Episode 20: Guardrails for the AI Frontier - Innovating Without Betting Your License with Travis Garland

Erik: Welcome to another episode of the Root Cause Business of Medicine podcast. Erik: And today it's going to be an interesting episode. Erik: It's a little more techie Erik: than we normally Erik: go, but we're going to delve into a conversation with Travis Garland Erik: , Erik: who. Erik: is from Erik: My Compliance Erik: Citadel.

Erik: And we had a fascinating conversation at kind of this intersection of healthcare innovation, compliance, AI. Erik: and Erik: even some of the safeguards that kind of are put into place. Erik: What were some of your big takeaways today, Davin? Davin: Well, I think it's great to know that there are people like Travis and his company that are out there putting together platforms and technology.

Davin: to Davin: enable us to use AI safely, right? Davin: Like they're Davin: they're um Davin: as you'll learn today, they're Davin: uh Davin: technology helps protect us, puts guardrails in place so that you can sort of determine how Davin: to use AI in a in an organization, even Davin: even a small one. Davin: um, you know, safely kind of push the boundaries, but not necessarily have it Davin: uh, you know, run the risk of breaking any policies or other Davin: regulatory um Davin: requirements.

Erik: Yeah, that was the same Erik: same Erik: concept that I I really enjoyed in today's conversation with Travis Erik: was the ability to Erik: Get excited about how AI can really enhance what we're doing Erik: from a system Erik: standpoint in a practice Erik: and gathering data and analyzing that data and Erik: giving that information to patients. Erik: But do it in a way that we aren't putting information out into a system Erik: or Erik: creating situations Erik: that can compromise Erik: Patient identity, patient Erik: privacy.

Erik: Um, so it's a it it was an important discussion today, and I look forward to you guys listening in and and seeing what you get out of our conversation. Erik: Welcome to the Root Cause Business of Medicine podcast, where we explore what's broken in healthcare and what we can do about it. Erik: I'm Dr. Erik: Erik Lundquist Erik: and I've been practicing functional medicine for the past 15 to 20 years.

Erik: I'm excited to co-host this podcast with my brother, Dr. Davin: Davenlundquist, who's just beginning his journey into functional medicine. Davin: We come from different points on the path, but we do share a common goal. Davin: We want to rethink how Davin: Medicine is practiced and helped others do the same.

Erik: The US healthcare system is in crisis, rising costs, declining outcomes, and physician burnout at an Erik: All-time high. Erik: But you know, we found a different way. Erik: Another way. Erik: A better way.

Davin: On this podcast, we dive into real stories from medical professionals who've stepped away from the traditional model. Davin: Kinda Davin: like me Davin: And have found a new purpose in integrative, functional, and alternative approaches to care. Erik: These are authentic conversations with practitioners and friends who redefine success Erik: not just for themselves. Erik: but Erik: for their patients and communities.

Davin: Whether you're a clinician feeling stuck, a student seeking direction, or just curious about what is possible, Erik: You're in the right place. Erik: This is Erik: the Root Cause Business of Medicine Podcast. Erik: Welcome to another episode of the Root Cause Business of Medicine podcast. Erik: And today we're going to veer into Erik: a new frontier Erik: in terms of our podcast.

Erik: We're gonna get into Erik: artificial intelligence and its Erik: impact Erik: on the business of medicine. Erik: And we have a special guest today. Erik: His name is Travis Garland Erik: from My Compliance Erik: Citadel. Erik: And we're gonna talk all things AI today.

Erik: In fact, Gavin, why don't you give us a little more of a background on kind of the impact that AI is having in medicine in your world and kind of some of the things you're doing and then we'll bring Travis in and have him tell his story. Erik: What what's been Erik: what's been the impact of AI Erik: on the landscape that you've seen and how are you utilizing AI? Erik: And then we'll get Travis to introduce himself and Erik: a little Erik: more. Davin: Yeah, well, you know, I think Davin: with my background as a CMIO, um, I always was looking for ways to, you know, Davin: find technology that would solve the problems as I saw them.

Davin: And Davin: I've noticed that Davin: even Davin: in this day and age, you know, the the electronic medical records Davin: ha Davin: while they've tried to to do this, I don't feel like that they've really Davin: gone far enough in terms of creating enough uh Davin: capability for people to really design their own workflow um and kind of figure out how they might want to um Davin: you know Davin: create Davin: uh Davin: tools um Davin: to to manage a practice. Davin: And Davin: um Davin: part of this is Davin: I think Davin: driven by the insurance companies.

Davin: and kind of what they need from doctors in terms Davin: so the EMRs Davin: have been built primarily to service the Davin: insurance industry, right? Davin: Um Davin: and Davin: and so then as physicians Davin: you sort of have to accommodate your workflow around that, right? Davin: And you know, your note Davin: has to be something that's billable, et cetera. Davin: Right.

Davin: And they've made it a little better over the last few years, a little more flexible. Davin: Um, but what I noticed as I Davin: um, you know, as my practice evolved in a non-insurance model where I Davin: Nobody needed my note in order to see how much the visit was worth, right? Davin: Um, and so then the note really just became a way for me Davin: personally to track Davin: what I was doing Davin: what I was doing with the patient. Davin: And again Davin: , Davin: when I think about my practice Davin: where I have a membership-based model.

Davin: And I'm working longitudinally with these patients, right? Davin: It's not about a particular encounter. Davin: It's not Davin: what did I do during an encounter? Davin: It's more like Davin: what Davin: are this Davin: what are these patients' goals, right?

Davin: Like Davin: I assess them holistically. Davin: Um Davin: I, you know, I have sort of a systems-based um Davin: scoring, if you will, of the patient. Davin: Um Davin: and then Davin: Um, we have interventions that Davin: hopefully Davin: begin to optimize each area of Davin: of their health. Davin: Um Davin: and so, you know, tracking those interventions, tracking their data that we're using to to see how well they're scoring and performing.

Davin: um, you know, becoming aware of symptoms or challenges, et cetera. Davin: Um, and so Davin: I just completely disconnected from the traditional kind of EMR model and and kind of soap Davin: note, right? Davin: And realize I wanted to create my own. Davin: And so Davin: that's where AI has Davin: has been a huge help Davin: because Davin: um Davin: it's been Davin: it's helped me sort of iterate on this process of sort of creating my own Davin: uh Davin: system Davin: and and so today I have Davin: I I use a database um Davin: called NAC Davin: that Davin: you know Davin: I have a a BAA Davin: with them a HIPAA Davin: secure arrangement Davin: So I've I've sort of begun to create my own EMR of Davin: sorts.

Davin: And Davin: um Davin: and then Davin: using, you know, I'll de-identify patients Davin: information. Davin: and leverage AI to help me sort of frame and and you know Davin: put together those notes and then I link them inside of my box. Davin: HIPAA Davin: secure environment and NAC, which kind of organizes the data in a way that's a little bit more kind of like an EMR. Davin: Um, I've been able to customize completely my entire system in the way that I want it to function Davin: um Davin: including, you know, setting up reminders and tasks and other things that I that I want in that um Davin: that are Davin: that are built on my system and the way that I interact with my patients rather than being dependent on Davin: um Davin: an EMR that's designed for for insurance billing.

Davin: So Davin: that's kind of where I'm at. Davin: And uh and Davin: uh Davin: without Davin: you know, going into too too much detail Davin: because believe me, I could I could, you know, speak about this for hours probably and geek out. Davin: Um Davin: but yeah, Davin: now Davin: with that Davin: as some context, Travis, I'm excited to hear your story and kind of Davin: where you've come from um Davin: and and and and how you've ended up kind of where you are. Davin: Oh Davin: uh Davin: so Davin: you you said a lot Travis: frankly Travis: Um Travis: there's there probably is like a whole series of podcasts and what you just said, Sally, for sure.

Travis: Um Travis: I I'm I'm Travis: before I jump into that, I would say I'm Travis: I'm really intrigued by two sides of this particular puzzle, which is Travis: From uh Travis: my time Travis: with uh Travis: with United Healthcare on the insurer side, my time Travis: with um uh Travis: UF Health Travis: Academic Medical Center. Travis: My more recent time as the founder of Kix Travis: in AI, where we were building agents and Travis: AI workflows for organizations Travis: to my current role as the co-founder and COO of My Compliance Citadel.

Travis: I very much so see Travis: the idea of one Travis: being innovative, bleeding edge of kind of stuff. Travis: I'm gonna build whatever is necessary. Travis: So Travis: when you say Travis: Davin Travis: I'm building my EHR. Travis: , Travis: I'm I'm developing my own AI capabilities.

Travis: My workflow is solid because it's custom to what I've done. Travis: That is music to my ears. Travis: I absolutely love it. Travis: I love the pressure it puts on the system, the disruption that's there.

Travis: And Travis: the other side of me is the sort of compliance side. Travis: It's the side of, oh boy, like what are we, what are we getting ourselves into with this custom stuff? Travis: And uh Travis: maybe to get a little bit back into the the storyline, the idea of spending Travis: a fair chunk of years in the space of a Travis: Um, let's ensure Travis: that we are working with the major players of the world, coming from Travis: largest global insurer, health insurer in the world.

Travis: having conversations with EHRs was a very common kind of thing. Travis: We did it often. Travis: We were partners with them. Travis: And frankly, the idea that they were rigid Travis: silos Travis: sitting out there was the mainstay for a long time.

Travis: They allowed smaller groups to kind of interject their respective pieces into and plug into the EHR. Travis: And Travis: an interesting tide started to happen, which was Travis: As the wave of AI really began to infiltrate all things healthcare, those holes Travis: within the EHRs and EMRs started to plug up Travis: intentionally from the inside out. Travis: There was a sense of what are these large silos Travis: going to be doing? Travis: How are they going to play in the market?

Travis: And for a little bit of time, there was sort of, hey, just build stuff, plug it in, we're great. Travis: Then it was, Travis: you know, let's have some conversations and let's see what we can do. Travis: Months, months, months pass, and now you see Travis: these large EHRs Travis: Um Travis: you could you could certainly name plenty Travis: of Travis: half dozen off the top of your head Travis: that Travis: as you begin to look at them now, they are deploying their own AI capabilities, whether it be ambient transcription, which is probably the most Travis: familiar and common thing within healthcare right now, or it's other kinds of capabilities such as their own use of AI with respect to referrals and part Travis: authorizations, touching insurers.

Travis: they Travis: they've Travis: kind of closed up ship a little bit in that Travis: that particular space, which um is no pun intended certainly for the naval officers sitting here on the screen. Travis: But Travis: The intention here is Travis: these large organizations Travis: are Travis: they're seeing that we're all trying to make this stuff work. Travis: They've opened it up a little bit and now they're Travis: They're saying, no, thank you. Travis: Let's let's close up, shop a bit, and figure out what to do next.

Travis: So when you say things like, uh, Davin, we've built our own custom stuff, we are in a very interesting space of being able to say, Travis: Oof Travis: boy, organizations can truly build their own stuff. Travis: Like Travis: we do not need to be in the space now where those are the only games uh in in the city to go after. Travis: So Travis: Um, I did not do a very good job of introducing myself uh in in that, but really just kind of hitting on those highlights that you laid out there at David in front of me.

Travis: The Travis: the Travis: my time with United Healthcare and with UF Health Travis: was preceded by my time in in spaces Travis: like foster care and adoption organizations where I oversaw utilization. Travis: And that's really where I I cut my teeth in Travis: operation space Travis: very heavy into data analytics Travis: in those spaces where Travis: um Travis: we Travis: we actually wanted Travis: a Prudential Davis Productivity Award Travis: for redesigning the foster care Travis: system in the state of Florida.

Travis: And that was something that we were very proud of, part of a a committee that put that together and that goes back more than a couple decades. Travis: But that really kind of created that momentum in me of Travis: how do we make systems that work really well for people? Travis: The foster care system and childwear Travis: generally isn't a space Travis: that gets a ton of attention Travis: financially or Travis: operationally. Travis: Um, but there are passionate people in those spaces and I was proud to have been part of of that work.

Travis: And then fast forward through uh through UF Health, where um I was a part of their Travis: medical or um Medicaid help plan that they had. Travis: And that particular structure, which uh Travis: which Travis: really Travis: helped um Travis: because it was a a a provider-owned Travis: network. Travis: those providers not only Travis: uh Travis: were receiving the revenue and the the dollars associated with the the actual uh Travis: principal payments for the insurance, but they were actually also the ones providing the care.

Travis: So they had a really interesting tie Travis: into Travis: What I'm I might say is kind of the precursor to value-based contracting Travis: and Travis: do Travis: I dare say Travis: even you getting into things like uh you know functional medicine, the idea that we're truly Travis: doing things that are best for our patients Travis: because Travis: it's best for our patients Travis: and not for the other side of the table. Travis: And that's something that Travis: As we build out AI capabilities and as we hear things like, hey, I stood up my own thing, or when uh when Erik talks about the Travis: the Travis: glass in front of his face as he's doing different procedures that Travis: Yeah, that's some cutting edge stuff to try and integrate AI, uh maybe augmented reality kind of stuff into the the ten Travis: by ten Travis: space of the physician.

Travis: But that's the kind of stuff that Travis: I love, love to press forward in and Travis: certainly Travis: I know we'll spend a good deal of time talking about. Travis: So a little bit about myself, give you a sense of Travis: the lens I'll be looking through during the next you know forty Travis: minutes or so. Travis: Um Travis: but thank you guys for welcoming onto the podcast. Erik: Travis, uh that's great.

Erik: Thanks for that Erik: that introduction Erik: and we're we're excited to talk about, Erik: explore this a little bit more today. Erik: Tell us a little bit about what Erik: is meant by healthcare compliance. Erik: Like why Erik: why did you set up a company Erik: That's geared towards that. Erik: And I'm guessing, you know, it says Erik: my compliance citadel, and Erik: the Citadel is always seen as a fortress, right?

Erik: A place of security. Erik: So maybe give us a little background on Erik: what Erik: what that means. Erik: Um Erik: you know, from a compliance standpoint, from a Erik: uh Erik: security standpoint, what you're seeing, what why Erik: why did you develop a company into this space and what are you trying to hope to accomplish with it? Erik: So Travis: a couple of quick nuggets to walk into this.

Travis: We know that in 2026, there's likely more than 90% Travis: of the organizations in healthcare plan to do something with respect to AI. Travis: That's that's not news. Travis: We also know that maybe a third or so actually have a plan. Travis: That's probably not uh Travis: great news either.

Travis: The interesting number for us is Travis: three and a half to five percent. Travis: That's the the percentage of organizations Travis: that actually understand Travis: have things in place. Travis: to govern or set up compliance standards for their automations and AI. Travis: That's a tiny percentage.

Travis: It's a tiny percentage. Travis: So Travis: our organization Travis: was really Travis: nestled into the space of how do we ensure that Travis: as the companies stand up and they say, hey, I want to build cool stuff, I want to really go after this, wonderful. Travis: You're on the front edge of those innovators, those three to five percent of folks. Travis: fast Travis: follows Travis: another twelve percent or so, right?

Travis: You guys know these percentages of what gets to that tipping point. Travis: We've now sort of gotten to that point where Travis: Organizations are now saying, I've sort of released something I wasn't quite sure about, which is I wanted my staff Travis: people Travis: to really Travis: drive forward on this stuff, and they have. Travis: And now I have no idea what to do. Travis: They're out there building stuff.

Travis: They're using whatever, you know, GPT thing that I put in front of them based on my company, whether it be Travis: uh Travis: a a clawed version, whether it be a a copilot if you use a Microsoft, doesn't really matter. Travis: In fact, I might even a really quick segue on this. Travis: I would say the vast majority of people actually use private GPTs. Travis: open AI's Chat Travis: GPT on their private devices to support their work internally to companies.

Travis: I'd love to know what that percentage is Travis: because that's where Travis: the space of a lot of that forward-looking innovation stuff happens and it's where Travis: the ripeness of compliance problems really comes from. Travis: People are trying to do stuff. Travis: If a company says Travis: yes, use it, great. Travis: They say no, well, guess what?

Travis: They're doing something to be as effective as they are. Travis: And people don't write 40-page briefs. Travis: um Travis: in an hour and a half, right? Travis: That doesn't happen.

Travis: I guarantee you they wrote it in five minutes Travis: with the chat GPT and Travis: sat around until the next meeting finished and then hit Travis: send. Travis: They may have even scheduled the Travis: sends, you know, it's sort of Travis: who knows. Travis: But Travis: the reality for us is Travis: our company Travis: understood that space, that there was a a Travis: a deluge of stuff coming our way. Travis: And we said, how do we make sure that we want to Travis: uh Travis: protect human data from badly behaving AI agents?

Travis: Plain and simple. Travis: AI agents Travis: are not Travis: ambient Travis: transcription. Travis: They are not the field that you type into on ChatGPT. Travis: They are not the kind of things that Travis: help schedule your emails that go out.

Travis: Right. Travis: AI agents and a Gentec AI, if Travis: I want to extend it just a tad Travis: bib. Travis: I might say Travis: they are things that Travis: within the guardrails that you give them Travis: provide an automation Travis: of Travis: uh Travis: forward thinking Travis: of capabilities to act on their own decision Travis: making. Travis: They have the ability to scrape data that's useful for them to achieve their standards, whatever that standard is.

Travis: So Travis: when Travis: we Travis: started walking into the space, there was a very clear position being taken of Travis: we've got to make sure that as people stand these things up. Travis: educationally they have to know what an agent is, right? Travis: Second, they have to understand how it can behave poorly. Travis: That Travis: that badly behaving AI agent space Travis: is a scary space.

Travis: It's a scary space to be in. Travis: Um Travis: the most recent piece sitting out there in the news maybe a month ago now Travis: is Travis: uh Travis: an organization that does car rentals Travis: had their entire production data set erased. Travis: because Travis: of a cursor clawed agent that admitted to all the stuff it had done wrong after the fact and said, hey, you know, I I appreciate that I did this. Travis: I was doing my best.

Travis: Here's all the things that I skipped over and assumptions I made. Travis: Meanwhile, the company's sitting there with an empty data set, right? Travis: Um, that Travis: scope Travis: is Travis: where Travis: my Travis: compliance citadel states. Travis: So when we talk about the fortress, I I don't know if I could describe it any better, Erik, so thank you for for describing it as a fortress.

Travis: Because it is the intended space for us to be able to say, when you think about what this fortress is, it's a it's a layer, it's a shield. Travis: Then Travis: make Travis: sure that Travis: as you Travis: do actual agent stuff Travis: It doesn't leave the boundaries of your particular organization. Travis: So Travis: if you have an agent that does something quite interesting, like, hey, go out and give me Travis: um Travis: the most recent articles on a cancer treatment that's sitting out there.

Travis: Aging goes out, scrapes a bunch of stuff together and it that's what it's intended to be. Travis: Drops summaries down to your providers. Travis: That way they can stay up to speed on the most recent kind of of research and in analytics. Travis: Very reasonable agent, right?

Travis: It's not doing anything besides that. Travis: What it will Travis: what it could do is it could go through and say, well, what's your current understanding within your organization? Travis: Well, my understanding is X, Y, and Z. Travis: Great.

Travis: Let me grab that and make sure I understand it. Travis: It starts to grab potential research that you've done in-house, or maybe you're doing clinical trials. Travis: Maybe it grabs a couple of Travis: patient records that aren't sitting inside of an EHR. Travis: Maybe you stored them on a Word document.

Travis: that's sitting in your system. Travis: Maybe you have them in Excel workbook, but it cures that together because it needs to know the best information to see what's out there. Travis: So it grabs it, goes out and says, based on this knowledge, give me some of the new stuff. Travis: And it comes back and it Travis: Tells you also it's a new stuff that isn't what you already have in your system.

Travis: When that happens, it's just taken all of your stuff and walked out the door with it. Travis: for sake of what it thought was the best thing. Travis: I'm gonna give you the best answer possible, but I gotta know what I know in order to give you the best answer possible. Travis: So Travis: it does those kinds of things.

Travis: It can do those sorts of things. Travis: So Travis: in the space of that, very inadvertent, right? Travis: Our organization puts that screen up and it says three things, three vertex Travis: you get. Travis: One Travis: Pass it through the wall.

Travis: It's fine. Travis: You get to put the jaw bridge down, you get to walk across the moat. Travis: Perfect. Travis: Second Travis: I'm not entirely sure.

Travis: So how about you just do the password, you know, knock on the door, we'll do an escalation up to a human in the loop situation and we'll make sure that the that the chief medical officer or the compliance officer sees it, understands it, says, Yeah, that's cool. Travis: Let's Travis: roll. Travis: The other side is a full prevention, right? Travis: So either pass, pause, or prevent.

Travis: And that prevent is what says, Travis: wait a minute. Travis: We don't want you guys doing any of this stuff. Travis: Don't release that information. Travis: That's a no-no.

Travis: Um, and then the most important piece of this, which is really where the compliance comes in, is that it drops everything down into a a tamper-proof audit chain. Travis: Similar to a blockchain, you know, it's a it's a secure sequence of events that happens. Travis: It is tamper-proof. Travis: You can only Travis: append to it or amend to it.

Travis: You cannot adjust it in any way Travis: And Travis: that drop Travis: down Travis: to that audit chain is what allows us to at the the click of a button Travis: send that out to an auditor or Travis: an investor or your board to say, hey, this is what happened with this particular agent. Travis: That verifiability of what the actions it took really Travis: is. Travis: is the difference between uh uh Travis: somebody out here who's just logging events in an Excel workbook Travis: and an agent that actively enforces what happens when you click the button.

Travis: And when it does that, it releases those three verdicts, one of the three, and drops it to an audit-proof chain and says, sorry, you don't get to pass the moat at all. Travis: You're out. Travis: Um, so our organization was built on those premise Travis: that we're running fast Travis: and we don't know where we're running sometimes, and sometimes we're faster than we should be. Travis: And Travis: we've got to protect ourselves to make sure that human data isn't just floating about Travis: unbeknownst to us.

Erik: Wow. Erik: That's a lot. Erik: I'm just gonna break down a few things Erik: just because I want to make sure our listeners understand. Erik: So when you say agent Erik: What you're referring to Erik: is Erik: a communication portal, right?

Erik: So if I'm using chat, GPT Erik: That would be my agent. Erik: If I'm using Claude Erik: and I'm accessing that portal with Claude, that would be my agent. Erik: Am I understanding that correctly? Erik: Um Erik: those would be Travis: um Travis: engines, I would say, to to Travis: to propel your request out into the world out to a large language model that contains the brain that you're trying to tap into.

Travis: um Travis: the agent Travis: would be in Travis: in in this context Travis: an organizational Travis: system Travis: that you've told Travis: When something happens, I want you to go and do this. Travis: You can use your own decision making as you so choose, but it's just a system Travis: that Travis: is brought together to be able to act on your accord Travis: for a specific mission. Travis: And sometimes that's research like I was referencing before. Travis: Sometimes it's rescheduling patients, like maybe a Travis: voice agent that's sitting out there as at a receptionist desk that can go in and look at your calendar and do different things.

Travis: So it's not Travis: it's not the field in a in a chat GPT Travis: text where you would type in a prompt. Travis: The prompt itself is an ask. Travis: It's just a, hey, let me see what I can get out of you Travis: And Travis: that Travis: um, Erik, is a really Travis: important space that you're talking about Travis: because Travis: oftentimes people say, well, how about if I just tell it not to send my sensitive information, right? Travis: The prompt is the gateway Travis: to the large language model that's behind it.

Travis: So why can't I just tell it to not do it? Travis: All the research Travis: it Travis: it doesn't really depend on Travis: Whether or not you tell it or not. Travis: It'll Travis: it will act in its Travis: best interest Travis: about 60% Travis: of the time following what you told it to do. Travis: generally, right?

Travis: Some are better than others. Travis: But what that amounts to is essentially like a pinky swear with a robot, right? Travis: You you Travis: probably Travis: aren't going to base your entire business on a pinky swear, right? Travis: I Travis: know I did with my best friend when I was seven, but that changes things now.

Travis: So Travis: Prompts are, you know, call it fifty, sixty percent of the way effective. Travis: Um Travis: what you really need is a s is Travis: a system Travis: that keeps those prompts from doing things that you can Travis: you can work behind. Travis: A good programmer Travis: will hack the heck out of that and will Travis: just convince Travis: through prompting to just w Travis: to work around something. Travis: You could tell an agent to go and do something.

Travis: But again, the prompt is before our filter. Travis: So you can try and tell it to do all sorts of stuff, but you really need something behind the prompt before it leaves the door and goes out to the large language model. Travis: Um Travis: and the agents Travis: are are really smart about working their way around those kinds of prompts because as soon as they're given the autonomy and the authority to do something. Travis: that's where we get ourselves into trouble is that they can go out and do that kind of stuff.

Davin: Yeah, I've noticed um Davin: in using, you know, some of these tools Davin: that Davin: um Davin: they've Davin: they start to develop Davin: um Davin: you know, your own rule set, if you will, right? Davin: I think Claude calls them Davin: skills. Davin: Um Davin: I think, you know, ch Davin: Chat Davin: GPT, you know, may have a different Davin: um Davin: uh Davin: name for it, you know, uh Davin: different G Davin: GPTs that you can create, you know, that kind of have like its Davin: own persona and kind of, you know, so you can kind of coach it and say, hey, this is what I want you to do, repeatable type tasks.

Davin: Yeah. Davin: Um Davin: but to your point, um Davin: I like there's been scenarios where Davin: um Davin: an agent is deployed Davin: Without me even asking for them to deploy the agent, like the two Davin: the Davin: the engine, if you will, deployed agents, you know, to to answer the question or whatever, right? Davin: And then Davin: the agents Davin: sometimes just decide Davin: we're not gonna follow your skill or we're not gonna even reference your skill or we're not gonna you know and then Davin: you're like, well why didn't you why didn't you do that?

Davin: You know, why didn't you reference my skill that's clearly there, you know, for this scenario. Davin: Oh Davin: well Davin: Because I send an agent, the agents don't usually foll follow it. Davin: Oh, well that's interesting. Davin: You know, so Davin: it is Davin: it is almost humorous, but not humorous Davin: if Davin: you're dealing with an organization and the risk to the to the organization, right?

Davin: So that that is very interesting. Davin: And Davin: Erik, I'm glad Davin: I think most of our users might Davin: Um Davin: I Davin: don't want to say most. Davin: I don't want to make any assumptions, but some of our users will think of it the way you did. Davin: Um Davin: because this is all new stuff.

Davin: I mean it's new territory, it's new terminology. Davin: And Davin: not all of the AI platforms or engines Davin: i even Davin: functions Davin: the same, you know, in the way that they interact with prompts or Davin: agents Davin: or other Davin: things, right? Davin: So Travis: you know it's it Travis: AI Travis: is uh Travis: it's Travis: filled with all sorts of Travis: new words that we just like make up sometimes. Travis: And sometimes you have to because it's a new I Travis: like the frontier, Erik, right?

Travis: It's a new frontier of stuff that's happening. Travis: Um Travis: and I I've yet to to be smart enough to come up with something that's so awesome that I get to name it. Travis: Like that hasn't existed yet. Travis: But I I keep waiting for that moment for myself.

Travis: Um Travis: what Travis: what I Travis: what I would say is you Travis: You know, in the in the space of you know Travis: deploying agents, just really quickly on that point, um Travis: agents can spawn other agents, right? Travis: So when you think of something as Travis: as basic as Travis: um Travis: Say you want to have your uh Travis: your front office desk Travis: staff Travis: know how to follow a procedure. Travis: And you type out a couple things into a GPT prompt and you say Travis: Be sure that they check X, Y, and Z when they're scheduling.

Travis: Be sure they introduce themselves this way and be sure they conclude the phone calls with this way. Travis: And Travis: when you do that, Travis: What comes back is way more information than what you ever told it in the prompt, right? Travis: We all are aware of this stuff. Travis: It can write books on content.

Travis: 14 Travis: pages. Travis: Yeah, exactly. Travis: Exactly. Travis: So it's wonderful on that.

Travis: Keep in mind that what you gave it was a very small amount of information. Travis: It just went out and did a bunch of stuff, right? Travis: So the idea Travis: that if you have an agent that knows even more about things, how much more capable it's going to be at saying, okay, I know exactly what you want. Travis: And I also know how to tie in ancillary Travis: content to that information.

Travis: You can give it personality and tone and directness and all sorts of those kinds of things. Travis: And Travis: to the extent that there is something called orchestration models now, right? Travis: Nothing new, but the fact that there is a supervisor Travis: of Travis: agents Travis: and orchestration agents Travis: that has been defined and built Travis: because agents do ridiculous things sometimes. Travis: They just Travis: they Travis: act badly, right?

Travis: So Travis: you have an orchestrator that says, hey, if you have fifteen agents and those fifteen agents do very specific things. Travis: You need to have somebody make sure that they're all doing what they're supposed to be doing and not getting out of control. Travis: So Travis: even in the concept of a basic prompt expanding to all this Travis: in an agent having to be supervised by another agent, Travis: you could see pretty quickly how this could get way out of control.

Travis: And that's that's where we find ourselves quite a bit. Travis: Um Travis: in Travis: in the the healthcare space Travis: I Travis: mentioned s Travis: ambient transcription is probably the most familiar, but it Travis: really Travis: I Travis: 'd say Travis: sort of to gain its traction in things like Travis: Yeah, medical imaging was probably one of the first areas where AI really started to get into Travis: the space of helping Travis: s at least visually Travis: better understand Travis: um Travis: what was being seen in a CT or an MRI or an X-ray.

Travis: Um Travis: those kinds of things Travis: were pretty prominent. Travis: Um Travis: clinical decision making kinds of stuff is is coming to be a big part. Travis: Administrative functions, I mentioned the front desk Travis: a huge part that's probably scaling um the quickest and probably the broadest Travis: because it's it feels Travis: safer Travis: um and you don't maybe need the technology of some of that that really deep modeling that's tied to the the imagery kind of content Travis: Uh Travis: but we see it in Travis: predictive Travis: analytics Travis: when it comes to deciding which of our panel members are the sickest.

Travis: And Travis: when those are the ones that are identified, how do we manage them, right? Travis: Do we go after them in a different way? Travis: In the insurance space, maybe value-based contracting, uh, but in other spaces, maybe we say, hey, we need to bring you in for an Travis: an you know Travis: extra three sessions of acupuncture and massage therapy or make sure we're looking after your nutrition in a way that Travis: that food as medicine becomes a a live and viable option here as we understand more and more about the science, those elements Travis: are Travis: part of AI.

Travis: And I think that's the really cool piece Travis: is that Travis: It Travis: 's no longer just in the hands of Travis: the large corporations who are harnessing Travis: eight, nine Travis: figure Travis: budgets to fund AI. Travis: Organizations now with relatively small budgets, let's say five, ten Travis: , fifteen thousand dollars Travis: can stand up incredible Travis: capabilities to run an organization and Davin, your Travis: your comments early on about Travis: your ability to to build these things and help your patients in a way that Travis: you know, ten Travis: years ago you'd probably be like, what the heck are you talking about, right?

Travis: Um Travis: it's very different now. Travis: It's it's Travis: um Travis: all because of of AI and the capabilities that are out there. Travis: It's really just about harnessing in the right way and protecting yourself Travis: So that uh you don't get to the point where you've got inadvertent Travis: problems, right? Travis: Those are really expensive problems.

Travis: Um Travis: I can tell you Travis: more about those stuff pieces, those those elements certainly, but um Travis: , Travis: for Travis: changing. Erik: I do want to Erik: I do want to get into a little bit of that, but before I still want to Erik: I want Erik: to Erik: Get a little more granular with some clarity aspect Erik: because Erik: I'm sure that there's others who maybe still are saying, I'm not getting this whole agent thing. Erik: Right. Erik: Now you just called superintendents.

Erik: And now it's like Erik: So how Erik: how Erik: how does one Erik: walk us through how one Erik: creates an agent? Erik: And Erik: you Erik: even said sometimes that there's agents that spawn agents or just uh Erik: GPTs will Erik: will Erik: just Erik: ca Erik: already direct an agent to to conduct uh Erik: a system. Erik: So maybe maybe help us uh Erik: because I, you know, I Erik: Yeah, I'm visualizing, I think a lot of us do, right? Erik: Star Trek and you have data, right?

Erik: You know, he's out there doing his thing and he's Erik: he's, you know, this this ultimate AI Erik: uh Erik: being. Erik: Um Erik: so I think I Erik: when you say agent, I'm immediately thinking, oh, Gata's out there doing some work for me or something, right? Erik: So Erik: tell Erik: maybe, maybe Erik: more specifically, let's get clarity Erik: about Erik: What's the Erik: what's the difference between an agent Erik: and the GPT in terms of if I'm just doing a prompt Erik: How how does it go from just me giving it a task to now there's an agent that's running a whole system Erik: Um, d Erik: then Erik: then is Erik: creating other agents and then is deciding that it doesn't want to even follow the prompt.

Erik: I I I Erik: I guess I'm I'm not and Erik: I'm not clearly understanding Erik: ha Erik: that whole process. Erik: And I think it's important Erik: because in order for us to have Erik: protection Erik: and safety around that Erik: Uh Erik: I think we need to understand what it Erik: what exactly these Erik: components are. Travis: Uh Travis: so the Travis: the difference between a GPT and an agent. Travis: A great place to start Travis: because the Travis: the prompting, we've been Travis: you know Travis: hit over the head with all sorts of things, even the commercials, like just prompt your way to your new website, right?

Travis: You know, that's out there. Travis: Just Travis: a few sentences and you can make your own, you know, b Travis: Hollywood movie, right? Travis: It Travis: does not work that way. Travis: I don't know if you've tried it, but it doesn't, right?

Travis: Um, but here's Travis: here's the difference between the two. Travis: Um Travis: we tend to Travis: uh Travis: express Travis: the best prompting within the scope of three things. Travis: what Travis: 's the intent that you want to accomplish, what's the context that you want to operate within, and Travis: what's the constraint Travis: that it needs to understand. Travis: And those three things together are really what Travis: create a really good prompt, right?

Travis: There's other frameworks, but that's one that I typically use Travis: And it is Travis: it is not a fancy Google, right? Travis: It is Travis: not a a mechanism to just like Travis: research, although GPTs are often used in that way. Travis: They are Travis: by nature Travis: generative models Travis: that tell you essentially what's the next best word and it puts all this context together and the constraints and Travis: operates Travis: within the scope of millions and millions, billions of data points and gives you an answer Travis: based off of those three things Travis: Uh Travis: an agent on the other hand Travis: is something that you provide the same kinds of uh Travis: content to, but you provide a whole lot of other things.

Travis: So if I were to s to define it with respect to the robustness Travis: A prompt might be two or three sentences. Travis: A good one might be, you know, five or six, might be a paragraph long. Travis: There's lots of ways to do this. Travis: But Travis: an agent might be built on, say, a five Travis: page prompt, if you would say Travis: if you can say it that way, because it defines Travis: the scope, the Travis: the authority that it has, the places it should go, the connections it should make.

Travis: And agents are are very prescriptive in the sense that you need to define one specific task. Travis: So whereas a prompt might be a search for information, an agent does one thing. Travis: You do not want it doing multiple things because it gets confused. Travis: It's sort of like Travis: um Travis: a five-year-old with a PhD.

Travis: Right. Travis: It it it will act like a child if you let it and it's so smart that if you put those two things together, you've got problems. Travis: Hence Travis: our organization. Travis: But if you start pulling these things together, you can get to a better place.

Travis: So an agent Travis: functions with one task and one task only. Travis: And you define what those parameters are, what the guardrails are to that agent. Travis: So you tell it where to go, what to do, how to respond, on what frequency it should respond, how it should consider things, what tone it should use. Travis: and an agent Travis: functions within those guardrails.

Travis: Now, the interesting extension of that is the agentic Travis: AI space. Travis: That's the space where it truly has autonomy Travis: to not only do those things within those guardrails. Travis: But to then start making decisions about what to do next. Travis: And if you stop and say, here's what I'd like you to do, I'd like you to go Travis: and assess for all of the current records that are in my EHR Travis: and give me the top 10 patients that are going to the ED most frequently so I can reach out to them.

Travis: And ensure that we put them on a special white glove list. Travis: Right. Travis: It can go out and answer that within the garbers. Travis: The next level of that is that an agent has the capability to make a decision of Travis: What do I do next?

Travis: Right. Travis: And it can do things like, well, let me split out the information, put them in the special protocol batches, and then let me make a couple Travis: outreach phone calls or SMS messages to define Travis: um Travis: what they should be doing and Travis: how they should be outreaching back to the clinic. Travis: And let me make sure that I'm also assessing for the clinic's ROI on that outreach. Travis: How much marketing is is the clinic spending on Travis: developing relationships with those clients.

Travis: Are they membership paying clients or are they not membership? Travis: And it starts to ask questions and act on its decisions. Travis: Then you go backwards. Travis: Let me go back to the agents that Travis: acts in the guardrails Travis: that doesn't do that.

Travis: It just says, let me go and get this information. Travis: And then back one farther, the GPT prompt that you put in. Travis: that does nothing besides Travis: just give you an answer Travis: within the Travis: the Travis: intent, the context and the constraint of that prompt. Travis: And you could see now the bookends of Travis: here's an answer I'm giving you Travis: versus Travis: I'm now making decisions Travis: as an agent based on what's coming back to me.

Travis: And I'm gonna act until you tell me not to. Erik: Mm-hmm. Travis: And if you don't tell me to stop, I'm going to be that agent that just destroyed that company's data set in nine seconds. Travis: Because Travis: That's what I Travis: that's what I'm allowed to do.

Travis: And if you don't give it those guardrails, it'll continue to act. Travis: Well I'm not sure if I answered your question. Erik: Okay. Erik: Essentially an agent Erik: is a Erik: program Erik: that Erik: you Erik: create guardrails for Erik: but allow it to make Erik: decisions, right?

Erik: It is Erik: it is a decision-making program Erik: that you set Erik: up Erik: Where Erik: when you just do a prompt, it's a you're just getting an answer. Erik: You're asking a question, you're getting an answer. Erik: Where an agent is gonna be a program Erik: that has decision-making Erik: power. Erik: to Erik: provide Erik: a system Erik: of Erik: information Erik: um Erik: that you Erik: set guardrails Erik: to Erik: So I think that and Erik: that that then makes sense Erik: to accomplish some Davin: or to accomplish some task.

Erik: Right. Erik: Right. Davin: That you've given it Erik: But you've given it you've given it some parameters Erik: and it has decision-making ability to Erik: decide what Erik: information to pull Erik: and and how to Erik: organize it Erik: and present it. Erik: Um Erik: but in Erik: it it has some Erik: decision making Erik: and I think that's the key component, right?

Erik: The agent has decision Erik: making power. Erik: And Erik: where the where Erik: the Erik: GPT prompt Erik: is just going to answer your question. Erik: Did I understand that correctly? Travis: For our conversation, you nailed it.

Travis: But Travis: um Travis: the Travis: the decision making piece, uh Travis: sort of the Travis: the swing into the compliance and governance, the protection spot. Travis: If you Travis: if you give your Travis: your agents Travis: the capability and authority to do something and it does, you have to be able to answer to that. Travis: Right? Travis: You are the deploying organization of that agent.

Travis: It's not the developer. Travis: It's not the architect of the agent. Travis: So Travis: if a client Travis: deploys an agent Travis: And the agent does something. Travis: The regulators are going to come back to the deploying organization and say, hey, what do you guys think about this?

Travis: Prove to me that what it did Travis: Not what the policy said it was supposed to do, right? Travis: Not what the SOP or the governance Travis: standard is. Travis: You tell me what it did and how did it make its decision, then unless an organization can do that Travis: You're going to be in trouble, right? Travis: And that's where the rules and laws and in Travis: our standards are going.

Travis: And here in the US we have Travis: Um Travis: we have no AI act that's in play. Travis: We have governance standards. Travis: We have compliance standards. Travis: Um Travis: other Travis: organizations Travis: do.

Travis: Um Travis: the EU Travis: is rolling out with a a relatively strong one. Travis: China has one. Travis: But when we think about how we define that through legislation, there's a Travis: there's a very clear position taken by the current administration, which is unique to previous administrations Travis: that Travis: we should be able to just sort of all Travis: allow the innovation to occur. Travis: Prior administrations had on the books very different kinds of standards.

Travis: And Travis: as the administration's changed, it's shifted the Travis: the space of AI and how we define it Travis: and Travis: organizations that Travis: are Travis: pushing for the definitions to have consistent definitions, um, are Travis: really Travis: trying to say Travis: this shouldn't be a Travis: an agent defining itself. Travis: It shouldn't be independent. Travis: I shouldn't be able to just ask you Travis: Erik, Davin, did you guys actually follow the rules?

Travis: And you say yes. Travis: And I say, okay. Travis: Right. Travis: That's Travis: that's a that's a Travis: that's a self-verification, right?

Travis: It means really nothing to auditors Travis: Second, independent. Travis: It's Travis: hey Travis: Travis, did Erik and David, Travis: sorry, Davin, release this agent Travis: and did it do what I was supposed to do? Travis: And I might say, let me look at the books, lift up the hood, check it. Travis: Cool Travis: Then there's a cryptographic.

Travis: And outside of, you know, being sitting in front of people who really want to know that kind of stuff, that's just the highest level. Travis: It's not even fully defined. Travis: It's most certainly not within Travis: the Travis: true legislation, you know, executive powers of our governments right now who would say this is how it's defined. Travis: But you can think of things like Travis: It has to be interoperable, it has to fit regardless of the system, it has to tell you every single step, it has to be independently verifiable by machines, not by humans.

Travis: All these different standards that go into this kind of standard of governance Travis: That's the kind of AI that we're moving toward. Travis: And Travis: we're gonna get there quick. Travis: Um Travis: I I do feel like there's some administration stuff that's going to push it one way or the other, especially as we move into the next couple of years, but Travis: You know, i Travis: the AI space, especially in healthcare, because it's so heavily regulated, it is a space to absolutely be Travis: um Travis: aware of the fact that just because there's not an AI act, a single thing that tells you Travis: Don't make the assumption that you're safe, right?

Travis: Because the EEOC has their thing, right? Travis: FTC has their thing. Travis: Every Travis: every Travis: lettered governmental agency has Travis: their standard that they're living by right now for AI. Travis: And depending on which one of them you want to talk to, the definition changes.

Travis: And there's no unified thing. Travis: So that's where we're we're sort of in our space right now. Travis: The big takeaway, and then I'll stop because I'm on a I'm on a Travis: I'm on a Travis: bad pedestal right now or soapbox. Travis: Um Travis: be careful, right?

Travis: It's a it's a moment of me saying on a risk side Travis: Do it, be innovative, press forward, incredibly powerful, use it to your heart's content, but don't Travis: don't risk the point that because there's no one single law that tells you not to, that that's the way that it's gonna be Travis: Build it with uh a conservativeness into your system Travis: and always have a way to ensure that you can track what's being done and if nothing else, ensure that you can stop stuff before it goes Travis: if it contains Travis: PII protect Travis: or personal identifiable information or protected health information.

Travis: Um, and that's that's the name of the game right now. Travis: Just push, but uh Travis: make sure that you Travis: got some protection. Davin: That's good. Davin: I Davin: I really like that.

Davin: I think um, you know, i Davin: innovation Davin: uh Davin: is is so amazing. Davin: You know, I think in healthcare Davin: for years Davin: uh Davin: this was the challenge though, right? Davin: Is the Davin: um Davin: the privacy, the security, the the level of Davin: um Davin: uh Davin: you know, the kind of data that you're dealing with, right? Davin: It's not a it's not a dinner reservation, you know.

Davin: So Davin: um, you know, your ability to schedule that, you know, is Davin: is so different, right? Davin: And I think that's why it's taken so long for technology Davin: that we see everywhere else in our in our ecosystem Davin: to hit healthcare, right? Davin: Because of the complexity and the sensitive nature of of the data, right? Davin: And then also Davin: all these, you know, systems that are disparate Davin: and use different standards and and and whatnot Davin: as well, right?

Davin: Um Davin: and so Davin: uh Davin: yeah Davin: I think it's it's it's an Davin: it's gonna be interesting to kind of see Davin: how Davin: AI, um, also, you know, i Davin: interacts with with healthcare. Davin: And to your point, there's a lot of people using it Davin: not Davin: officially Davin: under Davin: the guise of of a setup like yours, right? Davin: Um Davin: and and they're doing it innocently, not with bad intentions, but Davin: you know, to try to do the right thing, to try to Davin: be more effective at their job to help more people, right?

Davin: I mean, I think that's been my experience in healthcare Davin: is you have lots of people with great intentions, um, but their solutions don't all Davin: uh Davin: aren't always thought through in a way Davin: um, you know, Davin: that aligns with Davin: those good intentions. Davin: Yeah. Davin: And there's unintended consequences that happen in other things. Davin: And so Davin: um Davin: if people Davin: wanted to, I mean, your Davin: company Davin: um Davin: sounds, you know, like a a really good option for people.

Davin: Is it designed for large enterprises? Davin: Is it designed for Davin: individuals, business, small businesses, small practices. Davin: Yeah, help us understand that. Davin: And then Davin: um Davin: and then Davin: if it's an option for our, you know, users Davin: uh Davin: viewers to to take advantage of, you know, make sure we know Davin: they know how to do that.

Erik: Yeah. Erik: And then I wanna I wanna give you two scenarios Erik: because Erik: you share Erik: answer Davin's question, but I want to give you two scenarios Erik: then Erik: And maybe then you can tell us how, you know, whether or not those are safe Erik: exercises, activities Erik: And w Erik: and how Erik: what you guys are doing would maybe help protect against bad outcomes within those scenarios. Erik: Okay. Travis: Okay.

Travis: Uh Travis: so Travis: to hit on your your question, Devin Travis: Um Travis: our Travis: primary focus is Travis: really smaller organizations. Travis: When I say smaller, I mean like Travis: less than 500 staff. Travis: Um Travis: we Travis: the idea that we're not having conversations with Travis: um you know, the Travis: the Mount Sinai's of the world. Travis: The Travis: the reason behind that Travis: is because most of those large organizations are throwing their millions of dollars at it Travis: Right.

Travis: They Travis: they Travis: have teams to do these kinds of things and they're thinking about it and they're protecting themselves with Travis: uh Travis: boardrooms of attorneys and compliance officers. Travis: where Travis: we're at is there are a lot of things that happen in smaller organizations, clinical practices Travis: that need Travis: that support. Travis: That should be just as protected Travis: at a price point that makes a lot of sense for them. Travis: And that's an important space for op Travis: for us to operate in.

Travis: Our Travis: Our tools are agnostic. Travis: They can Travis: uh Travis: they Travis: can allow for hundreds of agents to be bounced up against it at any given time, with very little lag, if any at all, talking 0. Travis: 06 milliseconds, like tiny amounts of time Travis: Nobody would experience that slowness. Travis: That Travis: space, because of its ability to manage it, it can be deployed in the largest of organizations.

Travis: But that's that's not our focus. Travis: If a health system said, hey Travis, we'd love to use you guys. Travis: Okay, that's great. Travis: Our priority is Travis: organizations that we know are pushing forward, striving to be innovative, continue to stay up with what's currently occurring, and make no mistake about it.

Travis: Every single person that either uses and stores your medical records or pays you because of the work that you do Travis: is using AI. Travis: for their purposes, right? Travis: So Travis: remembering that there's a very clear other side of the table here that that Travis: as smaller organizations are seeking to to stay up and stay protected. Travis: They also have to be aware of the fact that the other side is using that information in ways that support their missions Travis: and it may not always follow yours.

Travis: You need some capabilities to ensure that you have agents to act quickly, act autonomously Travis: And you need to be able to have that same space be protected Travis: as it does that. Travis: So Travis: our focus is on the smaller groups, just to give you that. Travis: I feel like I'm ready for the the Jeopardy round here. Travis: I'm ready for your examples.

Travis: Let's see how we do. Erik: All right. Erik: So here's the two scenarios. Erik: So let's Erik: the scenario Erik: the first scenario would be a provider Erik: who is interested in getting some Erik: a different perspective, maybe a little more information about a patient they're caring for.

Erik: They're in the visit with the patient. Erik: and Erik: they then Erik: put information Erik: of that patient into, say, let's say Erik: chat Erik: GTP. Erik: Um Erik: but it's Erik: it's not protected. Erik: You're just putting the information in there.

Erik: Um Erik: and Erik: and let's say they upload Erik: uh Erik: labs into that because they want to get s a different perspective on how to interpret the labs. Erik: Um Erik: is Erik: that Erik: a HIPAA Erik: violation? Erik: Is Erik: that Erik: a Erik: a Erik: is Erik: what ends up happening with that Erik: information Erik: and who has Erik: who has the capability of accessing that through a Erik: chat GPT? Travis: So Travis: uh Travis: yes to the first part, which is Travis: is it a violation?

Travis: Yes. Travis: Um Travis: it is not a fineable violation until somebody says that it's a violation, right? Travis: So Travis: I can say yes, but you know, if no one knows, no one's the wiser, if I can put it that way. Travis: Yes.

Travis: Um Travis: the idea that Travis: um Travis: content Travis: that gets into that space Travis: and Travis: um Travis: needs to be accessible Travis: or how could it be accessible? Travis: I think was sort of the the tail end of the conversation or the question. Travis: Once it gets there, who can access that content? Travis: It's not a space that any of us can just tap into the GPT model and get it.

Travis: But here's the interesting part of this. Travis: One, um, if you're using a GPT, Travis: uh any Travis: GPT that's out there and you're just using it in the open space, you don't have a membership or a you know Travis: paid subscription for a pro Travis: model that's a little more secure, you just you know Travis: Chat Travis: GPT. Travis: com anywhere as you Travis: will. Travis: That is pure open source.

Travis: Which means the Travis: the Travis: actual Travis: content that's submitted to the large language models of that engine Travis: is usable to train the engine itself. Travis: Now, in and of itself doesn't sound like a bad thing, but what happens if you start continuing to do that and you do that over and over again? Travis: Nothing's really happening until you start actually asking questions of it. Travis: You're training the model to do very specific things.

Travis: And if it understands that you keep doing this, as it starts to respond, it's going to understand that you may be thinking more about that. Travis: It creates things like biases Travis: and its Travis: responses. Travis: It can be used publicly to find stuff. Travis: And that's where Travis: if a really smart individual were to start accessing Travis: information, you can query Travis: prompt Travis: a GPT from outside of your clinic, knowing what your clinic does, the name of your clinic, potentially even patients, right?

Travis: And Travis: can ask or prompt GPT to query content that's in the trained model. Travis: It can actually, through prompting only, pull content out, essentially coax the GPT responses to present information back that is absolutely PHI or PII from stuff that you've loaded up. Travis: So Travis: because that content goes into a public space, what has to come out eventually is that same data. Travis: It's mixed in with all sorts of stuff, but Travis: There's Travis: because Travis: you can, you know, develop agents that will go out and Travis: call the data and scrape stuff and query over and over and over again, there's ways to find that information if Travis: somebody's really motivated to do that.

Travis: So Travis: yes, it's a hippo Travis: violation. Travis: No, you don't want to load up things. Travis: Can it impact the biasness of the response? Travis: Yes.

Travis: Last. Travis: the individuals who can Travis: especially in open source content can go out there and find it. Travis: You can eventually get it to respond Travis: with data if you have enough ties. Travis: um Travis: or specificity about the data you want, you can definitely scrape that back out of the system.

Erik: So before I go into Erik: to scenario two, then Erik: would your Erik: does your company act as a filter to help Erik: Like Erik: if I wanted to do that, I want to use the Erik: the GPT Erik: and I let's say I ha I have a membership or I Erik: paid, you know, so it's a little more secure, but still i Erik: is there Erik: Is there a way to have a layer of protection so that I'm not getting myself into trouble just because I'm curious and I'm asking questions? Erik: I'm trying to get information.

Erik: I'm not really thinking. Erik: about Erik: the Erik: broad implications Erik: of what I am I'm putting into this database Travis: We have Travis: agents Travis: that Travis: enforce the point of action. Travis: So as soon as you hit the go button on Travis: that prompt Travis: or that agent going out to do something, the agent will go out to the Travis: to the large language model itself. Travis: It can bypass the the front-end Travis: prompting that you might have.

Travis: It's at that point that it will determine based on whatever agent we put in between Travis: there. Travis: Does this make Travis: does this matter to your SLPs? Travis: Does this matter to your policies? Travis: Does this matter to your internal documents?

Travis: That Travis: sort of company compliance is the first toll gate that we that we have at all agents. Travis: Second one is Travis: does it release PAI, PHI, does it violate high trust or HIPAA Travis: concerns Travis: As you get further into those different compliance packs, it absolutely stops it, renders Travis: one of those three verdicts Travis: of Travis: do I pass it, do I pause there, do I prevent it? Travis: And then drops into that auto chain to escalate it out to whatever regulator you might have.

Travis: So Travis: if somebody in this case were to say, hey Travis: agent, go out there and and Travis: each Monday Travis: go and find me all of this content, right? Travis: It will absolutely act at that point and say, nope, this Monday you guys tried to do something, it wasn't correct, or you had a front desk person try to load up all of this. Travis: um Travis: patient record contents. Travis: I Travis: they Travis: probably wouldn't do that.

Travis: Maybe Travis: load up a schedule that they printed. Travis: So they downloaded a schedule for the last month and said, I need to do a better job of organizing the schedule because of Travis: next month so many people are going to be out. Travis: Help me out with this GPT. Travis: In that scheduler, probably a bunch of patient names, right?

Travis: Maybe dates of birth, some other identifiable information to help link it up to medical records, possibly. Travis: Um Travis: if that were to be the case, those agents would say, mm-mm, like we gotta you gotta stop this. Travis: If there's content that it will Travis: that it can pass through, great. Travis: The content that it shouldn't Travis: It'll stop it a wholly, it'll uh redact it, Travis: scrape it out, or it'll tokenize it if we need to do some sort of other fancier stuff to make sure that the data can move into other spaces and be used.

Travis: But Travis: absolutely stops it. Travis: So Travis: that's Travis: that's the target Travis: you've got to pass Travis: is Travis: those agents that say, mm-mm, Travis: thank you Travis: Like a hall monitor, you know, of school, right? Travis: Either Travis: either you have the key with a two Travis: by four attached to it for the bathroom, right? Travis: That you have to get permission to walk in the hall with, or you don't.

Travis: So Travis: you either get to walk down the halls or you or you don't or you get pushed back and say, Hey, you better go get the, you know, the wooden stick with the bathroom key on it Travis: because uh you're not passing me. Travis: And that's Travis: That's the hall Travis: monitor that we'll that we're really talking about. Erik: Awesome. Erik: Okay, here's scenario two.

Erik: Okay. Erik: All right. Erik: So Erik: uh Erik: my clinic, we have multiple providers Erik: and a provider Erik: decides that they want to collect data Erik: um Erik: across the board. Erik: They want to know how many Erik: how many of our patients Erik: Their A one Erik: C's are greater than seven percent, and they want to know how many of those we have in our practice.

Erik: Can Erik: they use Erik: a GPT Erik: to access our EMR Erik: to be able to comb through the data and be able to find that within Erik: our Erik: system Erik: or is Erik: the EMR kind of protected against a GPT Erik: entering Erik: into that space Erik: and combing that information? Travis: Um Travis: so Travis: you can have it piped directly into your EMR. Travis: That Travis: space is getting tighter and tighter, as I mentioned at the beginning. Travis: They're sort of cutting those portholes off and saying, hey, we're doing this.

Travis: But Travis: where Travis: where Travis: they'll allow it, yes. Travis: Um Travis: I'll connect a comment, Devin, that you made a little bit earlier around the skills. Travis: So I'll reference Claude just to be specific. Travis: So Claude Travis: has skills Travis: It has cowork, which is one of the three areas, chat co-work and cloud code, that it can use to help make things easier for people.

Travis: If you, um, Erik Travis: said, um, I'm going to create a Travis: a skill Travis: that tells my system to go into my EHR and grab this particular data set Travis: and then throw it into Travis: a a GPT model of some sort. Travis: Those kinds of integrations are absolutely possible. Travis: They absolutely happen and they are very, very effective at what you do. Travis: So Travis: a as a Travis: maybe kind of stepping from healthcare for a second, if you just said Travis: I'd like to have my Travis: my GPT connect with my email Travis: and then connect to my Travis: um Travis: CRM so I can send emails directly out to my patients.

Travis: Those integrations that you see companies standing up and they say, hey, let me just connect all the pipes together, those are essentially the same kinds of things that skills do on an automated and a much bigger platform. Travis: So they just sort of say, hey, what do you want me to connect to? Travis: You want me to connect to your AHR? Travis: That's great.

Travis: Let me go get permission from uh Travis: Athena Travis: and Cerner Travis: and Travis: name all those epics, right? Travis: And then let me make sure I connect you up with my Hubspots or my custom CRM. Travis: And I'm going to go in, you know, email these folks, Travis: very Travis: possible. Travis: Now, the question whether it's protected is a different question.

Travis: Um, that's where you would say Travis: if things move between spaces, most of those pipes are as protected as they can be. Travis: But Travis: you have to get Travis: permission to allow it to happen. Travis: So if you allow for Travis: the connection between your EHR Travis: to the GPT Travis: to a SMS service that's going to send a text message to people. Travis: That Travis: connection is your Travis: responsibility Travis: and likely a violation of all sorts of rules Travis: because Travis: you might not have business associated agreements with those various organizations to protect Travis: the data that's that's moving Travis: between those systems.

Travis: So Travis: as soon as that happens and it moves into a space that you don't have a BAA Travis: with. Travis: Now all of a sudden you've released to them protected information that they don't have permission to have, even if it's Travis: something as benign as a name and an email to send us something out. Travis: So can you Travis: absolutely Travis: how Travis: how would our organization manage that? Travis: We would have agents that would sit inside of your Travis: skill Travis: That would say Travis: if you launch this, then here's as far as you're willing to go.

Travis: As soon as it comes back and touches a GPT Travis: We say Travis: stop Travis: at that moment. Travis: As soon as you develop a connection to another space, you may be able to get around it, right? Travis: So it's a Travis: it's Travis: it's important for us to be able to Travis: um Travis: be within what we call the runtime space, the activity that happens for your agents. Travis: And when you integrate our code into that Travis: Then it knows.

Travis: I can't go further than this because you're about to stop me. Travis: And the agents during the run of the process Travis: will say, uh Travis: oh, I Travis: I just hit something that's not going to be acceptable and that stops it. Travis: um Travis: drops it down to that tamper Travis: tamper proof audit chain and says, here's what's happening. Travis: So Travis: the cool thing is Travis: you can get, you know Travis: Carragon or a make.

Travis: com Travis: or a zapier Travis: to connect all sorts of pieces. Travis: Um Travis: the difference really quickly between a make. Travis: com and a zapier Travis: that's sitting Travis: or a carragon, excuse me. Travis: Carragon does have business associated agreements.

Travis: So Travis: it knows Travis: that when you want to connect things, it needs to have those agreements in play for all of its pathways. Travis: And it does a wonderful job of that. Travis: make Travis: more robust, but it doesn't have BAAs. Travis: Zapier Travis: definitely doesn't have BAAs in those spaces.

Travis: So when you think about the protected health information Travis: And you want to connect stuff, you just have to be aware of which organizations you're using to help connect the pieces, and then you're a little bit better positioned. Travis: But Travis: um Travis: Boy, that was a long answer to simply saying yes, you can connect it, but you're about to get into trouble if you're not careful. Travis: That's Travis: I keep going back to the same thing. Travis: There's always ways to protect yourself, but the cool thing about AI is that almost anything is possible.

Erik: Yeah, absolutely. Erik: Well, that was Erik: that was great, Travis. Erik: I think that was really helpful Erik: in Erik: we don't really Erik: often think Erik: about Erik: the impact that AI is having Erik: Sometimes in the background Erik: or, you know, in in our systems, we just Erik: think about, oh, EI makes my life easier because I can Erik: uh Erik: write a document quickly, I can search for some data points, I can Erik: I can Erik: access and put together Erik: collate something quickly.

Erik: Um Erik: much Erik: like uh you know the early printers and and fax machines allowed us to start communicating and organizing things in our office, I think we Erik: a AI is just doing that at a trillion times faster Erik: and Erik: uh Erik: but the complexity of it, the risk associated with it Erik: is a real Erik: uh uh Erik: entity that we have to really think about and and process. Erik: So thank you for coming on the podcast and sharing Erik: Does Davin Erik: any Erik: uh Erik: final Erik: questions before we close up today?

Davin: Well, I think you know Davin: something we ask uh a lot of our guests Davin: is Davin: um Davin: where they kind of see the future of Davin: of healthcare going, um Davin: kind of through their own lens, right? Davin: And um Davin: you know, if you if you were to sort of Davin: take your lens and and where healthcare's headed in terms of AI Davin: and and what it might do Davin: for Davin: uh Davin: for this Davin: even just the science of health or the technology of healthcare or the Davin: Delivery of healthcare, um Davin: any Davin: big predictions you want to make?

Travis: Um Travis: you know, I Travis: so I might have a I might have a a couple. Travis: Um, the first one is Travis: the Travis: agents Travis: are very good at giving you back choice, right? Travis: The choice is whether or not you want to use the time Travis: to Travis: take care of yourself Travis: to have a relaxing afternoon, to give your front staff a little bit of extra time, an hour off because they've gained back the time and their efficiency. Travis: Or if you want to use that choice to just fill it with more patience, more work to get done, make more revenue, whatever that might be.

Travis: So Travis: I I've stopped kind of saying AI replaces people Travis: because it doesn't always replace people. Travis: Sometimes it just gives you the choice. Travis: It always gives you the choice because it does create that capability. Travis: Um that's just a very blanket Travis: sort of AI Travis: comment.

Travis: Um Travis: the other space I would say Travis: um Travis: it Travis: it is going to Travis: advance clinical work in a way that um Travis: is Travis: sort of befuddling to me. Travis: Can't predict what that's gonna look like, but I would say that every time I have a conversation with Travis: folks in the pharmaceutical space or the clinical trial space, the capabilities of AI and the models that can contribute to the speed at which new treatments come about Travis: is Travis: powerful Travis: um Travis: on a personal basis.

Travis: I'm not a big medicine fan. Travis: I'm very much a an individual that loves food as medicine as a concept. Travis: Expand it from there. Travis: And so Travis: what I would love to have happen is actually for Travis: AI organizations Travis: to stand Travis: up capabilities to advance the science of how Travis: how food and nutrition Travis: and Travis: Better health, how you take care of your body Travis: actually impacts your future health.

Travis: And less about whether or not we can develop the fastest. Travis: you know Travis: research Travis: to Travis: pill dispensary Travis: way to cure some illness, right? Travis: And that's that is a Travis: a Travis: thing that I really believe AI Travis: should be tackling. Travis: I I don't want AI to tackle the solve the problem Travis: after it's happened, the sick care Travis: space.

Travis: It needs to solve the how do we just be healthier humans. Travis: And Travis: Part of our organization is is protecting human data from badly behaving agents. Travis: I think AI Travis: should be one of protecting human sickness and protecting human health so that we end up in a place where sickness doesn't even come in until you just fall off the cliff and die. Travis: Like I I don't know if I've mentioned that before, but that's like Travis: it's my thing.

Travis: I want to climb Mount Everest and then fall off the other side and just die. Travis: Like I don't I don't want to slowly just like Travis: wither away. Travis: You know, I want to Travis: I want to live until I'm just not healthy enough to live and then Travis: and then that Travis: my that's my day. Travis: So those are Travis: that's a couple predictions, a couple statements, and a couple wishes Travis: maybe.

Travis: Perfect. Travis: I like it. Erik: Well, thanks, Travis. Erik: It's been great.

Erik: We really appreciate you having on Erik: you on the podcast today. Erik: And we look forward to hearing what you guys think about uh what was shared today. Erik: So please leave any comments Erik: if you are so inclined and we look forward to having further conversations about AI in the future. Erik: Yeah.

Erik: Thank you, gentlemen.