Healthcare Is Broken. Can AI Fix It? | Petar Lazic

Using AI, generative AI, specifically large language models and agentic structure, reading these documents and extracting that information is a huge lift off of the hospital's uh administrative back. Uh this this would take something like 10 to 20 hours, a team of four to do before and this is reduced now to an hour and the the accuracy is very very good.

I think you used the words broken. Did you say broken? That there is something broken in health care.

This is a very interesting angle. Tell me more about what's broken in healthare.

So yeah, and then what I've noticed is that the the smaller your context window is the better.

So almost small language models are more effective than large language models if that makes sense.

AI is it hype or help? Good day and welcome to the show. I'm your host Pete Cooper and our topic is AI hype or help. Today's guest is Petar Lazic from Axion Advisory. He helps health care organizations make the transition from AI as a wishlist to something that's real execution. We go deep into the health care. So hang on for the ride. Let's get into it. Welcome Peter. Welcome to the show. Um Peter, you're with your company called Axion Advisory. Um first of all, welcome. And secondly, um, tell us a little bit about your business and what you do.

Absolutely. Yeah, thanks for having me.

This is quite exciting. It's one of my favorite topics. Um, so yeah, I'm the founder of Axion Advisory. Uh, it's a healthcare AI and operations advisory firm. Uh, my background spans in healthcare ops, uh, revenue cycle, uh, analytics, automation, and then most recently with AI enabled workflow design. And really the focus of the organization is to help organizations figure out where AI can genuinely improve their work. Uh what has to be fixed first um creating uh efficiency through optimization of current workflows and whatever technologies available

and then how to move from uh interest to governed uh practical execution.

So whereabouts are you located Peter?

Uh I'm based in Chicagoland downtown specifically.

Okay. And so I guess you have a a ready pool of uh health care organizations to work with. Do you narrow it down? Is it hospitals? Is it clinics? Is it some some other healthcare organizations you focus on?

Yeah, my background has been uh really kind of the gambit.

It's been uh practice management from physical therapy clinics um all the way up to large hospital systems uh both nonfor-profit commercial um public health uh in the consulting industry we we really focus in the larger scope of things but we we also had uh large physician groups etc etc but uh Axion is really focused um in that uh mid uh midsize range where the folks who are going to try to uh optimize uh and bring new technological uh innovations within their practices.

Right. Right. Awesome. Well, that sounds fascinating. So, I mean, how did you get into this, Peter? Let's take a step back. Where did you come from and how did you arrive here?

Uh, yeah, I have an eclectic background.

Uh, I initially wanted to be a doctor. So I worked uh with an X-ray which is a really interesting uh radiology is a fascinating part of the hospital simply because you work in uh several different departments. One day you will be in the X-ray department, the next day you'll be in surgery, the third day you'll be in

um ER, fourth day you're going to orthopedics, fifth day you're doing portables.

So you kind of get this plumbing view of a hospital for lack of better expression.

Um and something clicked there. If you have the plumbing, you also have a satellite view from my perspective. Uh, but I had a an interesting opportunity to go sell exotic cars in my early 20s.

So, I did

and uh helped build a dealership with what a mentor of mine

and that's where I caught like the finance and business bug. Um, then I spent some time in logistics working with friends.

Uh, and eventually decided, hey, what do I really want to do? And I wanted to take some of the pragmatic and practical experiences that I learned through business development in the logistics world and fix healthcare. Um, and that's how I got into practice management and there we went from paper to digital and I was hooked and I wanted to do that for large hospital systems and I got into consulting.

Nice. So exotic cars that's a very very different field to this.

I mean when we we talking Ferraris and Lamborghinis that type of thing.

Yes.

That must have been fun. Yeah, it was awesome. We started with six cars and grew it to 60.

It was a wonderful experience.

Yeah. Yeah. You'd get to drive them and, you know, be with them. Yeah. As a as a young man, I always wanted to have a Ferrari. I don't know why it was a Ferrari. That that desire doesn't seem to be there now. Um, but I do watch some podcasts with some guy in the UK that rebuilds these um broken and damaged exotic cars. But that's a side issue. So you mentioned also um that there is I think you use the words broken. Did you say broken? That there is something broken in healthcare.

This is a very interesting um angle. Um tell me more about what's broken in healthcare.

[snorts]

Um well there's a lot um both from a operations perspective and then also from a philosophical perspective with some of how our insurance is uh essentially the burden of proof is on the entire physician side

right

of the practice. So, the the money is almost like an escrow with these insurance companies, and they're they're basing their uh their decisions off of claims. And to get to the claim to be as uh proficient and accurate as necessary per each insurance, per each plan, for each group, it's quite a lift on the hospital side of things.

M

and there's this misconception that oh there's this there's all this billing that some of that bill most of that billing is dictated by the insurance side of things.

So there's there's a firstly an imbalance from that perspective

and then secondly there is the administrative hygiene within h within hospital operations is subpar.

There's no other way to say it.

Yeah. So coming from outside of the US, I found it extremely confusing and also um moving into that area with some of the products and services we offer.

I found there was a number of $265 billion um which is money that goes into the administration of health care which is about I think it's a third of all of the costs um for running healthcare in the US is administration

and I I believe there's some active federal initiative to try and change that but it seems like an enormous amount of overhead because none of that goes to the uh none of that goes into the pockets of those providing the service or those that receive the service. So, it's a big really big bolt-on um just to make it run. Now, is this something that we're talking to? Are we talking about the same thing? This massive amount of money?

Yeah. This inefficiency.

There's this huge gap that causes it's almost like an iceberg versus the treatment that you're you see at the top versus actually where the the money is funneling and and going into costs that are just large inefficiencies truly.

Yeah. So you mentioned this other thing which is surprising to me that I think you said the burden of proof lies with the physician. So it lies with the doctor or the person who's providing the service. So it sounds to me like what you're saying is they have to take on responsibility to provide the service and then justify that they can get paid for it after the fact. Mhm.

And then the the the defining layer of complexity within that is that each insurance company, so like the major ones, let's say Blue Cross, Blue Shield, Healthcare,

Etna,

uh Medicare, Medicaid, they all have specific standards that are set to specific plans and there are millions and millions upon plans.

M

so you can have a Blue Cross Blue Shield uh patient that's has Bluec Cross Blue Shield of Illinois that pays the same amount as a Blue Cross Blue Shield of Connecticut but has an entirely different um getting their uh like primary benefits checked and how how payment is dis dispersed

and now those nuances still fall upon the burden of that physician or that hospital administration group

and the the complexity is just too vast to keep up with And then everyone changes from year to year.

So it's always hard to explain complexity simply. I often struggle with this um as well. I'm going to try and unpack it. So I just maybe talk some of the numbers.

I think there's a 100 plus insurance companies in the US. Maybe there's more. Would each one of those have its own method of communication? Like it's not a standardized way to communicate with that insurance company. [snorts]

Correct.

So that's one complexity, right?

Yeah.

So you got to kind of know your insurance company and who can know 100 insurance the interface between each of those is slightly different. Like you said there's a portal. Um I mean we all I was without insurance here in the US for a while and finally got sort of got over the line on it.

Um having such a difficult insurance and over overhead system does mean that there's people out there that aren't getting access to service, right? Because either there's these problems with the administration they can't get over or they just can't afford it. So a lot of people and you know for the c this country which is you know the wealthiest in the world it's still shocks me that some people just can't get access to the basic needs you know of health care

right

and I guess that's the tragedy of of this it's it's fine if you can afford it all but if you can't afford it all um then you're out um so that must be uh that's another rabbit hole we can go down as well. Oh yeah, 100%.

And that solved with solving the inefficiencies first.

Yes. Right. So all right. Well, let's jump into that. So

you work with a range of healthcare providers. Tell us a story if you can about where you've maybe solved an inefficiency. I guess ideally with AI that's the topic of our conversation today.

Yeah. Um a lot there's been several inefficiencies that we I've come across. Um revenue cycle is a funny thing within the health care system because um it is a cycle but it is treated as silos. So it's split into three pieces uh which is access and then mid-revenue cycle which is like coding and grabbing actual interaction with the patient and then backend.

Um so there are a lot of aspects that go into this. So that patient access point is where you collect the patient information and then you also have to run their insurance. So do they need pre-authorization? Do they need certain pieces of information? And this is something that uh the hospital organizations and and physician management groups and MSOs lived and died by because if one piece of information was grabbed incorrectly at that entry point of how that person's insurance was read, well then downstream a lot of things go wrong.

So one of the biggest uh inefficiencies that AI has been able to solve is reading these complex payer contracts that can be anywhere from 50 to 200 pages

and you get one at the beginning of the year. So all of your you know the hospital creates deals with united health care specifically with this specific branch of united healthcare. Everything follows that subsequent contracted rate and all of the rules alongside of that. Using AI, generative AI, specifically large language models and agentic structure, reading these documents and extracting that information is a huge lift off of the hospital's uh administrative back.

Um this this would take something like 10 to 20 hours a team of four to do before and this is reduced now to an hour and uh the accuracy is very very good.

So these are just basic basic issues that using basic generative AI can solve kind of right off the bat.

So you so you said there was a team of four taking a 10-hour day I think was that what you were

employing 20 hours a team of four to do this. So about a half a weeks of work, a team of four.

And that came down to to what again?

Hour [snorts] and a half, two hours.

Wow. For the same team.

Yeah. And you're just using one uh model. You're just running a essentially an automation. Yeah.

That has a large language model attached to it, which is a f the fancy word is an agent.

Yeah. It sounded like about a 10 to 20 reduction in time and effort.

Yeah.

Wow.

That kind of makes sense. So, just for our audience to sort of understand that this to simplify it, there's a lot of text that needs to be read so that you can apply the correct uh access and get it reimbursed. But someone for someone to read and understand all of that, I can't imagine that job. Honestly, I think it would be awful. But um LLMs are very very good at reading text especially if you prompt them well and you say I'm looking for this and I'm looking for that.

So it kind of brings the context into the reading. Um and it it seems like almost a perfect solution provided and this is always the kicker. It does it accurately. Did you ever come across any accuracy problems? Uh I've noticed that the accuracy problems uh are only happen more frequently in very large context

situations. So if I have for example a very demanding research prompt

and I wanted to pull information from hundreds and hundreds of sources

um and those are varying sources. That's when I've noticed uh

mistakes but it's not like it used to be.

Um and uh hallucinations were solved because the the issue was how the system was being fed by the data scientists in the back

where a non-answer was treated as a punishment versus any answer. So the system was taught through like a X's and O's and it decided it's better for me to make up something than to not give something and and say it's not that I don't know it

and that in the last two years has been been ironed out by like the anthropic folks, the open AI folks. It's it was the reward s mechanisms that were in the back of the models of how they were trained.

Wow.

So yeah, and then what I've noticed is that the the smaller your context window is, the better.

So almost small language models are more effective than large language models if that makes sense.

Okay.

Yeah. Like clinical language models are awesome but you they won't know anything outside of that.

So that would be a a model that is isolated in some way and then dedicated specifically for clinicians.

Does that mean it's a model running privately u on its own server or its own computer

or yeah they could be renting cloud space within a server spot but yes it's okay

they're running uh there there are certain models that clinical models that run specifically within that organization's network a lot of them are like the for radiography

and that one's a very high uh impact use case because there's hundreds millions of hours of actual recording of the radiologist reading the image. So they can overlay the language over the image of what was diagnosed and train the models that way.

Yeah. Wow. That's a whole other area of uh of imaging.

Um so um this was fascinating how you can um so just to summarize here we've got the ability to take a large language model and put it into a private space and tune it fine-tune it for the clinical space so that it doesn't react like a human. Sometimes humans do the same thing don't they? They say they make up an answer rather than say they don't know.

Correct.

Yeah.

And it can be a cultural thing. Right. I think uh in some cultures, I think Scandinavia, Europe a little bit, they they don't mind saying I don't know. Uh in other cultures, uh maybe in Asia, um you know, that it's very uncomfortable for them to say no or they don't know. Yeah.

And I think it takes a certain level of self-confidence in in humans to say they don't know. And then what's the next step? Well, you don't know. You can maybe find out. Maybe you can't find out. Um, I guess the I noticed this with using large language models as we all do and how friendly they are, how nice they are, how willing they are to please and as you say the reward system is there um for

uh for for pleasing um so producing something that looks good um which might complete completely wrong.

Yeah. I mean guess their reward system is contained within the chat. It's not a human being has its reward system which goes a bit further on than that.

Like they have a long-term career, they have credibility, integrity, things like that which are part of their reward system.

Anyway, it sounds like that that particular problem is solved. Um and um it's producing results.

So that means we could have a very big impact on this 265 billion dollars because we can 10 to to 20 times reduce at least a piece of that. Is there another area that um I mean the imaging side of things tell us more about that? Yeah, there are so many areas. Uh, a lot of the coding and documentation is repetitive and a lot of the human error comes in there.

But even before

um even downstream from that is uh ambient listening technology

um which is I I think probably the the single best application of AI in the healthc care space. Um the the the problem of that entire burden of the organization being on the clinician in that space with a patient and their entire interaction is taking the notes.

And what they're doing they're they're pro they're going through a checklist kind of like a pilot uh crossing things off almost avoiding the possibility of getting sued

because that's another burden that they have.

And then they have to be able to reference to be able to diagnose later to actually get the code that's going to go onto the claim.

Oh,

so they have this really intense space where they're focused on just taking these notes and what h and what ambient technology listening technology does is it removes that. It just listens to the conversation. The doctor can go and treat and it hears everything and automatically through kind of the same thing with radiologists. We have hundreds of years of CPT codes, diagnosis codes, ICD-10, ICD-9, all the codes that before that of here's what was documented, here's what was what happened, here's how it was treated.

So now we have all that training data. We have the audio and now it just it it removes that necessity of having to do documentation right there. And then it removes afterwork documentation

because that's where the doctors get killed. Their bandwidth. They have to go home and then they have to go through their charts and write up the whole notes, the plan of care. They're called soap notes, which is like exactly what the insurance is going to get that's going to justify this care. So, the level of work is like, okay, they're at the hospital for 40 to 60 hours, then there's another 20 to 25 of documentation at home.

Oh, wow.

And in this so there's a huge clinician burnout and simultaneously the smartest people of this country don't want to have to work like this. So they don't go into the field and what it's caused this vacuum where we're losing this great generation of doctors, highly capable doctors and they're not sticking around to train the new generation fast enough.

Wow.

So we there's it's almost a race to be able to augment some of this crazy workforce and burden that's a uh causing this clinician burnout to happen

and this ambient listening technology into the coding AIs and automations is the light at the end of the tunnel for the

wow I I didn't really realize I mean um I guess I'm ignorant to the life of a doctor here in the US but it's it seemed to me a lot of people would be drawn to being a doctor because they wanted to provide care to others. They they naturally wanted to do that. They wanted to understand people, work with people, help people.

Now, they didn't expect going into that that they would have to do so much tickboxing and dealing with insurance claims and and the rest of it. So, I can definitely understand um there's a misfit between what their desires are and what everybody essentially wants. Well, not everybody, but most people who receive care want the attention of the doctor, and the doctor wants to provide the care, but yet we've got this thing jumping in the middle there, creating this overhead and burden.

Um, yeah. What was the name of that tool that um the AI tool that helps with that?

It's called ambient listening technology. There are several uh

right

vendors in the space that do that.

Ambient listening. Something. So that it sounds to me like it's really just sitting there

um and collecting the information, whatever's being said and turning that into text and then making good use of that to help guide him to making the proper insurance claim.

Correct. Uh it could be on their phone, on their laptop. Uh places like the Mayo Clinic and Cleveland Clinic are putting like microphones in their ERS. And here's the other fascinating aspect of this

is that if the the powers that be within these hospital organizations, the the most prestigious ones play their cards correctly, they take by by having these recordings, you know, a litigation is completely solved.

If something happens, you have the ability to pull up exactly what occurred, but b it takes away the power of the insurance's data play. Because right now that's where the insurance that's what makes it so American is that we have a third party private company that has our healthcare data not our government

right so that's what keeps it American in our world

but if the hospitals have the entirety of the interaction kept as is that takes away the validity of the data that the insurance has and it sweetens the power to the providers now.

Right. Right. See that if they play that correctly,

right? That's that's f well that sounds like great news.

So, it sounds like AI really has a place here in addressing a really key pain point. Uh and and is there [clears throat] and it sounds like there's a path forward and it's intuitively it intuitively makes sense to me that administration is something that AI can help solve because I know in my own world when I'm working with AI it often cuts down a lot of administrative burden if I have a lot of documents to read I can get it summarized easily I can interact with it through the chat is very easy way to interact act and very satisfying in a lot of ways. Um, so it's presenting things in a way that's digestible to me.

[snorts] Um, you know, I come from a a generation that didn't have computers, believe it or not, and it was all on the the written pencil and paper, and reading was a very important skill um that you had to do a lot of reading in order just to keep up with your basic education, right? Not being a fabulous fast reader, it was always a bit of a problem for me. But with LLMs, I can outsource a lot of that work and still have all the other things that I think I am good at.

The idea generation, the creativity components, the bringing in of new information, the challenging and the being structuring of an argument, all that sort of stuff is now is is amplified because I've been able to solve that small piece of the problem. And I think you've illustrated really nicely how that works in in the field of physicians and doctors. Great news. So, um, quick question I've got here for you in our preloaded questions. What are some of the mistakes that people are making when they're doing, uh, AI roll out and implementation?

Uh, I think this applies to multiple industries, but definitely in healthcare, they start with the technology instead of the work.

So, they kind of ask like, what's a quick win? Uh where can we use AI before they've defined a problem that they're trying to solve.

So they'll they'll overlook things like uh workflow stability uh data quality which is data [snorts] quality is the probably the most important cornerstone

decision ownership um baseline metrics

um they kind of gravitate towards these shiny pilots uh that really can't survive production. The pilots are great. They show off a really wonderful piece of technology. Um, but it's uh the technology isn't as malleable as the organization has to be.

It's kind of like uh in that movie Hitch when he's teaching him how to date

and he says you have to go 9010 with a kiss. Mhm.

It's like as an organization, we have to change 90%

to fit the logical systemic build of how AI operates.

It's because we're inefficient, not because it's inefficient. It's really efficient actually.

Yeah. So, one of the things you said there which we need to unpack a little bit was was data quality.

Um, explain to me and the audience what data quality is. Uh data quality is um the both the input collection and uh data display of your information. So uh it's are your are the frontline employees uh entering the information correctly?

So is the information accurate and truthful? Mhm.

And then that's the first part. And then are there multiple underlying systems? So it's like do you use one CRM or one EHR but then have multiple vendor solutions that are plugged into it.

So where is the source of truth?

That would be a huge data hygiene point. Mhm.

Um and that's a lot in the health care space that is going to be one of the hardest um parts to solve because uh 99. 99% of organizations large and small either have multiple EHRs or multiple uh technology vendors bolted onto their EHR solutions.

And EHR stands for what? For the audio

electronic health record.

So like there's Epic, Oracle, Athena Health, those MEDITECH, those are the big ones. Those are that's just like saying Salesforce,

right? A CRM or customer resource thing put in there. So So you're saying that they would have multiple ones, multiple of these different systems running.

Yep. Happens all the time.

Talk to each other.

Yeah. They don't. Sometimes they have

talk to each other. Multiple CFOs overseeing different systems,

right? So, is it possible that one patient could have a something in an EMR over here and a different thing over here?

Mhm. [clears throat]

Oh, and they don't talk to each other. So, which one is the right one?

Yeah, they uh in theory, the way it should be set up is that each patient has uh electronic health record. So, they have one number that presides over that system,

right? It should be ported over. But kind of similar to uh you know being a doctor isn't as attractive. Being an IT person in healthcare is like the least attractive after Silicon Valley and financial tech and then healthcare tech is kind of last in line because there's so much bureaucracy

and it's really hard to get things uh moved and it's it's slow and

yeah interesting.

I was working for a company a few years back where the IT guy was there and he was actually pretty good but everybody was treated him really really nicely. Uh and then they had this funny little thing. They had this office uh where the IT guys were. It was a little closed office with servers and everything running and there was a couple of guys in there and you'd knock on the door and they'd open the door like this and they'd peek out. What do you want? Can I come in? No, you can't come in. [gasps] Yeah, I've never seen that before. I've never seen such closed door policy. And um they would like shut off completely from the rest of the organization. I don't know what was going on there.

That's funny. There was that great show, The IT Crowd.

Yeah. Yeah. This was this was exactly that point. These guys were separated from the rest of the world. Yeah. Yeah. Okay. So, let's talk about um you know, where do we think the the lowhanging fruit is? Where where else can we go and make some big gains with AI? [snorts]

Uh I'm very optimistic about AI. So, I see this this upcoming time as the age of abundance. If if we can figure out how to

uh how to get our stuff together. Um I think the gap between AI ambition I think there's a gap between AI ambition and operational execution.

M

so that's why I think like the first steps are because a lot of these leaders

people don't know what they don't know and AI moves so quickly

that a lot of it is learning it is not so technical as it is uh mental like getting the uh approaching it with the right mentality versus it's almost like the way that we were using computers before was like a two like a bicycle with AI it's almost like we have a tricycle and people aren't realizing that it's actually gotten easier for us to interact and get more out of our technology.

So these leaders are surrounded by all these tools, vendors, agents and uh these pilots that can often lack a practical starting point.

So I think this this is a huge space like mapping these workflows starting to understand where the the the proper data that we need where the the sources of truth are to prioritize use cases define the governance um I think this is where our biggest gaps to fill are right now something I've noticed especially in healthcare um is a lot of these folks who want these solutions who who see these really cool things. Uh it's like some of their technology stacks. It's like they have a Fiat, but now they want to put this Ferrari engine in.

It's the the wheels will fall off the moment you hit the gas.

And the these are the kind So this is where this lowhanging fruit is is this uh the show Madman uh Don Draper is like if you don't like what's being said, change the conversation. That is the lowest hanging fruit is changing the conversation of how this is done and not using these uh these really aggressive words and it's the the part of AI right now is not going to be pretty. It's going to be we have to tell you you might have to rip out your entire system

and all of your investments for the last like five years. This isn't going to cut it with how the technology is working

or so. Yeah. So, some of those conversations have to happen.

Yeah.

So, it seems like we've gone down a path. U we're so far down and so invested in it, now we have to sort of backtrack or backtrack the path or cut across the bush to get to the because that path is is is getting very very hard to go forward on. And um yeah, so this sort of undoing and unwinding a lot of uh a lot of ways of working and changing it into a new to a new way of of operating. But I don't think that's necessarily unique to healthcare. I think it just from what we've even this just discuss this discussion, it's deepened my understanding of healthcare and how complex it is.

Um if if I just use AI in some really simple way, um you know, I want to cook something downstairs and I turn on the voice activated thing for ChatGPT and I ask, you know, how long should I put the chicken in the oven?

And it comes back to me, bang. It's a very simple use, very easy for me to adapt my way of thinking because I would have normally just done looked it up on the internet

on my phone or the back in the day I would have come upstairs to the computer or going back further I would have asked somebody or gone to the library to get a book or you know done all those different things. It's just it's just made that process so much sleeker and easier for me.

Um and it was a small step. Unfortunately, these complex systems, it's that small step is really hard to take. Yeah. Yeah. I would like to be able to sort of walk away with the promise that, you know, that if companies made a small investment, they would get a big return, but it almost seems like we it's an inevitable big investment. Um, but it's just it's a painful one that we're going to have to take

unless I mean, do we think this is a really broad question. Do we think AI is going to have a big impact on that 265 million? I mean, is that realistic?

Yeah. Um, 100% is.

Something that I always get reminded of uh is there's a really wonder it's electric cars is the first place where I wanted to say but there's a really wonderful image from like 1908 of downtown New York City and it has all the horses and then it's like 1913 and it's all cars in the same

same place. And then the same thing happened with uh electric vehicles like how long the batteries go, how cheap it is to produce them. Um right now the investment should be to optimize get your workflows to first principles.

It shouldn't be just jumping for this technology right away.

Um because the truth is is tokenization is expensive and you know people think like these agents are replacing a worker. They're not. An agent is just a workflow, but you're still paying like you're going to be paying a tokenization for the use use of that compute power,

right?

That's that's still not cheap, but that will flatten

will drop the same way everything else has.

I've been telling some of the leaders I talked to, you know, they're always talking about EBITDA and I've been like EBA, you know, earnings before taxes, depreciation, amortization. Saying we have to add one more T in there because it has to be earnings before interest, taxes, tokenization, depreci.

Yeah.

And like yeah, they can some of them are like we're just going to wrap it up into RFD

but and because it costs a lot to build the agent too.

So okay, so most people would be familiar with say ChatGPT and and having a free version of ChatGPT you can do something for nothing, right? Or if you have paid version, it's $25 a month or something like that. But we're not talking about that, are we? We're talking about tokenizations. We're talking about going in through an API.

Yeah.

That's much more expensive, right?

Yeah. That's where you're using legitimate compute.

Um just going through chat, it's not that's so so low usage.

That's why they can afford it.

It's when you start calling the APIs and pulling them in there and bu actually building things.

Yeah. Yeah.

That that it's using a lot of compute power.

Yeah. Yeah. Well, I I mean we saw we've seen this before, haven't we? We've seen this when computers came in and the internet came in. We ended up substituting one thing for another. Um we substituted, you know, people could do a lot more with a computer. They wouldn't need, you know, their assistant. For example, I used to, and believe it or not, going way back, there was an older gentleman who had a secretary and he did not answer his emails directly.

He had each one that was irrelevant printed out for him

and then he would write his replies on them and hand them back to her.

Can't imagine anyone doing that today. But that's an example of a very different workflow. And that was the way things were done before computers, right? The messages and mimos were collected for the executive. He would answer them and then they would go back. So you had that person as that doing that [clears throat] piece of work. Computer come computer comes along, everyone's on the computers, they're answering their emails directly. Okay now we we go into the next level which is somebody was doing that job that the token is now doing. Right?

Um so the money that's coming into the top of the business isn't going to necessarily give someone a job. It's going now to it's going out to this third party this uh model these big model makers, you know, the anthropics of the world. So, they're intersecting revenues is what I'm trying to say. The the money isn't going to provide someone a job. It's going to [gasps] to fund the AI model. And I guess this is why we've got such investment in this. People can see that coming. The workflows are changing. Money is going to be redirected from the workforce to the AI workforce who's intersecting that value chain and that's why these things are nearly valued at a trillion dollars.

You know, I think Anthropic went up was it 900 and something

something billion dollars. Yeah. And SpaceX become an AI company, would you believe? So, you know, they're all jumping on that because they realize this is what's happening.

They just acquired Cursor, which is like a an AI coding tool. One of the first ones. So that's very exciting.

Was it $80 billion or something?

65 billion. There you're up with the news, too. I saw something like that this morning. All right. So [clears throat] interesting. So where else should we go with We talked about metrics. Metrics is another thing. What's some of the key metrics that we should be using in when we roll out AI in healthcare?

I [clears throat] I mean you have to start with the accuracy. Uh there was a huge misconception in the very beginning where it's like we wanted these unbiased models

and uh those were just uh signaling words. You want your model to be trained on a large subset of data but you want it to be very biased to you

because there are very different treatment cases.

So accuracy um rework uh error rate

uh escalation volume

um cycle time uh throughput uh financial impact downstream consequences user burden adoption that's all I can rattle off

that's a lot um

wow straight off the cuff you really do know this really well um I Um, escalation was one that seemed to be easiest to understand. This is where

a human has to get it back involved and to to figure something out. Right.

Yeah. And you want to minimize that because you want the AI to do as much as it can.

Right. And that's that that escalation is actually the the like the red herring and how all of these systems are going to work.

Um a a big thing that so many people roll their eyes on when I mention governance,

but it's treated like a box that you check, but it really isn't. It's an operate. It's a human being operating system.

So, it's like we're creating very similar. We're creating a cognitively conscious uh biological system, operating system that is overlaid over the AI and will be able to intercept almost with like swarm-like logic, but the the people themselves don't realize they're a part of it and it almost works autonomously.

Yeah.

And that's that's that's not the lowhanging fruit. That's like the holy grail of what we need to develop in this space.

Yeah. Yeah. Gotcha.

So there's a lot more metrics we can dig into and I think we we haven't really got time to do that. So coming to the end of this very interesting discussion and there's a lot we could unpack here. I think Peter, we need to have you back. Um do you have any questions for me or things that you wanted to talk about here today?

No, this was wonderful.

Thank you. So one final question for you. AI is it hype or help? Um, it is help when it's applied with operational discipline

and it is uh hype when it's treated as magic and like a sol.

Yeah,

that's a really nice way of phrasing it. That's the best I've heard on all these podcasts. I'm gonna have to bottle that up, Peter. Awesome.

There's a lot of really good stuff here. Look, thanks. Stick around um and uh we'll have a quick chat after the show, but um thanks for coming on.

Yep. Thank you for having me.

Well, we went deep into AI and healthcare. Now, we talked about how big the problem is in the US. 265 billion dollars goes in administration. That's about a third of the cost. So, that doesn't go to the physicians and the practitioners and it doesn't go to the patients. It's just a bolt-on. It's enormous. Now, there's a lot going on to try and improve that and AI certainly has a very important place to play. Now, if you want to listen to the whole podcast, there's lots of good stuff in there.

If you want to know more about what Skillion does, look to the links in the description. Same for Axion Advisory. I hope that helps and thanks for watching.