A Product Market Fit Show | Startup Podcast for Founders
00:14:01 · Pablo Srugo, Rahul Shivkumar
[00:14:01] S0: But your point was that that part of it was very saturated, a lot of big players in there. The first step is you got to get a doctor onboard it so that the whole RCM stuff can start. And that's where you decided to kind of begin your journey. [00:14:14] S1: Exactly. Yeah. And it was a problem that again, when we evaluated solutions at Dawn, this is not something when you're running a clinic, you want to build in-house, right? Like, these are things that if you could buy versus build, you'd always pick kind of like buying. And on the revenue cycle side, there were a lot of good solutions that we could buy. But on the kind of like provider onboarding site, there weren't any. Like, we actually have to staff people internally for it. [00:14:39] S0: This is a tangent, but because you touched on it, like, build versus buy. That's just changing in such a big way. I don't know that anybody has the answer, but because you just mentioned it, you said this is not something that people would want to build. And in the past, I was just thinking face value. Why wouldn't people want to build this today? And obviously, maybe some people are not tech forward, these sort of things, but like, assume all this stuff gets easier, cloud code gets easier, more, it penetrates more, like, why not?
00:15:04 · Rahul Shivkumar
[00:15:04] S1: Yeah, it's a great question. I was referring it to a time when we didn't have this tooling, so no sense to build at all. I think right now, there's an argument for maybe some of these workflows it will sense to build. But the big thing, I think, with a lot of these AI agents that are promising outcomes, right, that are telling you, hey, we can get a resolution. As an example, I think there's a bunch of customer support tools that say, hey, we want to build you by resolution because we're very confident. I think to have that level of confidence, there's a lot of evals running in the back end, but there's a lot of edge cases being tested. I think going from zero to 80%, cloud code makes it significantly better. But the last 80 to 100 is where I think a ton of the value is today, right? Like, for us, if our enrollment workflows were only 80% accurate, like, one out of five of your doctors wouldn't be a network, right? Like, and that is a large amount, right, for an organization that could rack up in missed revenue very, very, very quickly. So I think, like, we, like, from the size of our engineering team as an example, like, we need that many people to make sure that what we're building is actually tested for edge cases. We're constantly updating it with edge cases that are coming in. And as new models come out, we're evaluating what are they specifically good at, what are they not good at, right? Like, we're constantly testing these things that I think in very, like, high, reliable, like, systems, like, the build option is so difficult because you have to account for a lot of the edge cases and a lot of, like, the last percentage, like, accuracy that you need.
00:16:37 · Pablo Srugo, Rahul Shivkumar
[00:16:37] S0: And so maybe related to that, like, when you do decide to launch in March 24, or is that when you launched or is that when you started working on this? [00:16:44] S1: We were working on it lightly through kind of, like, I'd say, like, late 23, like, winter break, like, November, December, like, the holidays we were working on it. And then our first couple of customers committed in January were like, okay, let's sprint. And then I think March is when we, like, incorporate it officially and we had, like, the first build of the product. [00:17:02] S0: And those customers were customers you had before from Dawn? Like, you knew them? [00:17:06] S1: They were friends that we met through Dawn. They were building similar types of businesses to Dawn. Yeah.
00:17:06 · Rahul Shivkumar, Pablo Srugo
[00:17:06] S1: They were friends that we met through Dawn. They were building similar types of businesses to Dawn. Yeah. [00:17:10] S0: I think that's one of the advantages of, we talk a lot about, you know, moats these days, right? And we think a lot of it obviously is BTs and it's like, one of the big ones is domain expertise has always been one, right? But when anybody can build anything, distribution just kind of jumps up, I think, in terms of importance and that's just one, it's a small example, but it's, why do you get customers faster? Why do you get more customers than the next person who has, the next company I should say, that has a similar product, right? Well, if you're from the space and you can get, people will trust you and you have built-in credibility, then it just helps you, right? Get those customers before you even have anything which helps you iterate faster, which helps you get the next feature faster, which helps you get the best product and now it's like, okay, I have the best product and the best distribution and it just becomes this execution mode more than anything more fundamental. [00:17:53] S1: Totally. I think distribution has always been king. I think that's like more so true now than ever,
00:17:53 · Rahul Shivkumar, Pablo Srugo
[00:17:53] S1: Totally. I think distribution has always been king. I think that's like more so true now than ever, [00:18:00] S0: tell me a bit about V1. Like, how do you scope it out? Because you said yourself, this can't work 80% of the way and I agree, by the way, like building with, as we all do with quad code these days, you know, the thing you get at first is so easy to get but the thing that is production grade, whether that's production to customers or production as an internal tool, that's where you spend most of the time but at the same time, you need to get something out fast, right? So how do you scope it so that you give something that is 100% even though maybe it doesn't do everything?
00:18:24 · Rahul Shivkumar, Pablo Srugo
[00:18:24] S1: I think for us, we focus a lot on human in the loop to close some of these workflows, accuracy on workflow X, right? Let's have a human on our team or on our customer's team depending on the model in which we work with them, like check that off before the work actually went out and any changes that were required kind of went back to us in two forms. One, feedback for the engineering team and then also feedback for the kind of like agent itself to like improve, right? So I think human in the loop is very underrated, right? Like I think YC had this old thing which was like, hey, initially your backend can just really be people, right? Like where the fulfillment and things like that are actually being done by people and then you figure out ways to automate it. I think we're lucky today where some of the tooling has gotten significantly better that the first version of your product can look like something like 60% done by AI and then 40% done by people, right? And then you're trying to like narrow that gap of like how can I go to 80, 20? How can I go to 90, 10? Like how can I keep the reliability really high while increasing that like ratio? But I think human in the loop is very, very underrated. [00:19:30] S0: So even your V1 actually was able to deliver onboard doctors very fast at 100% accuracy rate. The only drawback was you had to put humans in the loop to kind of make some of this thing work.