Decoder Podcast Live! Host Nilay Patel interviews CEO of ZocDoc, Oliver Kharraz

Voice AI in general is going to be a commodity in the future, but then you have to infuse it with the domain expertise.

Good to be here. Always ready.

Good to be here. Thanks.

So Zocdoc is really a platform that connects patients and doctors wherever they are. Obviously, as you point out, the marketplace, the app, is really well known, where people can just do that self-directed. But we are making sure that wherever you are as a patient, you can get access to care. We have a partnership with some health insurance companies, Blue Shield of California, for example. When you go to their website, you can get access to care. We help veterans get into care. You know, we have other surfaces that are very annoying, right? Like the phone, which seems weird for us to do given that we started out to eliminate the phone from the healthcare process. But we've recently released a product that allows you to call into your doctor and schedule an appointment, you know, with an AI agent, you know, completely autonomously. So our current trajectory is really how do we make getting access to care easy for any patient anywhere.

So I think we're definitely an app model and we have figured out how to do access to care better than anyone else in the US. Right? When you pick up the phone and you start dialing for doctors, it takes you on average 30 days till you can actually see one. On Zocdoc, the plurality of all appointments happen within 24 hours. Nearly all of them happen within 72 hours. So that's an experience that's, you know, an order of magnitude better than what you get through the phone and sort of the old modalities. But we're not trying to take sort of the platform captive. We are opening it up for others as well. You know, some of the health insurance players that I mentioned before, but we are generally thinking of ourselves as something that can be useful in meeting patients where they are and allowing them to see their doctor.

So we offer telehealth, but if you're totally honest, and this was visible, you know, early on, patients just don't really want it, right? So we offer telehealth options, we offer in-person options in everything but mental health. 95% of all appointments are in person. And here's the interesting thing. Even doctors that offer both telehealth and in-person visits, you know, they get more bookings than doctors that only offer one or the other, but the bookings are all for the in-person visits. So the patient really only values the option of, okay, maybe in the future I want to see that doctor in a telehealth visit, but right now I have a body, right? Like they want to look at my mouth, they want to listen to my heart, they want to poke my abdomen. And one of the things about somatic medicine is that, you know, telehealth is a little bit like tea pizza. It's great, except you can only eat the pizza while you're in the same room with it. Right now, mental health is very different. In mental health, the picture is exactly reversed. Nearly all of it is happening remotely and it just has, you know, tremendous advantages for both parties to do that. So I think it's a very nuanced picture and like one blanket statement isn't going to do it complete justice. We offer that as we offer all other modalities, like we offer urgent care and, you know, primary care and 250 specialties, you know, all the way to cardiac surgeons and oncologists. So you can find really any type of care on Zocdoc.

Yeah, so I have some skepticism about that future. Mostly because I do think there will be more self-medication, right? Dr. Google is going to be replaced by Dr. AI and the patient will develop their own judgment where they think that an AI is good enough to give them guidance and where they actually want human judgment. And I think it would be maybe misleading to blur the line and say, 'Oh, you're talking to AI, but I make it look like you're speaking to a human,' because the patient self-selected into, 'I want human eyes on that because I think the potential for an error is too great and the change in outcome is too significant.' Right? So this is where I think we just need to be honest to ourselves. Not everything that is possible is actually useful.

So yeah, obviously they know it's an AI and they can opt out of that experience. Frequently playing scheduling Tetris on the phone with another human isn't actually that fun, particularly when you have to wait 20 minutes to actually talk to that person, and people are okay with that. But you know, one of the big misunderstandings on how AI solutions work is that, oh, we're just automating the work of, you know, the receptionists or the call center agent. I think if you aim for that, you're aiming too short as an AI enablement company, right? Because what you need to think about is, hey, now that I have this AI and I have essentially unlimited bandwidth, how would I design this job from scratch, right? And so for example, for us, it's not, okay, how does our AI compare to human agents, but it's actually measuring the effectiveness of all the human agents, knowing the effectiveness of the AI for every type of patient, and then connecting the patient to the right resource. So if you call in for a routine thing, you just want to confirm the office location or you want to reschedule an appointment you've already made, well, do that with an AI because it's so straightforward. You'll get faster service, it will be super friendly. But if you have a complex question, well, let's connect you to the human that is best informed about that in the practice, right? And the AI can know that and can dynamically triage these patients to come in and give you a much better experience than you had before. So you should really rethink your call center, not as how do I reduce my expenses in a cost center, but how do I actually turn this into a profit center where I now lose fewer patients and have less leakage on the front end and make sure that patients have a great experience when they call me.

So it's actually very interesting because what you say makes absolute sense, minus the fact that as a patient, your experience is actually you have hundreds of different logins to all these different doctor systems. Obviously, I hope everyone uses Zocdoc, then you have only one login, but in reality, some patients still go the phone to make the appointment and they don't think about the app as an alternative. So you'd be surprised what percentage of calls that come in are actually simple things like scheduling, and they clog up the pipes for the patients that are coming in and calling about complex issues. So there is just probably a transitory period until everyone uses Zocdoc where these reschedules still happen over the phone. But then in terms of the insight, so what we see is actually that humans don't perform equally on all complex issues either, right? And so we can measure, you know, what's sort of the successful conversion rate, you know, for a call that comes in to the average human, you know, to Zoe, to other AI solutions, to the best humans, right? And when you look at this, you know, and there's been an independent study that has been done on that recently, but they found, okay, Zoe among the AI solutions is actually the best, you know, conversion rate of roughly 52%, where everyone else was below 40, and humans are typically like, the average human is in the high 40s, so comparable to the AI. The best humans are 65, right? So they are dramatically better. But are they 65 for everything? And should you use them for everything? No, you should make sure that whatever they are doing, you teach to all the other people that are answering your phone. So you up-level in general, but then also you want to make sure that you route the patient that actually has this problem that this, you know, call center rep is expert in, like that patient and that expert need to talk to each other, not, you know, some other random person on either end of that.

Well, so the AI needs to be sort of self-conscious in that way, right? And that's why you can't just leave it to the AI. I think anyone who uses LLMs finds that they are too confident when they shouldn't be, and they're not curious enough when more questions would actually be adequate to get to the correct solution. And so we have solved this in a completely different way, where we have sort of the deterministic orchestration layer that then uses LLMs selectively to make sure we parse the answers of the patient correctly. But we have a master plan and we know when a conversation goes outside the bounds of the master plan and should be transferred over to a human, and therefore, you know, like we can sort of take accountability for that. You know, like this is very different from just dumping everything in the context window of an LLM and praying for the best.

Yeah, so we're a little bit over a thousand employees and we are still functionally structured. So we have a head of sales, a head of marketing, a head of government relations and what have you. And the reason why that works for a company our size and why I think it's going to work is because of our quite unique history. We didn't have a straight line up, right? Like we've been around for a long time. We went through a major, you know, business model transition, turnaround you could call it, and it had created a kind of cohesion that sort of a 'one Zocdoc' philosophy still works, right? Like everyone in leadership is gold against the same number, and it's a number for Zocdoc in its totality. And this is why we can bring functional teams together and we don't get the typical, you know, corporate politics that make us not work. So when you...

So it's a revenue number, it's a profitability number, and like we sort of fuse that together into one score.

No, it's actually... So obviously some doctors didn't like it, and some doctors liked it a lot. And so the interesting thing about marketplaces in general is that sort of the utilization follows a power curve, right? And so, as you may imagine, if you have one flat fee, the people that are on the top end of the power curve, they're getting essentially value for free. So obviously, and the people that are on the low end of that distribution, they don't get enough value. So everyone who was sort of to the left of that distribution of our sort of new price, they love this model. And you know, a lot more, like an order of magnitude more doctors are on Zocdoc today than when we started that. Obviously, some doctors had to pay more, right? If you were getting 10,000 patients from us a year, and we had a $3,000, you know, fee, that's cents per patient, like there's no way you're getting that anywhere, including Instagram. But also, obviously, it's now that we ask you to pay a fee per patient, it's going to be a lot more. So clearly there was some adjustment. What is super interesting is that despite sort of the fact that we had to have conversations, 'Oh, your price is going up 100x,' which if you ever had that conversation like that, it's not fun. But all of these doctors, all the big spenders actually came back to Zocdoc, except for one. And they came back to us and said, you know, the quality of the patients I'm getting, the volume I'm getting, the predictability for my business is such that there is just no alternative.

Yes, for the matching process, absolutely. Yes, we do that.

So we understand a lot about the patients, and obviously they also answer questions for us, and we understand a lot about the doctors, right? And there's in some ways layers of information that are not broadly documented, really. These are things that we know between the doctors and Zocdoc, between the patients and Zocdoc, and that's the information we can use to make that match as efficiently as possible. And there's a lot of like just public information also that you need to take into account for that, you know, which doctor accepts your insurance card, right? Like which doctor actually accepts new patients? What type of patients does this doctor see? How long does a doctor typically take for a patient, you know, with your chief complaint, and how many do they see them in the morning? Do they see them in the afternoon? How many of those can they see consecutively, etc. So like these are all like meta information that we have about the doctor, and we have the direct connection to their schedules to see, okay, given that those are all the rules, which slots are even, you know, potentially available for you. And then obviously there's sort of clinical fit questions which we tackle, and actually is, I think, a very, very interesting area of growth for us.

Yeah. So first of all, you know, we are not making referrals, you know, the patients are using us to book with their doctors. But within the scope of that, from day one, from day one, the challenge was how do you make this match better, right? And like for anyone who is doing a business in the actual physical world, understanding all the outliers and all the ways in which this can be off is a critical piece because if you apply the 80/20 rule, you're going to piss off 20% of your customers and you cannot do that very often. So you constantly need to sort of zoom in and say, okay, great, like what are the remaining edge cases where this doesn't truly work. This is a problem that's a little bit like the coastline of England, right? Like if you look at it from a map, it seems like, oh, I can just trace this and I can measure that. But as you zoom in and you say, oh, but here's like a little bay. So I really like, it's really going in there, and in the bay is a rock, and so there's like another surface, and in the rock there's a crack, and then I go to the crack and there are micro cracks, and like the smaller you go in and measure, the more you realize, oh god, I will never be done with that. Like there's just too much to do. Now AI is great because we can accelerate like the kinds of problems that we can solve to make this an ever more seamless experience for the patient and for the doctor.

So we had a team on that since day one.

Except that obviously back in 2007 they were not using AI, but we were using machine learning and other techniques to improve the quality of the match. So we have a belief actually that the quality of the match is a huge determinant. Like we are, we're not trying to optimize the number of bookings in any given moment. We're trying to optimize the experience that the user has because we believe that's determinant of do they come back and use us again. Do they have a preference for Zocdoc because that's the tool that just works.

18 years later, we're still here.

On Zocdoc? Yes. On AI specifically. So yes, absolutely. So I think there too we're thinking about ways how to use AI to not just make what we have already been doing or what has already been done more efficient, but what new things are now possible that AI exists that were just not possible before. And so there's interesting things coming out in the future and I'm happy to chat when we're ready to announce them.

So I am not in founder mode, if that's the question. I actually think I only make like three types of decisions. The first one is who are the people that I trust and I bring on the bus, right? So what's the senior leadership team and like who do I think can actually help us get to that next milestone? And once I have these people in place, if I chose well, they should know their area better than I ever could, right? If I hire an enterprise sales executive, if I, you know, have to teach them how to do their job, I mishired, right? So this needs to be on autopilot, and the only way that can happen is if I don't get into their hair. The second type of decision is, you know, where risk is involved. I think organizations, they tend to drive people to take not enough risk, and that is something that as a founder you're uniquely positioned to say, you know what, I'm going to absorb all the blame if this doesn't go right. You could say I instructed you to do that, and if it does go right, it's all yours, right? You came up with it, let's go forward, right? And so this is when I see that there are areas where we should be taking risk, I get involved and I make sure that everyone knows that there is absolute license to take the risk if it's a smart one. We're not trying to jump off buildings, but you know, like there's a lot of opportunity there. And the third type of decision is really when it comes to where's the puck going, right? Like we talked a little bit before the show. I think this is a thing where you need to integrate a lot of different inputs, right? That there's obviously what's technically feasible. Then also, you know, like I talk a lot to our customers. I understand how they are thinking about the world, where, you know, they sort of have pebbles in their shoe. And then I spend a lot of time in DC and I understand, okay, what does the regulator want, right? And then you need to triangulate all these things and say, okay, great, given that, like what do we need to do, what new capabilities do we need to bring in-house to be able to manage that next challenge. And you know, I'm a believer that companies can evolve and develop new capabilities. I don't think core capabilities are boxing you in in any way, but you need to know what you want and what you need, otherwise you can't build it with confidence.

So, we'll integrate with these agents, and the reason is that I think that fear, the DoorDash fear, might be slightly flawed thinking. And here's why I think that. Here are the questions you should ask yourself. Question number one: are these agents simply going to completely displace yourself? Right? Anyone who's running a business that interacts with the real world knows that that's not going to be the case, right? Because of that learning curve, because of all the edge cases, all these things, like even if the AIs, you know, were to start learning about them, we're so much further ahead, we can always deliver, you know, a better experience. So this is the, you know, coast of England problem. We have just, our cartographers have been at this for 20 years. There's no way that anyone would catch up to us anytime soon. So they're not going to put us out of business. Now the second question is, are they going to drain the profit pools, you know, for these things? So you could say, well, you know, there's a world where you could imagine this happening, where consumers pay a subscription fee to people that build these agents, and then the agents find the optimal price for you. That flies in the face of the entire monetization model of the internet. If you look at it, like everything has been monetized through advertising, and so you'd have to believe that there's going to be an anthropological change where people suddenly say, 'Yeah, I'm actually happy to pay upfront and then maybe like collect rewards over time where this is potentially giving me better deals.' But if that were true, like everyone would be eating healthy, working out, taking all preventative tests, etc. So we just think, I just think that that is not how humans actually work. And so then the third thing is, okay, fine, the profit pools will not be completely drained, but are they going to take most of my profits away, right? And I think there we are all anchored in these last 20-plus years where Google was a monopolist and could sort of ask for these tolls. I think the tables have actually very much turned, right? There's five major LLMs or AI companies that are competing to be your agent. Imagine you had the one that doesn't let you order a sandwich, that doesn't let you book an Airbnb, that doesn't let you call an Uber, that doesn't let you book a doctor. Would you use that one? No. And so the providers of these services actually have a lot of leverage right now to sort of negotiate the kinds of relationship with these AI agents that they never had with Google, because Google was already the monopolist when they came up.

work. That destroys your leverage, right? You have to detect their agent and say you can't do automated browsing, and there's no framework. There's no negotiation framework for that.

So why they do that, right? They are not making any money, and I make money as I used to. So that's actually cool. Give me...

Well, who, how do you know? Right? Because I might know which agent...

Let's look at Uber, right? Like Uber is making money from the driver. If that wasn't the model, right? Uber would be getting all that free traffic from Perplexity. I'm sure they love that, right? And I'm sure Airbnb, if you book through Perplexity and no money flows to Perplexity, I'm sure Airbnb will love that. Like DoorDash, oh, you order through my DoorDash app and I don't have to pay you for traffic. Great. Like why wouldn't people want that?

Yeah. But if you look at the companies that create the most value, right, they are not trying to do this through pure advertising, right? Like I'm obviously, advertising is a part of everyone's revenue, but they take you transaction fees, right? You order that sandwich, you pay a service fee to DoorDash, right? You book this Airbnb, they're taking a cut on the booking fee from you. So, yeah, use the website, right? Like that is a totally fine mechanism. And if less money, I mean, Airbnb doesn't even have advertising, but if less money comes in through advertising, you'll take that right back, you know, in other ways. So, I don't think there's really a threat there. And if they are going to negotiate, right, if they do want to have some of that money, I think these companies that are the Ubers, the Airbnbs, the DoorDashes of this world are in a unique position to dictate their terms in a way that they could never do with Google.

Look, I think we are always looking to help patients wherever they are in whatever way they want to interact with us. Like we even work with, you know, health insurance companies whereas Zocdoc is completely hidden. You know, it's you log in with your health insurance company login. You see the doctors that are in network with your health insurance, you book, you use the Zocdoc pipes, but you know, as sort of the patient, as the member of that insurance company, you don't need to...

I think these doors are more open than ever. That's exactly right. And I think as Gemini is trying to be your AI agent, and ChatGPT and Grok and Perplexity and, you know, Claude to some degree, right, well, do you want to be the chat agent that uniquely doesn't have the capability to use Uber's pipes, Dash's pipes, Zocdoc's pipes? Like that would put you at a competitive disadvantage. And I think that is sort of a reality that all these companies have to grapple with, no one more than Google, who has historically enjoyed this monopoly.