Latest / Serious People Podcast / When AI Answers Your Doctor's Phone

Transcript

When AI Answers Your Doctor's Phone · Serious People Podcast

• [00:00:00] Alex Cohen: We actually remove empathy for the most part. Do you? Yeah. And so, you know, in the beginning we thought we'd have to be like super empathetic with an AI. Like, "Oh my God, I'm so sorry to hear about your cat scratched your back and you're bleeding, and that must be so painful," and all. Like you don't want that. You want like, "Hi, thanks for calling Urgent Care Clinic. How can I help you?" And you go, my cat scratched me and I'm bleeding." And it's like, "Sorry to hear that. Let's get you in." [00:00:23] Noah Levin: Welcome to the Serious People podcast. Today, I'm talking to Alex Cohen, Co-Founder and CEO of Hello Patient. Alex is working on the unglamorous version of AI, putting it on the phone with actual patients inside actual clinics where a bad call can have real consequences. We talked about what it takes to make AI work in the regulated high-stakes world of healthcare and some unexpected lessons he's learned along the way. If you follow Alex on X, you know that he does not mince words. I love this conversation, and I think you will too you left Carbon Health when they sort of cleaved off software and just became clinics, right? [00:01:01] Alex Cohen: I left Carbon, January 2024, so two and a half years ago. And I wouldn't say that we, like, clipped off product in the sense that, like, they didn't sell off the software assets. They didn't stop building software or running on their own software. But they, very much deprioritized it and moved it into maintenance mode away from- Yeah ... you know, can we go build all this stuff to go make the clinics run faster, better, more efficient, all these things. [00:01:25] Noah Levin: I first came across your name not the way everyone else does, from your Twitter shit posting. Yeah. Although maybe I, I think maybe I connected the two after I read it. [00:01:31] Noah Levin: You wrote an article, like a blog post for, a Talkdesk case study. [00:01:35] Alex Cohen: Oh, yeah. [00:01:36] Noah Levin: Yeah. Oh, yeah. You, you were talking about like telephony and- So long ago though ... so long ago. But that was... I was exploring phone solutions for, um- Honor? ... Season Health. [00:01:45] Alex Cohen: Okay. [00:01:45] Noah Levin: This is, uh, around 20, uh, 2020, 2021. And you had written this case study on, the complexity of the phone system at Carbon Health- that you had built, and I think it was a success story for Talkdesk, how you implemented them. That was like the premium tier- Yeah ... uh, we weren't quite ready for. But I remember being really attracted to being able to drag and drop, my IVR. [00:02:03] Alex Cohen: Yeah ... [00:02:04] Noah Levin: What was the lesson from building, telephony at Carbon? [00:02:07] Alex Cohen: So it's funny 'cause that, that is almost like the origin story of how we got started at Hello Patient now. Yeah. In 2021, I might've been six months into the role. I started Jan 2021. I had just left like a bad co-founder breakup for, my second company that I was running, and I was taking time off living in San Mateo trying to figure out like what I wanted to do with my life. No, San Mateo is like a really dark, dark place to, have to- Yeah. ... like, to stare at a stucco ceiling, eat some chilies, and think about- Yeah ... [00:02:34] Noah Levin: What you're doing with your life. [00:02:35] Alex Cohen: Yeah. You know, would go for walks around the San Mateo Park, whatever it was called at the time. Yeah. And, one of our angel investors became my boss, so Ayo Omojola, who you may have come across with his writing. He writes a bunch about like fintech and health tech and whatnot. So anyway, I joined Carbon Jan 2021, and I got brought on to build out a team to go do internal tooling inside the company to go make all of our teams more efficient. So we called it the productivity pod, which is a weird name for a product team, but it was really like, you know, the front desk people need this stuff, the call center people need this stuff, the, you know, practice managers, the operation. Like anyone who was not a physician, we would build software for internally. I think I might have been like three months in, somewhere around that range, and Will Abbott, our COO, he calls me and he's like, "Alex, I'm pretty f*cking sure the call center's broken." And I was like- ... "What do you mean?" And he's like, "I just don't think we answer our phones in the clinics." And he's like, "But I can't prove it yet. I'm trying to go like get a grip on the data." At the time, we'd been using Dialpad. Mm-hmm. And we had super outgrown Dialpad for a number of reasons. But, he was like, "I need you to go focus on this as like the main project of your team right now." And so I went within like two weeks, like super deep on, talked to Steven Monn, our, call center operations lead at the time, or he was like VP of support. [00:03:47] Noah Levin: This is in-house call center? [00:03:48] Alex Cohen: All, yeah, all in-house call center. But it was, it wasn't just the call center. It was like, the clinics wouldn't answer either, and then I think we had an overflow call center at the time. Mm-hmm. I'm trying to remember 'cause we changed models so many times. And I was like, "Cool, let me go deep and like figure out what's going on." So interviewed him. I, I like started inter-interviewing all of our call center ops people and I was like, "What?" Like, "What do you guys use? What are you doing?" All this stuff. Like, "Can I pull metrics from, from somewhere?" And we did. We pulled all the metrics out of Dialpad, but that required, exporting individual reports from every single clinic individually, compiling all those back to sort of the data problem that we'll talk about. No Claude at the time to just like put it all together. And I remember being like, we miss like 30 to 40% of our calls right now." Like actually- Wow ... like for every one call or 10 calls that's coming in, we miss three to four of them coming into the clinics. And, it's kind of like almost the answer to how do you get more average daily visits up. Because if you're missing calls and patients are going to Next Level or Exer, another urgent care down the block, you're losing that patient who went out of their way to find you and call the clinic and come on in. [00:04:44] Noah Levin: Honor, which is an elder care business, had a very similar there was like a 15-minute window we noticed- [00:04:49] Alex Cohen: Yeah [00:04:49] Noah Levin: Post someone calling us where if we didn't get back in touch with them- [00:04:52] Alex Cohen: They're gone ... [00:04:52] Noah Levin: They're gone. Yeah. Because the urgency of that moment, you think about what you're, what's going... I, I imagine urgent care is- [00:04:57] Alex Cohen: Yeah, [00:04:57] Noah Levin: Urgent ... a similar situation like, "I'm bleeding onto the floor right now." [00:05:00] Alex Cohen: Yeah. [00:05:01] Noah Levin: Are you open?" [00:05:01] Alex Cohen: Yeah. And highly unlikely that the team had the time to call them back anyway. And it'd be- Yeah ... like, it's not like they were coming through missed calls. Like we didn't even know dispositions of calls at the time. We didn't know- why patients were calling in. So I created a Google form and I was like, to the call center team, I was like, "Look, for the next two weeks, every call that comes in, just like log here what happened on the call." And they did, and we started getting patterns and signals out of why patients were calling in, and most of it was scheduling, rescheduling, canceling, and then a ton of questions around how do I get the COVID vaccine? What about travel testing? All these things that now are very easily answered through a, an AI agent who has context of your knowledge base. Mm-hmm. But at the time we had to go, figure out how to automate it. And I remember that was the first question I asked was like, "Okay, well there's 40% of calls we miss. We can't staff 40% more people. We can't force the call center team to, or like the clinics to pick up all their phones 'cause you never know, it could be 9:00 a.m, 30 calls come in and they're clustered because you just opened or 8:00 a.m. And, you know, they're checking patients in, they're checking patients out, taking payment, insurance, taking people to the back. So they just like physically can't pick up the phones. [00:06:03] Noah Levin: Yeah. [00:06:03] Alex Cohen: And the answer was like, I remember asking, "Can we just automate all this?" Like, there's gotta be a way that, patient calls in and we can look them up with a little like patient pop and, or data pop, whatever we called it, and say, "Okay, we know who you are. We see your appointment. When do you wanna move it to? Press one for this time, two for that time." And I remember like we had switched to Talkdesk by that point. [00:06:22] Noah Levin: I was like, "Guys," like, "I wanna build this automation." And they were like, "Well, you gotta go build it custom 'cause we don't support this today. Like, we have all the tooling and the scripting, but you gotta go build, you know, your own APIs and go do all this work and go build the decision tree." [00:06:35] Alex Cohen: And I was like, "Yeah, we're not doing that." And so what we got to with them was what they called an IVA, so an intelligent voice agent that was like version 1.0. That's when you call into AT&T and It goes, "I can understand full sentences." And you're like, "No, you can't, you f**king idiot." And so you try to say a thing and it's looking for keywords. So it's basically just like keyword matching. You say COVID test and then it goes, "Okay," like, "I think that I can match this word to COVID and so therefore like I can read you an answer to a question." But a lot of the times it's wrong because it's a contextual question. [00:07:08] Noah Levin: Okay. [00:07:08] Alex Cohen: Yeah. And it's all keyword matching. There's a specific term for it. But, uh, so a lot of people who we hire who used to work on traditional, like, voice agents would mostly be doing, keyword matching. Mm-hmm. And so literally over time, you listen to more and more calls, and you pull out keywords, and you add keywords into this, like, semantic layer that then determines what question you're trying to ask, and then it spits out a pre-canned response. Most of the call centers still do that today. But that was our only option, so we added that for travel testing and for COVID vaccines and second vaccines and all these things. And that was the bulk of the calls, is someone just like, "When am I getting my COVID results?" And so the answer was, "If you had your COVID shot within the last two weeks, we'll reach out to schedule your second shot," and whatever. Yeah. And so same thing with COVID testing and whatnot. But that didn't really, like, solve the issue. We still had a big call answer problem after that, and the team was just like, "All right, staff call centers." That is your only option. Most of the time, if you're not doing it onshore, you've got to go use, like, a healthcare-specific BPO in the Philippines or somewhere like that. So anyway, we did that migration and it was successful in the sense that now, like previously in Dialpad, if I wanted to change an IVR across 90 clinics, I had to go update every single IVR individually. You couldn't have like shared IVRs at the time. Talkdesk, I could have like a universal... They had the best, what I would call campaign, like inbound IVR drag and drop builder, which is why we went to them. Where now if I needed to make a change, I could just like tweak a singular IVR with some conditional logic for some locations or some clinics [00:08:33] Noah Levin: Can I tell you the experience I'm having when you, when I listen to you like share the stories? I, I, you know the, you know when you watch, movies from like the '90s and the whole plot, hinges on, not being able to call someone's cellphone 'cause cellphones didn't exist, right? Yeah. I... there's like so much of the story just would not be po- A [00:08:51] Alex Cohen: Thing today. [00:08:52] Noah Levin: It wouldn't be a thing today. Yeah. Even like the three months it took you to learn, this setup, so much of it feels like anachronistic now. [00:08:56] Alex Cohen: Yeah. I mean, it's fu- it's funny how much has been solved with just access to AI that can answer a lot of these questions and can help run the RFP process and can go, you know, do even a lot of the configuration, right? It's like now- [00:09:07] Noah Levin: And can answer the phone at the end of the day. [00:09:09] Alex Cohen: Yeah, exactly. It's funny, like I'm sure Talkdesk has an MCP and, Dialpad probably does, and they all call themselves AI companies now and all these things. But at the time, it would've been great. Like, there is a world here where I just go into my terminal or Claude Code or, Claude Desktop, and I say, "Hey, I need to update the IVR for like this location and that location." That is how we do things today. Yeah. Everything's like agent scripting and, tool calling and the ability for me to go tell an AI like what I want it to do, and I don't have to go touch a UI in 90% of my job right now. Back in the day, it was all drag and drop, but it should be way easier to just tell Claude like, "Hey, Pasadena location is closed for the day. Can you just go like mark that one as off?" And then it like goes and does that work. [00:09:50] Noah Levin: Today's episode is sponsored by Valet. Alex and I talk about what it takes to put an AI product into the middle of a real operation, and one part that frequently trips up companies is what to do after the AI makes something useful. So you have a dashboard, an analysis, an internal site, or a small tool. It still needs a place where the team can find it and use it without throwing it onto some random public URL. That's what Valet does. Valet lets you publish AI-created dashboards, analysis, websites, and tools to a URL. New sites are private by default, and members of your organization can access them after signing in. They stay at the same address, so the useful thing that your team made last month doesn't disappear into the terminal or a Slack thread. Valet is the founding sponsor of Serious People. The team took a bet on this show before it existed, and I'm grateful to have them along for the ride. Go to valet.dev to try it out now [00:10:42] Noah Levin: W- we gotta come back to like what is Hello Patient in a second. Yeah. But that, that way of controlling the, configuration, is that something that your team is doing or is that something that you're giving control to the office staff- [00:10:54] Alex Cohen: It's mixed. Like responsibility fall 'cause we don't, in most cases, like we don't replace their phone numbers on their websites. They have a lot of SEO associated with it. They've got, business cards and partners who all have these numbers. So what they do is they keep their CCaaS provider or their VoIP provider, and they forward calls into our system. We give them a unique like 10-digit number for that location or that call center, and that there's an AI hooked up to our own telephony. Like we run our, all of our own telephony stack on top of Telnyx and a lot of stuff in the middle. That call comes in and a, an AI answers the phone, and then it has context 'cause it knows like what number is being forwarded to us and like which destination that's going to. Yeah. But the way that we configure all of our stuff is all agentic. And so, for the most part now, all you need is a dumb answering system that can forward calls after 10 seconds of rings or even 24/7, like no rings and just have us answer all the time. So it doesn't matter how bad your, your contact center platform is anymore. We have a lot of groups who use Intermedia as an example, and Intermedia is this like legacy IVR call center platform, I guess. And, sometimes we'll just go in there and make the changes for them. Mm-hmm. And, and it's really as simple as, they have no AI. Like I've looked, I've seen it, I've tried to go under the hood. They have no AI. They say they do and it's all a f**king lie. I like hate the industry and how much they pretend that they have things that don't exist. So we just had a call forwarding rule and an IVR option and that forwards to our number and our AI picks up the phone. [00:12:15] Noah Levin: Tell the people what Hello Patient is. Yeah. [00:12:18] Alex Cohen: I would say agentic communications for healthcare clinics. And so what that means is we go deploy our own AI agents purpose-built for a healthcare group, to go do all of this work like scheduling, rescheduling, canceling med refills, insurance questions, eligibility, things like that, both on the front office side. So as you call into a clinic and now you don't have to talk to a person, you can talk to an AI to go say, "Hey, I need to make an appointment for Thursday at 3:00 PM." Or even as complex as, "I've got a runny nose, like, should I come in?" We work with ENT and allergy groups all day that have a lot of really complex triaging workflows to figure out what type of appointment you should go in for. But we do all the front office workflows there over voice and SMS, and we have a chatbot now that can live on their websites and do all the same work. And then we do a lot of, outbounding work. So what I would call like patient recall campaigns, closing care gaps, marketing campaigns, following up with inquiries that come into the websites, doing post-visit follow. I mean, there's so many reasons why you would call outbound to a patient or text- outbound. And then we do some back office work, like we process referrals now for the groups, and we reach out to the patients to get them scheduled. We are going into some prior auth workflows and we work in some RCM queues right now. We also handle, like we take over full inbound billing call centers now. So we do all this like agentic patient-facing communication work for outpatient healthcare groups. And, we build, deploy, manage, and QA our agents end-to-end for our customers. We integrate with their systems and, all they have to do is partner on making sure that we have all the right information we need to go like build an agent that adheres to their protocols. [00:13:49] Noah Levin: What it takes to pull off what you just described- Yeah ... is wild. Yeah. And I think, I wanna ask you a little bit about what makes like the inside of that box hard. Yeah. Right, because, a lot of the people who I hope will be listening to this are people who are trying to solve these problems themselves and trying to figure out build versus buy- and who are also trying to sort of clear the hurdle of belief that these problems can be solved in the first place. You listed three things that you do not do. So, I thought we should go through them. [00:14:15] Alex Cohen: Yeah. [00:14:16] Noah Levin: Tell me if I'm getting these right. So number one was, no clinical AI without humans in the loop. [00:14:22] Alex Cohen: Correct. [00:14:24] Noah Levin: I think I know why not, but why not? [00:14:25] Alex Cohen: Our job is not to give medical advice to patients or, you know, there's a, there's always a gray area in what we're doing. Like, someone calls in, they're having an emergency. We work with some mental health groups. They could say, "I'm suicidal," or have suicidal tendencies, whatever it is. Our job is not to diagnose and to, provide treatment. Our job is to find when someone needs to be escalated to a human and escalate them over there so the human can hop in. Obviously, there's always the potential that an AI misses a signal and things like that, but for the most part, our goal is to make sure you're getting booked for the right appointment- Mm-hmm ... Like, that's up to the provider to go do, and we, we believe that. And so I think, you know, there's a lot of companies that are like, "We're replacing nurses, we're replacing, you know, doctors. We're doing all these AI..." You know, one day it'll just do virtual visits with AI, which like, sure, I do believe in, you know, the 80% of happy path cases where you have a common cold or you have an STI or like whatever it is. Maybe an AI is good for doing that work if you're willing to, as the patient, take liability for the outcome of your care. But, the amount of complexity that goes into understanding human healthcare... It's not that I don't believe AI can't do that work. I just don't know who takes liability for that and how you ensure... Can you really put the burden back on a consumer to say, "I'm willing to take the risk for what I'm about to get treated for by an AI"? Some people would argue, "Yes," like, "We're, a free country. Let me do what I want." But I think, I think healthcare's so complex that like you lack all this context, that you don't know if the AI's right or wrong. [00:15:51] Noah Levin: Not knowing if the AI's right or wrong, fair enough that it's not gonna give you clinical, advice and you've sort of opted out of that use case- Yeah ... for now. Why do you trust it? How do you know you can trust it to do the routing well enough that you can at least have it be the first line of defense? 'Cause I think whatever AI is doing, there's gonna be some subset of cases where it's not confident, not equipped, where you want the human in the loop. But if you're putting AI in the critical path, you've gotta have some like Six Sigma-ish, confidence that- Yeah ... it's going to get the routing correct in that moment. [00:16:22] Alex Cohen: This is why, you know, to your point on like what any business should go build or buy, it's like fundamentally why I believe they should not go try to build the voice AI piece because there's so much, like there's so much that we do that is not just answer the phone and determine where this person should go based on routing or what type of appointment. It is, upstream evaluations against a whole bunch of like thousands of medical keywords and signals, and simulations then run on all of our different scenarios. Every time we make a prompt change, new simulations get run to look for regression testing [00:16:52] Noah Levin: A simulation is literally you are, sending fake calls to the system- [00:16:56] Alex Cohen: Yeah [00:16:56] Noah Levin: And, and seeing what it does? [00:16:58] Alex Cohen: Correct. And then there's evals that will pass or fail all those. Mm-hmm. And so you'll see very quickly if you have a, you know, an agent that is, regressing because you made some change that you can't track. And so if you are, trying to build this on your own, you're now probably stitching together four more tools to go do this work. And, you need some level of confidence that your evals now as the non-technical person, is going to work properly in all the contexts. And so we have a bunch of layers there that track. So I would say the first thing we do is we have like our global guardrails, like red flag screening in there, chest pain, shortness of breath, things like that. Can we detect if someone's got, you know, sort of depressive signals or suicidal or whatever it is. And then there's a whole bunch that we do to make sure if someone asks for a human, they get to a human. Mm-hmm. And so typically we, you know, obviously there's always the disclaimer at the beginning of every every healthcare call that's like, "If this is an emergency, hang up and dial 911." Most of the calls, like 99% of the calls that we answer are non-emergent calls. Like, you're not calling in because you have a gunshot wound and you're trying to determine if you should go to your ENT doctor. You're calling because, like, you may have an active nosebleed, and that's one thing that we will catch, and we have an entire, model essentially built around how do you triage an active nosebleed, and what does this clinic want to do regarding active nosebleeds? For example, some otolaryngologists will see, a patient with an active nosebleed. Some will tell you to go to the ER, and we actually have to build those protocols in at the clinic level. And there's a whole system around doing that, which is why it gets so complex and, and all these things. And then we have our, like, rules engine and our, and our AI and our evals and sims that will flag like, "Hey, we missed this signal and we, did the wrong thing." And then you go and, harden the system and you flag it to the clinic and all these things. And so, again, this is why we touch administrative work, not medical work, because there's always that question of, like, what's the surface area of things that you could miss? [00:18:49] Alex Cohen: Your risk is proportionally, like, significantly higher if you're doing medical triaging of things that you could miss compared to administrative work that you could miss. [00:18:58] Noah Levin: Yeah. We, at Honor, we had some similar triage when, clients were calling inbound and when care providers were calling inbound. And, we used sort of an Eisenhower matrix of, like, urgency versus importance- Mm ... to map it. But, the thing that I thought was sort of interesting is there was this, trade-off of the, the most acute, urgent, risky, situations are both the thing that you might not want to trust AI for yet- [00:19:22] Alex Cohen: Mm [00:19:22] Noah Levin: And the thing where having someone that will answer the phone 100% of the time- Yes ... right now actually matters quite a bit. [00:19:28] Alex Cohen: Yeah. There's almost, an exercise to be had as well. I mean, if you're, sort of a healthcare business and you're deciding like, "Where do I put AI into the mix?" On the voice AI side, in like the front door of your clinic is the most sensitive part of the stack- Yeah ... which is why I would not vibe code your most sensitive part of the stack. And it really is one of those things where like we actually had this today. We've been working with this ENT group for a year now, and we've done 50,000 calls with them Majority of the time the calls go great. We've had some ups and downs when we were like building early on and all this stuff. But then like one patient today, was like threatening towards the AI and then threatening towards their staff 'cause they got frustrated with the AI. And we heard it from the team, like they were not happy. And it's like on one hand it is one out of 50,000 calls that we've ever had an issue of like that level of a complaint. On the other hand, like it's so sensitive that one call could like f*ck up your entire relationship with a customer. Yeah. And so it's like walking that balance of how comfortable are you as a healthcare group knowing that AI is not going to be right 100% of the time. It's not deterministic. Knowing that, you will end up with a subset of patients who are frustrated and don't like the AI, and like how willing are... Like is that trade-off acceptable for plugging a hole in your shortage, like labor shortage staffing and for all the other issues that you have? [00:20:48] Noah Levin: We're still on number one on your LinkedIn list. Yeah. But number two, was don't pretend to be human. [00:20:52] Alex Cohen: Yeah. [00:20:53] Noah Levin: So is this why you don't want to masquerade as a human when someone calls? [00:20:58] Alex Cohen: There's a bunch of reasons why, but I would say first is when you don't pretend you're human, you get a little bit more grace for errors. And so like people will feel like if I'm calling you and you work in the front desk and like you mess things up three times, I'm like probably really frustrated and I'm like, "This person should be competent enough to go do this thing." If you're an AI who messes up two times, as a patient you're kind of like, "Eh, well I tried it. F*ck it." So you know, whatever. I'll, I'll now talk to the human. So you get a little bit more like leniency there. Second is we don't wanna trick patients because of all the negative effects that come from that. It's like we don't want them to believe that they're talking to a human and that, you know, that they could have like all this context dump. Like, we actually need content somewhat structured from the patient in order to do things. Like, it is better when you're just like, " Make me an appointment." "Okay, what are you coming in for?" "Sick visit." "Okay, great. What location?" "This location." Right? It's like you almost want the person to deterministically go through the flow that'll yield the best outcome majority of the time. But the further you err into, "I'm talking to a human," the more unstructured your, your content becomes. [00:22:03] Noah Levin: So you're giving off like old school IVR vibes to try and get that- [00:22:06] Alex Cohen: No ... [00:22:07] Noah Levin: That behavior back? [00:22:07] Alex Cohen: No, we can handle the full like, like you can say whatever you want. It's just that there are edges where like we had one patient call in and she was like, "I need to reschedule a visit for my son, but it needs to be next to the visit of my daughter and they both have soccer practice on these days." And it's like that is a really hard thing for an AI to disambiguate and like break down on a call in real time and respond in two seconds, right? A lot easier if they're just like- ... " I need to make an appointment for my son." Great. [00:22:32] Noah Levin: Pause. Like, " When do you need that?" Yeah, ex- Yeah ... and so it's more just like, can you get a person to treat it a little bit more robotic, but knowing that it is fully capable of doing all these things. It's like actually a really hard, How do you signal that you're, AI? Do you just tell them? [00:22:45] Alex Cohen: It answers as like, "Hi, I'm a virtual assistant with-" Okay ... you know, Happy Tails Clinic or whatever it is. [00:22:50] Noah Levin: Do you see th-there are, clients or I guess, patients who would rather talk to an AI than a person? [00:22:55] Alex Cohen: Yeah, for sure. And then there's some who, have a preconceived bias of what it means to talk to an AI. Yeah. And so they say, "Human, human, human," and then you have to add like one layer- ... of, pushback, which is like, "Oh, just so you know, I can do all these things." Yeah. "And it'll be faster than sitting on hold." And then you get a lot of people who are like, "All right, f**k it, I'll try it." [00:23:13] Noah Levin: Can I tell you my human, human, human story? [00:23:14] Alex Cohen: Yeah. [00:23:15] Noah Levin: I, uh, about two weeks ago, we had, part of our air conditioning system in our house- Oh, yeah ... rebuilt. And it was working totally fine before they got there, and they spent the whole day rebuilding it, and then that night it shut off So, I called them and, I was fooled for a little bit. Oh, interesting. I thought, I thought I was talking to a person. And the two signals that, it was not a person were, one, they did not seem to understand the urgency of not having AC in Austin- after it was working totally fine before the service people got there. Yeah. But actually the tell for me, and this is like because I had played with ElevenLabs in the past, you could hear, this ambient noise in the background. And it was louder than it would normally be in a call center. And I, I realized that they had checked the box for insert call center ambient noise to make it sound like it's a real call center. [00:24:00] Alex Cohen: Yeah. [00:24:00] Noah Levin: And that got me mad. So I went- Yeah ... "Human, human, human." And then, the, the coup de grâce was that, I did eventually reach a human, but there was no context handoff whatsoever from that system to the human. [00:24:12] Alex Cohen: Sounds about right. Yeah. I mean, it's funny. We, we're trying to win a deal away from SoundHound actually, like SoundHound's in the AI, voice AI business. I'm happy to call them out 'cause like their AI is really bad. So if they're hearing this, like my, my emails are open, But you know, you call in and every time the AI looks something up, I sh*t you not, it sounds like R2-D2. Does a whole bunch of beeping. And so it, first of all, it's super robotic. Like, "Hi, thank you for calling," you know, insert clinic name. [00:24:35] Noah Levin: And not as a style, but as, that's just how it actually sounds. [00:24:37] Alex Cohen: It's just how it sounds. I don't, I don't, maybe that's the style they went with. And then you're like, I need an appointment." And it goes, "Beep, beep, beep, beep, beep, beep, beep, beep, beep." Like every single time it does a tool call or like does a data lookup. And so, actually emailed the CEO of the healthcare group that's running on them, and I was like, "Can I show you AI that doesn't sound like R2-D2?" It's just like Some of them are just so frustratingly bad that, you look at it and you're like, you're just pushing us back another like five years 'cause patients are, are gonna get frustrated with the AI. So we don't try to fake that it's AI. [00:25:07] Noah Levin: But [00:25:08] Alex Cohen: We do have to convince people that it is competent, and that's like a hard, thing to do within the first 10 seconds of a call. [00:25:13] Noah Levin: How do you signal competence? [00:25:15] Alex Cohen: The pushback, like, "I can do all these things." Yeah. Or like they say something and it's like, "Look, I know you want a human." And so then you talk a little bit more human, like, "Look, I know you want a human, but I'm also like really smart AI." I will say that, OpenAI and Anthropic and Gemini and all these companies like kind of building mass market, mass consumery facing AI tooling has been sort of the best thing to like lift everyone up into people feeling more, comfortable trusting an AI to do some hard work. [00:25:42] Noah Levin: Yeah. [00:25:43] Alex Cohen: So we don't try to fake that it's human and we don't even add ambient noises right now or, or, um, you know, clicking or typing or any of that. We actually remove empathy for the most part. Do you? Yeah. And so, you know, in the beginning we thought we'd have to be like super empathetic with an AI. Like, "Oh my God, I'm so sorry to hear about your cat scratched your back and you're bleeding, and that must be so painful," and all. Like you don't want that. You want like, "Hi, thanks for calling Urgent Care Clinic. How can I help you?" And you go, my cat scratched me and I'm bleeding." And it's like, "Sorry to hear that. Let's get you in." [00:26:12] Noah Levin: Yeah. [00:26:12] Alex Cohen: Like that's actually empathetic, is getting someone to the thing that they want as fast as humanly possible, not adding a bunch of fake em- like who gives a sh*t if you have fake empathy from an AI? Like it's not real. [00:26:23] Noah Levin: Yeah. You know, it's interesting as well, we hire people who have linguistic backgrounds and can actually think through some of these things like how do humans talk? What do we need to ask? And how do we make it more personalized? How do you just like figure out how to prompt the AI into what you need it to do? [00:26:36] Alex Cohen: That goes back into what should you build versus buy. Like, if you're not a linguist and you're not a prompt engineer and you don't know how these systems work, like sure, you could throw up a-- Like it's super easy these days, like go use Bland or Retell or one of these to go throw up a, you know, a voice agent that answers the phone. For it to be like exceptionally good requires like this level of domain knowledge and consulting, not just on your healthcare operations, but on how the f**k do voice AI agents work actually. [00:27:02] Noah Levin: Yeah. Should we do number three? [00:27:04] Alex Cohen: Yeah. [00:27:05] Noah Levin: Don't try to automate 100% of calls. [00:27:07] Alex Cohen: Yeah. No, we don't. We track two metrics. One is total containment of calls that come into the AI. So like if we answer 100% of your calls, maybe we're at like 50% containment rate or something like that across all of them. What we do track though is in-scope containment. So for all the calls in which the AI should have been able to resolve the call, how many did it resolve? That number's been like exceptional recently, like 60 to 80% across the board with our use cases right now. And, that just came from like two years of heads down building a platform that could switch context at any given point in time and call a bunch of tools. It's very goal-oriented, all these things. But there's so many things that an AI doesn't have context about- [00:27:48] Noah Levin: Mm-hmm ... [00:27:48] Alex Cohen: The data problem, that it can't resolve 100% of calls. So I'll give you a good example. If we're answering calls for your billing call center, like you get a bill in the mail from Carbon, you're all unhappy. You're like, "Ah, why do I owe $300? I paid a copay at my time of visit," and blah, blah, blah, "I should be in network." You call in and our AI now can actually look up your EOB and look up your statement and see what the, what we billed out, what insurance paid, what CPT codes we used, and explain it all. But if you say, "I need to dispute this charge, actually. Like this is, I believe that you coded incorrectly. I believe that you're charging me for services that I didn't agree to." That is an entire new set of context to go give to an AI to go figure out, is this a proper dispute? Now, can you do it over a long enough period of time? Probably. Like, should you optimize for every, you know, 5% of, like use cases basically? Like, probably not right now. [00:28:39] Noah Levin: Yeah. [00:28:39] Alex Cohen: Over time, we will go chase denials and go help do disputes and go do all these things. So like you end the call and it triggers another AI, which can be longer thinking, have more chain tool calls, have more, reasoning, all these things to go look at the different systems asynchronously and figure out, did we do the right thing? But the amount of context that you need to do that, like that's why there's RCM companies to begin with. [00:29:00] Noah Levin: Yeah. So one of the things you said when you were talking to Molly, this is like, what, nine months ago at this point? [00:29:04] Alex Cohen: Yeah. Wow. [00:29:05] Noah Levin: When you did your A. [00:29:05] Alex Cohen: Yeah. [00:29:06] Noah Levin: You were talking about how every, office has its own SOPs, its own workflows. Yeah. And, the hot topic in, applied AI is how to, take your context graph, your domain knowledge, the latent SOPs that are in your company, in people's heads, and get them encoded so that an agent can do the thing. So when you go into a new client, I assume they're not just handing you the binder of how things actually need to work each time from scratch. [00:29:32] Alex Cohen: I mean, sometimes. Sometimes they don't have the binder. I would say, like, we build, vertical specific templates now that we know, like exactly, [00:29:39] Noah Levin: When you say vertical, you mean sub, subvertical. [00:29:41] Alex Cohen: Like healthcare vertical. Yeah. Yeah. Like primary care versus urgent versus ENT versus mental versus whatever. There's a really interesting dynamic right now where, like, let's say 80% of your agents are repeatable and reusable. Like the core structure and the scaffolding of that agent is the same if you're doing scheduling at one urgent care versus the other urgent care. [00:30:00] Noah Levin: Mm-hmm. [00:30:01] Alex Cohen: The 20%, which is all your custom SOPs and all of your rules and your FAQs and your binders and your insurance plans that you accept, and how you handle answering questions about insurance versus someone else, is 90% of the work. And so you have this, like, last mile problem where every implementation has a last mile problem, and we are doing a whole bunch of things around, building agent harnesses to go do that work and do the configuration on their own with the EMRs for ones that we're already integrated with, and then go actually take all this unstructured data, compare it to... We have so much data now on, like, how these things are supposed to look when, when you set them up for an agent to run. [00:30:38] Noah Levin: Can you give, like, a really narrow example of this? 'Cause I think it's, it's sort of conceptual- Yeah ... and hard, hard to imagine. [00:30:43] Alex Cohen: Um- [00:30:44] Noah Levin: What's a last mile gap? [00:30:45] Alex Cohen: Yeah, like, let's say we're integrating with eClinicalWorks. They're a big EMR provider for- they do a lot of like, you know, ambulatory care. But- Like let's say you run, Urgent Care Group One or ENT Group One, and I run ENT Group Two, and we both go to Hello Patient and we say, "I need you to go build me scheduling." You might have the same services that you see across both, but we have completely set up ECW in two completely different ways. And so what we would expect the EMR to return from the first five implementations is not guaranteed to look that way in the sixth. Mm. And so you might have 100 different appointment reasons 'cause that's how you've done things. You might have a bunch of providers who haven't been there in 10 years that we have to go filter out now. You might have a bunch of locations that don't exist. You might have like lack of metadata in those locations. You might have no appointment template rules and just the girls at the front desk like know exactly what appointment to book, but they don't block anything off on the calendar and there's no templates. And so, that's that last mile configuration problem where we have to go uncover like what the f**k are you doing with your EMR in every implementation. [00:31:45] Noah Levin: So, so is there a structure to that discovery? How do you, how do you... That sounds like a mess. [00:31:49] Alex Cohen: Yeah, it is. It's usually an... We call them agent PMs, agent product managers. They're the ones who are really, like, working with a customer directly to go get them live. That's their job, is to go and do discovery. Like, how do you guys book? What appointment types do you book? How do you do it? What questions do you ask? How do you make an appointment? What's the minimum amount of information you need? And then we go hook up the EMR and, that integration, and then we start pulling back the data, and it's like them, Claude, and God who have to go figure out, like, how this shit's supposed to work and look and be like, guys, there's like 16 different annual wellness visits in here. Like, which one do you use?" And sometimes they're like, "We don't know." And, this is like more of where we're building our own... We don't fine-tune any models or anything like that, but like building our own sort of like trained-ish AI agents on doing a lot of that probing and- figuring out that and disambiguating to come back with like a bunch of findings and say, "Hey, like here's how this EMR is configured and here's what we need to do to, you know, to set it up for this customer." So that's one layer. You have some EMR that a customer is using that you have to integrate with. We have our own practice management software internally. That's our own middleware that the AI uses for context. And so now you can imagine, "Cool, you don't have templates in your EMR? Well, we can just go build them in our system because, we built that." [00:33:03] Noah Levin: My experience as a SaaS customer and, nascent SaaS builder is connectors used to be like currency And now they're just commodity. Like, That was like the path into SaaS products- Yeah when you used to Google what systems connect to my HubSpot? Right. And so if HubSpot was the one that had 100 connectors, 100 SEO pages that they got to brag about. And now connectors, it seems like, it's sort of cheap and easy to connect to, new systems. Is that- [00:33:29] Alex Cohen: CRMs, yeah. EMRs, no. And that's kind of the distinction is like, yeah, HubSpot has an API, Salesforce has APIs, Zoho has APIs, you know, Fin/Intercom, like whoever. The EMRs do have APIs. They are not commoditized. Like, they are so hard to build against. And I think that's actually, you know one of the things that you wanted to talk about was just like how businesses should think about using AI inside their practices. And the biggest limiting factor will be, the system of record being unavailable in a way... Like, they may have APIs. They probably don't do what you want them to do. eClinicalWorks, just to give you an example. They have FHIR APIs for clinical documentation. They do not have practice management APIs for us being able to go do scheduling, so we actually use, RPA under the hood to go do scheduling. [00:34:14] Noah Levin: RPA? [00:34:15] Alex Cohen: Remote process automation, so like web scrapers and stuff like that. That's true for a lot of the EMRs, or they don't wanna give you access, or they wanna charge you as the customer to get access to your own data. Everything I would say comes down to if you were to distill everything down to, what is the hardest problem to solve, it is truly credible, create, read, update, delete, CRUD, for those who don't know what that means. But, like, truly credible access to, the systems and access to clean, normalized data. Like, those are the two biggest problems. And so, that's a lot of the work that goes into the middle of how do you make agents actually work for a bunch of workflows and use cases. Well, you need access to go read and write back to a system, and you need access to the data to be able to go figure out, like, what you're actually gonna go do. [00:34:58] Noah Levin: If you're sitting as an operator of a medical practice right now, is there something you can do to be more, AI-ready? I guess there's two versions of that question. Mm-hmm. Is there a, uh, a best-in-class, EMR to, migrate to? And if you're stuck on whatever the one was that you just said that where you have to open a web browser, [00:35:16] Alex Cohen: They're super popular too. They have, like, a million physicians on the platform. [00:35:19] Noah Levin: So if you're one of those million physicians, like, is there anything you can do to, help bridge the gap? [00:35:25] Alex Cohen: Well, we can do work on ECW. If you're trying to get AI-ready to, like, let's say you want-- Let's say you said, "Hey, I want Claude Enterprise for work so that my team can do some stuff in Claude a lot more efficiently. There's two questions to ask, which is: What do you want to do? And like, what are you trying to automate away? And then which systems are you trying to do that work in? Yeah. Now, if you said, "I want Claude Enterprise to be able to go do, work inside the EMR," that's probably really hard to get it to go do on its own. If you wanted it to go create Excel files and run financial reports and things like that, that seems way more reasonable to be able to go do right now. And the way to get AI ready for that is figure out how to get access to all of your data and export it and, you know, get it into a consumable format for another system to get access to that is much smarter than your EMR. Or go yell at your EMR and say, "I need an API access," and then go throw Claude at building an integration to it, which it may do like some pieces fine. It will probably not, but like, you know, that, that's like where I would focus on is, yeah, and do we switch EMRs? ECW, people really like, um, they've been around for a long time. Athena, way better to go integrate with. They have an open marketplace. We're in the marketplace. Like they have APIs to be able to do everything. They have like 900 endpoints. They are very marketplace open first. They're also an Austin company. And then, ModMed, same thing. If you're in specialty care, we love coming across a customer who uses ModMed because ModMed has full APIs for doing everything. Same thing- Mm-hmm ... we're in the marketplace. Veradigm, we're integrated with like fine. I would put them at like mid-tier of accessibility in terms of, hard to integrate with is what I would say. You know, then you look at, like, NextGen, they won't respond to my emails. eClinicalWorks, no API. There's a big urgent care one that's been fighting me on access now for, like, six months, and I'm still trying to work on us getting access. [00:37:08] Noah Levin: What do, what do you think that's about? Is it, a territory thing, like a fear of giving up control of the read/write surface? [00:37:13] Alex Cohen: I think it's two things. That's one, is they believe... They also feel the commoditization, and they don't want everyone, you know, extracting value where they could capture value as the system of record. [00:37:22] Noah Levin: Yeah. [00:37:23] Alex Cohen: The second is, they're very protective over their data because they don't want anyone having unfettered access to all this patient data. But even you as the cust- like, as the healthcare provider, the conversation goes, "Hey, we wanna do some stuff- "... with our data. We own our data. You should give it to us." And the EMR company is like, "Well, no, you don't f**king know what you're doing with the data, and we don't trust you to have access to all these." And, like, "Where are you gonna go pipe all this data to?" Right? So they have a responsibility to patients at the end of the day to make sure that none of this stuff, gets leaked and you have, like, a bunch of, HIPAA violations. Right. So I think that that is part of it, which is why we go through certification processes with a lot of the EMRs and show that we have, like, our security posture in check And, like, You know, if you're a 10-location urgent care group, like, you're not, getting SOC 2 Type 2 compliant 'cause you're not a software company. But we are, and we're spending the tens of thousands of dollars a year to go keep all those things up to date So I think that's like, you know, if you are a plumbing business or a home services business, fine. Like, that data is way less sensitive, and it's way less regulated than, access to your medical records and charts and appointments and doing things like that, which is why I think healthcare is, like, the hardest industry. Finance and healthcare are the two hardest industries to build in. [00:38:33] Noah Levin: Yeah. I worked at, Season Health, which was a healthcare company, and we were HIPAA compliant. And I swore after leaving there that I would never work in healthcare again. And then I ended up in charge of HIPAA compliance for, uh, Honor. [00:38:44] Alex Cohen: Yeah. [00:38:44] Noah Levin: It's a slog, man. Yeah. It's a slog, especially if you're not built from scratch to be that. What,What's been hardest about maintaining like a strong HIPAA posture? [00:38:52] Alex Cohen: It's not that it's hard, it's just tedious. It's like every single vendor we have to have a BAA with that's touching PHI, or we have to have a zero data retention policy with, and then you have to go report your sub-processors back to your customers. And then a lot of them try to negotiate contracts like, "If you ever change a sub-processor, you are obligated to let us know." And I'm like, "No, we f*ck around with a whole bunch of models, and we do a whole bunch of things, and like vendors that we are signing BAAs with and doing this work, but we may never ship them into production, so like there's no reason for us to go disclose all that. We will give you a sub-processor list upon request," is like how we renegotiate those things back in. Mm-hmm. You spend a lot of money, like there, there's full audit trails, everyone needs full role-based access control. If you have any international employees, they're logging in via VPNs and VPCs, and they're signing BAAs with the company. It's kind of just like prohibitive legal and compliance work is what it is to keep bad actors out of working in healthcare 'cause it is expensive. So I think that's part of it. We're starting to enter the UK actually with a customer who's signing that's in the UK. And like now we're going through, a whole bunch of UK regulatory compliance pieces and like probably gonna have instances of our stuff all UK-based. We have to have DPAs with our vendors again here. They have to be like a- assign SCCs and all these different things now for a different regulatory body. That's an enterprise customer who's in the UK, which is the only reason that we're doing that. But, you know, the US like to its credit, like has good healthcare security posture, even though a lot of it feels fake much of the time. [00:40:19] Noah Levin: Uh-huh. [00:40:20] Alex Cohen: Does AI make any part of compliance easier? Yeah. I mean, we can like go check controls really quickly, and we can go mitigate on, on, um, you know, like we use Secureframe, which is a, SOC 2 monitoring platform and, we just finished a really rigorous audit with a auditor to go get our SOC 2 Type 2 com- like it just got done this week actually finally. Congrats. Thank you. Yeah. I would say like, the AI helps as a co-pilot for the team that's working on just making sure all of our documents are in place, and we're just like checking the boxes and everything and, Yeah ... but it still requires, like our director of engineering has to go into AWS and make sure everything is turned on in the way that it should be, 'cause the auditors are gonna go in and be like, "Is everything turned on the way that it should be?" [00:40:57] Noah Levin: Where can people find you if they want to see more from you? [00:40:59] Alex Cohen: Twitter or X, whatever, @anothercohen. I suppose LinkedIn, hellopatient.com. [00:41:05] Noah Levin: I suppose LinkedIn. That, I think, I think, I think that's a T-shirt. [00:41:08] Alex Cohen: Yeah. I'm also not the biggest fan of LinkedIn. I just like sh*t posting on there. [00:41:13] Noah Levin: Yeah. How can people be useful to you? [00:41:15] Alex Cohen: If you're a healthcare group, come to us and, we'll figure out if we can bring AI into the group. We're hiring, which has been a lot of fun. I think we have like 10 open roles right now across engineering, agent PMs, head of product, customer integrations engineer, a biz dev person to go do some stuff on like the EHR side and, some other channel partnerships. So yeah, hiring and customers are like the biggest two things for us. [00:41:40] Noah Levin: Alex Cohen. Yes. Thanks for being on Serious People. [00:41:43] Alex Cohen: Yeah, thanks for having me. [00:41:44] Noah Levin: Thank you for joining us for the Serious People podcast. If you like what you heard, subscribe on YouTube, leave us a review on iTunes, or sign up for our newsletter at seriouspeople.ai/podcast. When we're not podcasting, Serious People helps businesses put AI to work in their daily operations. Visit us at seriouspeople.ai to learn more. The Serious People podcast is sponsored by Valet.dev. Go try building your own agent today at valet.dev or click the link in the show notes. See you next week.