vendor saying we're an AI vendor. So those chief AI officers, they're setting up AI review boards, machine learning review boards uh that have great intentions. They want to make sure they're not introducing biases to the system.
23:52 Speaker 0 They want to make sure that they are not uh losing ownership of their data. There's a lot of these uh checklists that are being put into place, but we're still at the first innings. A lot of them are thinking about this as one size fit all for every problem that's out there. Yeah.
24:08 Speaker 1 Are you finding that people are still viewing your product as a technology purchase? Essentially, you are kind of augmenting their staff in many ways. And obviously, as we know, labor budgets tend to be much larger in healthcare. Where are you getting slotted?
24:21 Speaker 0 We've never been part of the technology budget. Oh, interesting. Because from the very beginning, we were selling to the operations leaders. Yeah, yeah. Um technology had the IT teams had to sign off and do the integrations. Yeah, but it was a way of the operations team saying, how can we support more? And especially as we went through COVID, there was a huge shortage, and there continues to be a huge shortage. And so we've become part of that operational muscles of our customers. So there's a flurry of activity which called a blizzard that happens in the first three to four weeks of January where everyone wants to make sure their patients are still active, still eligible, still fully covered. So everyone just calls the payers. So the payers staff up in October and November, do test calls in December so they can support incoming calls in January.
25:09 Speaker 0 On the provider side, everyone hires up in September, October, November. And so time after time, when folks take us in, well, they don't have to have three months of just hiring and staffing and people being paid for ramp up to do three to four weeks of work. They can use technology in the way it's supposed to be in an elastic scalable way. That that example is absolutely the case. The second, we've seen a lot of attrition in the healthcare workforce in the last few years, and we've been able to just backfill that that group. And we've been able to, for some of our customers, I think the most exciting, and this is the most exciting part is we can help drive revenue as they grow without their teams having to worry about uh to worry about where are they going to find workers in a in a time of shortage, especially pharmacies are growing at a faster pace than ever before.
25:59 Speaker 0 Many of the largest ones are our customers, and every year, the biggest ask they have of the CFO is more headcount, and now they're saying I don't need as much new headcount.
26:09 Speaker 1 Yeah, that's amazing. So you're kind of like leaning, there's a lot of you know fearful talk about replacing jobs and all that, but what you're saying is you're basically that there's enough surface area where you're focused on things that we just couldn't even hire enough humans to do, things that um where there's natural, you know, kind of churn, let's call it in the in the workforce and no need to kind of backfill one-to-one, and then you know, shifting. You mentioned earlier, like shifting the humans to actually be able to work on more patient-facing things on higher value things that humans you're qualified to do.
26:37 Speaker 0 Or you know, many RCM groups uh spend more than half their day waiting on hold. Yeah, right. Every claim is bespoke. Uh that those workflows we we believe are, you know, when you can get the data digitally, great, through APIs is great, but the conversations actually get pretty complex. It's like, hey, you sent me a check for $15.85. It was supposed to be 150 dollars and eighty-five cents. Um why was that? Um, oh, by the way, you say you shipped it out. Can you give me the FedEx number while that's happening? You're going to FedEx's website, checking the uh where it is. Those people shouldn't waste half their time waiting on hold. And having an AI co-pilot that can do that drives a lot of efficiency, and no one joined healthcare to sit waiting on hold.
27:18 Speaker 1 Yes, yes, absolutely. Maybe just like a market landscape perspective question here. Like there seems to be centers of gravity forming around sort of the killer apps, right? With healthcare AI, you guys are working on one version of it. You have things like AI scribes that are more you know, provider-facing, you have a lot of these sort of non-voice, you know, kind of like autonomous, almost like the next generation RPA type tools. Um, how do you see all of those things either converging or diverging over time? Like, do you think each of those actually end up being completely separate markets that don't overlap at all? Or do you actually see um because of the innovations on the technology side, it being easier for a singular company to become almost like a super app platform across multiple use cases?
28:00 Speaker 0 Yeah, I I think this is an open question. Whether you ask the startups in the space, whether you ask the EHRs, the epics, the Cerners, or the Salesforces, uh, I think everyone wants to imagine themselves being the orchestration layer of all these agents that come together, right? So today, uh as we're at the infancy of this revolution, if you will, everyone's building a set of agents. We're just building a set of data exchange agents. Yeah, some clinical, some administrative, some voice, some electronic, others are building scribing agents, but the layer underneath it, that last mile integration problem that we talked about, it's still to be determined who ends up owning that layer. And I think the agent companies are gonna try to go down.
28:45 Speaker 0 The platform companies are trying to try to go up, and it's gonna be interesting to see how that plays out over the next couple of years.
28:54 Speaker 1 Yeah. I think it's a great thing that there is like real market competition here that's driving the bar up on the caliber of the products that we're seeing out there.
29:00 Speaker 0 And the talent.
29:01 Speaker 1 Absolutely. The talent, the I mean, just the amount of of spend that is being absorbed by the by the system, it's all very exciting. So let's actually end with a question about what is a use case that is intractable or you know, just not yet responsible to apply AI to today that you think will be addressable in five years' time.
29:19 Speaker 0 I I think healthcare data is still pretty disparate. Um and that means that there's a lack of context. In some cases, the data is missing. In some cases, it's in a thousand different places. As we put in better technology, better agents in all of these different places. I think we're gonna see that data come together. And that's gonna create experiences that we can't imagine today. Uh, a a good example of something that, you know, as I talk to uh some of our partners, is everyone sees the potential of integrating social determinants of health data, but it's missing 80% of the time. Data collection is getting easier, data aggregation is getting easier, and the big Biggest power of today's LLMs, and I I I think it's only going to get better and better over time, is the fact that they can understand different variations of the same thing.
30:10 Speaker 0 They're translation engines. And so as we standardize our understanding of the population of individuals, we will see more personalized care, more personalized communication and a context-aware nature of communication that all of us go, how did we live in a world before this? Well,
30:27 Speaker 1 we will look forward to that future. I look forward to doing this podcast with you in five years. Thanks for joining us, Ankit. It was great to hear about what you guys are building at Infinitas.
30:34 Speaker 0 Thanks for having me, Julie.
Previous Building Cluely: The Viral AI Startup that raised $15M in 10 Weeks w/ Roy Lee Next How AI Will Impact Emergency Response: The Tech That Could Soon Save Your Life w/ Michael Chime Page updated August 25, 2026 · transcript and summary are machine-generated
