21 Jul 2026 · 23m · Matthew DeMello, Alex Tyrrell · 25 quoted facts
Alex Tyrrell of Wolters Kluwer argues that agentic AI in healthcare depends less on how good the underlying model is and more on the systems around it. Ordinary chatbots create content, while agents act: they plan, reason and carry out tasks across connected workflows. That shift makes domain-adapted reasoning, sound infrastructure and security the deciding factors.
The main tension is that agents are new but need old foundations. Tyrrell says APIs, databases, access controls and compliance still matter, and matter more once software agents can send traffic far faster than people clicking buttons. Teams must therefore modernize backend systems and avoid monolithic designs. They must also watch for model drift and plan for harder-to-trace failures and new security risks.
Tyrrell frames the difference as 'create versus do'. An LLM produces text or images, but an agent wraps the LLM in extra technology so it can plan and take actions in the outside world.
Human users were slowed by how fast they could click buttons, but agents have no such friction. Tyrrell warns that traffic through MCP servers and tool integrations can spike far beyond what a SaaS product used to see.
Tyrrell says it is almost never one model: teams start with a frontier model, adapt it to the domain, and repeat as new model generations arrive. He also describes adapting the model on the fly to the facts of each case.
Trained professionals are best at breaking a job into steps for an LLM. Tyrrell says that fine detail lets you walk back through the reasoning and check each step was correct.
Tyrrell says big black-box systems lack proper endpoints and fine-grained entitlements. Agents then get whole payloads instead of only what they need, which weakens auditing and least-privilege access.
Alex Tyrrell says an LLM creates things like emails, poems or images, while an agent does things: it plans, reasons and executes tasks. The LLM is still at its core, but extra technology lets it interact with the outside world.
Tyrrell says agents can send requests to backend systems much faster than humans clicking through screens, so traffic can spike unexpectedly. Companies need robust APIs, microservices and well-scoped access controls to cope.
Tyrrell says building frontier LLMs is too costly for a company of Wolters Kluwer's size, and basic tasks like summarization or translation are better bought as managed services. He sees value in adapting models to a domain and possibly building small models for specific devices or settings.
Tyrrell describes a monolith as a huge black-box program without the right exposed endpoints or fine-grained permissions. Agents then can't be given only the specific data they need, which makes auditing and safe integration harder.
Tyrrell says connected agents enlarge the attack surface and raise questions about how to verify an agent's identity and who it should talk to. He adds that LLMs help attackers scale up, and that standards like HIPAA and SOC 2 must be re-examined for LLM use.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Wolters Kluwer is described as a global provider of information solutions, software and services across healthcare, tax and accounting, compliance, legal and regulatory, corporate performance and ESG.
“Wolters Kluwer is a global provider of information solutions, software and services across healthcare, tax and accounting, financial and corporate compliance, legal and regulatory, corporate performance and ESG.”
Alex Tyrrell says Wolters Kluwer has a print and textbook heritage and is a 189-year-old company.
“we were a print heritage and textbooks, and by the way, we were 189 year old company and there were no computers.”
Alex Tyrrell says Wolters Kluwer evolved from print to digital, online, cloud, SaaS and what it calls expert solutions.
“the evolution from print to digital, to online, to cloud, to SaaS, to what we call expert solutions, we have been continuously focused on.”
Matthew DeMello says agentic AI was the buzzword of 2025 for the general public.
“agentic AI sort of was kind of the word of 2025, especially for the lay public.”
Alex Tyrrell distinguishes a plain LLM, which cannot send an email itself, from an agent, which can act, plan, reason and execute.
“It cannot send an email for you specifically, right? That gets you to agentic, right? Agents do things, a act, plan, reason, execute.”
Alex Tyrrell names the electronic health record workflow as a prime healthcare setting for agentic AI.
“one of the most obvious places that that we talk quite a bit about is kind of the whole EHR workflow, right?”
Alex Tyrrell says handling a claim denial is part of a connected healthcare workflow where agents can play a role.
“If there's a, you know, a claim denial, that's actually a fairly connected workflow.”
Matthew DeMello says conversations in 2023 centered on foundational versus bespoke models, a framing that has since evolved.
“This was definitely the majority of the conversations we had in 2023. There was this language of bes- uh, foundational models and bespoke models, and that has evolved.”
Alex Tyrrell says teams typically begin with a foundational frontier model and adapt it using methods such as fine-tuning, low-rank approximation, parameter-efficient fine-tuning and in-context learning.
“start with a foundational frontier model, but you're going to domain adapt it, either through ins instruction, fine tuning, supervised fine tuning, low rank approximation, parameter efficient fine tuning, in context learning.”
Alex Tyrrell says Wolters Kluwer does not build frontier large language models because the cost is prohibitive for a company of its size.
“we're not building, you know, frontier LLMs, you know, it's cost prohibitive for a company of our size.”
Alex Tyrrell says early efforts at Wolters Kluwer focused on grounding and retrieval-augmented generation so answers were attributable to sources and did not hallucinate.
“model grounding and and retrieval augmented generation and making sure all the answers were attributable to source and didn't hallucinate.”
Alex Tyrrell describes a dynamic chain-of-thought approach that adapts the LLM to the exact set of facts in each case rather than only to the general domain.
“not just adapting the LLM to the domain, adapting to the LLM to the exact instant set of facts and patterns presented.”
Alex Tyrrell says breaking tasks into fine detail gives explainability and lets reviewers walk back through the reasoning steps to check them.
“you break down things to such fine detail, you get that explainability, you get that transparency, you can walk back through the reasoning steps and make sure that they were correct.”
Alex Tyrrell uses the 'Bob Jones' example to describe the long-standing NLP problem of resolving which real-world entity a name refers to.
“Like Bob Jones, how many times you're gonna find Bob Jones in a document? Which Bob Jones were you talking about, right?”
Alex Tyrrell says hyperscalers are bundling capabilities such as summarization and translation into managed services.
“And now the hyperscalers are are bundling a lot of these capabilities and kind of launching services.”
Alex Tyrrell says he did a PhD on image analysis, which he considers lapped by newer models, and would not build it in house today.
“Now I did a whole PhD on this, and I I'm not saying they should take it away, but it's really been lapped many, many times.”
Alex Tyrrell says distilling LLMs into small language models for industrial or healthcare devices is an area worth investing in and building.
“distilling LLMs and creating SLMs and creating SLMs you can deploy in in various, like maybe industrial capacities, or maybe in a healthcare setting on more of the the operational side and devices”
Alex Tyrrell says organizations that build their own models take on ownership burdens such as monitoring for model drift.
“when you start building some of these things now, you own it, so model drift, you know, things like that.”
Alex Tyrrell says agentic systems raise the problem of identifying an agent and knowing who one is talking to.
“how do you identify the identity of an agent? How do you know who you're talking to?”
Alex Tyrrell says LLMs are helping threat actors scale up and better hide their attacks.
“the LLMs themselves, this technology, is allowing the threat actors to scale up their attack vectors, hide them better, come up with new, interesting ways”
Alex Tyrrell says agents remove the friction of human button-pressing, so the rate and volume of requests to backends can rise sharply.
“Now, when agents are doing the work, there's no more friction. The rate and speed and volume that they can send to your probably similar back end”
Alex Tyrrell says fine-grained API entitlements let an agent receive only specific, auditable information, supporting least-privilege access.
“that your entitlements are at the right level of granularity. So you're not just sending the whole payload back to the agent, you're getting the specific information that you can audit and ensure, um, least privileged access.”
Alex Tyrrell says compliance standards such as HIPAA, SOC 2 and ISO raise new questions when LLMs are used, including for business associate agreements.
“HIPAA, SOC two, ISO, all these standards and compliance, they're interesting with LLMs, right?”
Alex Tyrrell defines a monolith as a huge black-box program lacking proper exposed endpoints and fine-grained entitlements.
“it's just this huge 1,001, black box, you know, large program that really doesn't have the right exposed endpoints, it doesn't have that sort of right granularity of entitlements.”
Alex Tyrrell says semantic web technologies such as SPARQL and RDF never fully solved describing and harmonizing data, which LLMs may now help with.
“you used to hear about Sparkle and RDF, and, you know, uh, they never really got the job done.”
Tyrrell gives a simple definition of the shift from generative AI to agents. It frames the rest of the discussion about acting in the real world.
“create versus do, create versus do. LLMs create things, they create emails, they create a poem if you like, a recipe.”
Tyrrell explains Wolters Kluwer's long focus on customer workflows and how agents fit into connected tasks such as registration, authorization and claim denials.
“the idea that it's really about the jobs to be done, right?”
Tyrrell explains why real solutions combine several models and why off-the-shelf performance is not enough.
“it's almost never one model. It's just not how it works, right?”
Subject experts are best placed to break tasks into steps for an LLM. Fine-grained steps make an agent's reasoning checkable.
“trained professionals, right? Just go right to the source.”
Tyrrell contrasts his early NLP struggles with what LLMs now do easily, and argues commodity tasks should be bought while workflow value is built.
“They know who Bob Jones is. If it's somebody important, give it a little hint, and they'll tell you.”
Connected agents enlarge the attack surface and raise questions about identity and trust between systems.
“how do you identify the identity of an agent? How do you know who you're talking to?”
Agents remove human friction, so backends must handle traffic spikes and be tracked more carefully.
“Now, when agents are doing the work, there's no more friction.”
Tyrrell closes by warning against opaque, siloed systems and suggesting LLMs may help harmonize data.
“large program that really doesn't have the right exposed endpoints”
Welcome everyone to the AI in Business podcast. Today's guest is Alex Tyrrell, SVP and CTO of Health at Wolters Kluwer. Wolters Kluwer is a global provider of information solutions, software and services across healthcare, tax and accounting, financial and corporate compliance, legal and regulatory, corporate performance and ESG. Alex joins Emerges Matthew DeMello on today's episode to outline how agentic AI changes operational realities in healthcare, by shifting from systems that create outputs to systems that execute tasks across interconnected workflows.
He explains how adoption now hinges less on model performance and more on infrastructure readiness, domain adapted reasoning, and safe integration into regulated environments. Just a quick note for our audience that the views expressed by Alex Tyrrell on today's program do not reflect that of Wolters Kluwer or its leadership. In this episode, we cover agentic AI in healthcare, to go deeper on AI topics and learn how to evaluate AI vendors by assessing leadership expertise and why funding benchmarks can signal product maturity and stability.
Download our free PDF report five ways to select the right AI vendor at emerj.com slash AIV1. That's M E R J dot com slash AIV and the number one. Now the conversation with Alex.
Alex, welcome back to the program. It's a pleasure having you.
Ah, Matthew, great, so great to be back.
Absolutely. We're talking about models, infrastructure, enterprise readiness for agentic AI on today's show. Generative AI capabilities have accelerated. We're hearing from many enterprise leaders across the healthcare space, um, that they've, you know, discovered that model performance alone is not the limiting factor. Instead, adoption is increasingly constrained by infrastructure readiness, domain expertise, and then that ability to safely integrate AI into complex, regulated workflows at scale. Uh, we we know that agentic AI sort of was kind of the word of 2025, especially for the lay public.
Maybe for for insiders, really more the word of 2025 might have been infrastructure, because you you gotta build the back end before you build what's what's really going to touch the patient, the customer. How are you seeing agentic AI adoption differ from early generative AI use cases we saw a few years ago?
yeah, that's a great question. So I think, you know, w when I have conversations, you know, w with senior leaders and and I try to actually, you know, absolutely bullseye this this very concept. I I say, create versus do, create versus do. LLMs create things, they create emails, they create a poem if you like, a recipe. They can create images, they can create music, they can create sound. It's a very transactional experience, and the LLM itself cannot influence the outside world.
It cannot send an email for you specifically, right? That gets you to agentic, right? Agents do things, a act, plan, reason, execute. They do have the ability to interact with the outside world. They could potentially park your car, things like that. And so the LLM is still at the core of the agent, but additional technology accoutrements have been provided around it so it has more ability to interact with the real world. And I I think that's a pretty good definition of the difference, yeah.
So far, twenty six may not be a whole new year, but so far I got away with that.
no, I I've I've maybe differentiated on another big step level change we see from agentic AI is the ability to jump between domains and programs. Yeah, I can have an agent go buy me concert tickets, but it's gonna have to know where I live, it's gonna need That's my credit card information,
right, that's right.
it's gonna need to go to Google and and tell me know that I'm gonna have to get there. It's gonna need to jump to Ticketmaster, you know, open its own web page. So, and and we haven't really seen systems, especially that the public uses, really jump those domains, uh, quite yet. So that's that's another gigantic step, uh, w with agentic. But go ahead.
yeah, so I think, and then what we focus on in the professional setting, and by the way, at Wolters Kluwer, you know, we we've we've been focused on this for a very long time. And that is the idea that it's really about the jobs to be done, right? I mean, you know, we were a print heritage and textbooks, and by the way, we were 189 year old company and there were no computers.
Back then, as we made that journey, the evolution from print to digital, to online, to cloud, to SaaS, to what we call expert solutions, we have been continuously focused on. What is the day in the life of a customer? You know, what happens after you read, you know, a journal article, or you look up something in the tax code, like, you know, what's the next step in that workflow? That's really where we focus on jobs to be done.
And the fact that we've been doing this for well over a decade, when agentic came on the scene, it was almost like, aha, what a catalyst, what an accelerator to this strategy where we take our domain expertise, expertise, our subject matter experts. We combine it with best in class, in verified, trusted information, and then we layer in technology that is going to drive outcomes, right? That is owning an outcome potentially end to end. And with agentic, it could be very profound, right?
In the healthcare setting, one of the most obvious places that that we talk quite a bit about is kind of the whole EHR workflow, right? You know, everything from, you know, registration, pre-visit, everything from things like pre-auth, the post visit summaries to the post auth. If there's a, you know, a claim denial, that's actually a fairly connected workflow. And agents can play a role in each step of that journey, each step of that workflow. They can reduce friction, improve accuracy, and generally allow all these folks in a healthcare setting to worry less about these administrative tasks.
And again, focus more on delivering care than the patient. And that's exciting.
oh yeah. Um, I I I think it's it definitely agentic has made healthcare. And, um, many, many industries a tremendously more exciting space, uh, that we've seen in the last few years. Comparing it to, and this is very useful, I I I think this definitely would have benefited what we saw in internet adoption, that it that it was so far away, or the dot com burst was so far away from the rise of social, the the rise of social media, those two hype cycles, people's lapse memory.
They couldn't take lessons that we learned from the the dot com burst over to maybe having a more managed approach to a uh, to to social media, uh, that we've seen kind of play out over the last twenty years, to put it mildly. But I think even a couple of years ago, people, it's fresh in people's minds. This was definitely the majority of the conversations we had in 2023. There was this language of bes- uh, foundational models and bespoke models, and that has evolved.
That, yeah, yeah, we we giggle about it now, because it's
This is my favorite. This is, you know, whenever I'm talking to key customers, they're always like, which model do you use? Do you have your own model? Are you creating proprietary models? And, you know, recently, you know, we've had some some really interesting engagements where we're we're now basically, you know, explaining, listen, it it's almost never one model. It's just not how it works, right? I mean, it could be many models and a single solution, especially if it's agentic.
Generally speak, I mean, you might, you might start with a foundational frontier model, but you're going to domain adapt it, either through ins instruction, fine tuning, supervised fine tuning, low rank approximation, parameter efficient fine tuning, in context learning. That thing is not going to look like what you took off the shelf, it's going to be that. And you're likely going to do this multiple times, and you're likely going to have to continue to work on the improvements.
And obviously, as the model generations and model families improve, you're going to have to stay on top of that. So where we look at it is, you know, there's a lot, you know, we we're not building, you know, frontier LLMs, you know, it's cost prohibitive for a company of our size. But we also cannot just accept off-the-shelf performance, nor do we need to, because who's got great content and information? In some in the early days of this, we looked at, you know, obviously, model grounding and and retrieval augmented generation and making sure all the answers were attributable to source and didn't hallucinate.
But even more important, as models evolved and began to exhibit this LLM reasoning, when you really look at how you evoke LLM reasoning through things like chain of thought, which might be a little technical. When you when it comes to providing instructions to an LLM, you know who's really good at it, trained professionals, right? Just go right to the source. And like, if you had to do this task, could you break it down? And what we find in an agentic workflow, it's not so just much about, like, the face value, like what a human would identify as a task.
You may have to decompose that more subtly, or you may fold together multiple steps that a human would have done and just do it in one go. And that is where this model adaptation really comes into play. And you know, you've got tools like chain of thought, you've got, uh, fine tuning. I also think a big one for us is really, really, really, we do sort of dynamic chain of thought. So, like, on the fly, adapting, not just adapting the LLM to the domain, adapting to the LLM to the exact instant set of facts and patterns presented.
And that is really unique, and that gives a really, really better, better outcome. And by the way, I made the comment before. When you break down things to such fine detail, you get that explainability, you get that transparency, you can walk back through the reasoning steps and make sure that they were correct.
Well, and and that kind of y and you were mentioning out of the box models, uh, that that term has still lasted all all of these time, uh, all this time from 2023. Another big one, um, from 2023 was it started out build versus buy. I actually, I think it's more ratio now. Everybody's building something, everybody's buying something. So how how do you think that conversation i is changing around agentic AI and enterprise?
yeah, I mean, so, like, so I, I'll go back to, like, you know, some of my my my twenty five year history here, right? So I, I had two long stints. One was computational linguistics, or what we call NLP, which is things like named entity extraction and resolution. Like Bob Jones, how many times you're gonna find Bob Jones in a document? Which Bob Jones were you talking about, right? This was a classic problem for decades and it was miserable.
And it led to, like, unending levels of embarrassment when you extracted, you know, a named entity and incorrectly identified it in a real world sense, right? I mean, I'm going back to, like, you know, accidentally calling something new kids on the block. And this is how long I've been doing it, right? The owner Apple, like, had an album out of nowhere. And, you know, I tagged a news story that, you know, Apple stock was like, you know, because it was the album maybe wasn't that good.
It was horrible, it was embarrassing, it was such a struggle, it was highly stylized. It took weeks and months to make model improvements, and LLMs just they are just in. They know who Bob Jones is. If it's somebody important, give it a little hint, and they'll tell you. It has real world sort of a patterns and, you know, knowledge is not a right term, but you can understand. It has context well beyond a document or a paragraph or a sentence where your traditional NLP was trying to glean insight, right?
And those types of tasks on document processing. And now the hyperscalers are are bundling a lot of these capabilities and kind of launching services. I wouldn't build those, those are getting really, really good, you know, things like summarization and, uh, also traditional NLP tasks like translation. The those are getting really, really good through managed services, which obviously are built on LLMs. And you can, unless it's absolutely critical to your value proposition, those would be the obvious ones.
The next one is is image analysis. Now I did a whole PhD on this, and I I'm not saying they should take it away, but it's really been lapped many, many times. But what's curious is, you know, the progression of, you know, the early days of computer vision to where we are now. And there was this deep learning, conver, convolutional neural net. I mean, these were the basics that were developing fifteen, twenty years ago to deal with images, and they just built into these really, really large, powerful models.
The fundamentals are still there, um, but they just got so much better. I would not build that, I would definitely look for a managed service or something. So those sort of commodity tasks now, the differentiation obviously lies again when we when we think about, you know, moving up the value chain and into these jobs to be done. Your agentic workflows, what problem you're trying to solve, where do you want to establish, you know where you play, where your strengths are.
The other interesting thing about agentic is it may not just be one solution provider, it may not just be one, you know, company providing. You may have to create an ecosystem with different components. And so you have to decide, am I going to build this step of the workflow, or could I maybe partner or leverage another existing solution? I still need it as part of my key workflow, but I don't necessarily have to build this piece, right?
And that's another build by decision. I think building LLMs, that's pretty tough. It's a lot of money
especially in house.
you need to do it, but I think distilling LLMs and creating SLMs and creating SLMs you can deploy in in various, like maybe industrial capacities, or maybe in a healthcare setting on more of the the operational side and devices and things like that. That's all very, very interesting, and that could be something you absolutely have to invest in and build and do yourself. But I think, you know, also, you know, when you start building some of these things now, you own it, so model drift, you know, things like that.
Y- that's a life cycle, you know, you are continuously monitoring and making sure that it's behaving correctly, and you have to consider the overall cost of ownership as well. Yeah,
yeah, yeah, absolutely. Uh, throughout both conversations we've had on the show, very evident. You know, you've seen a lot of change over even the last five years, the last ten years going go. You're at the ground floor of, uh, n uh, nlp academically, i in computer vision. W What have you seen change the least or more stay true, and especially in terms of AI and long-term AI deployment?
you know, the first one, and it's just like a lightning bolt. It it is cybersecurity, privacy, you know, the more these like tangled webs we weave, the attack surface goes up, right? You you start talking about, you know, how do you identify the identity of an agent? How do you know who you're talking to? How do you know this system's agent should be talking to your system's agent? And that is requesting technologies like, um, agent to agent, mcp.
They're allowing us to think about, you know, building ever increasing, ever larger ecosystems, again, focused more on workflow. But there's this sense of like, well, how do we maintain that security posture that we're used to in the past? Where we went from the data center, owning the network and the proximity, the physical servers, somebody had a key to the door, then into the cloud, where it was like, okay, you know, there was a little bit of a of a maturation there.
But but cloud has provided a lot of security in now agentic, so that's number one, right? And, you know, and and we're evolving quickly. A And also, obviously, I mean, you can't overlook the fact that the LLMs themselves, this technology, is allowing the threat actors to scale up their attack vectors, hide them better, come up with new, interesting ways that your typical, you know, scans and penetration tests that you might not see, and that's also unique and interesting.
And it has changed, but this is my favorite. What doesn't change, you know, what doesn't change is when you have your application with eleven screens and all the buttons. And I I tell my team all the time, you know, how many more buttons do you think we could put on that screen, right? Like, in a sense of like, what are we solving for, right? Behind all those buttons are APIs, and behind all those APIs are likely, you know, you know, you know, crud and databases and all this transactional processing.
When you had, you know, all these screens and these buttons and human operators and people were going through workflow, the volume of traffic was almost somewhat metered by how quickly you can press those buttons. Now, when agents are doing the work, there's no more friction. The rate and speed and volume that they can send to your probably similar back end through, you know, either through, you know, mcp servers, or, you know, all these new, um, tool integrations to llms, you better be, you better be ready for that.
You better be ready for the volume of traffic that can spike in ways that you didn't see before, when it was a lot of human operators using SaaS products. And so the notion of, you know, having a robust API and microservice strategy is still very, very operative, it's still very, very important. On a certain foundational level in your IT stack, you are going to have databases, you are going to have transactional processing. It's really sort of in the upper echelons where these are orchestrated and stitched together.
But that foundation, that IT foundation that we built for twenty years, that doesn't go away, in my estimation, right? It's really, really critical. In fact, there's probably more focus to make sure your IP, uh, uh, APIs are at the right level of granularity, that your entitlements are at the right level of granularity. So you're not just sending the whole payload back to the agent, you're getting the specific information that you can audit and ensure, um, least privileged access.
These concepts now are really alive and well on almost another level, even though they've been with us for a long time, right? So, like, not everything changes, right?
yeah. And I I think that's gonna be comforting for folks on some level, even if those places are are really, really challenging and and giving people I a work
think, Matthew, we still have outages, right? yeah, all the time, Yeah, who? really. Uh, You know, you still there's no have firewalls, you still have your CDNs, you still have your reverse proxies not working. And if it was an application, it was a button, you might be able to track it down quicker. Now that it's kind of abstracting these agents, it's a bit of a challenge on the operational side. Where did the spike from come from?
where did this error come from? The observability, the monitoring, it's a challenge, so you have to adapt. Yeah,
yeah, and and that makes sense. I mean, for the first part of your answer, you sounded more like our our retail folks, uh, and guests for a second, just in terms of, oh, hey, we got the arms race. There's, you know, fraud, you know, uh, bad guys will never leave us, you know? And it a lot of the same thing applies, uh, to, you know, the healthcare space. Might not be n nefarious necessarily in the same way, but you're always gonna be having to deal with in any security factor that people who wanna do harm or things that can do harm will get more complex along right along with the technology.
And that's that's not a a watch that's ever gonna go away anytime soon.
I mean, HIPAA, SOC two, ISO, all these standards and compliance, they're interesting with LLMs, right? I mean, you're using some of these LLM, so, you know, if you're going to get a BAA in a healthcare setting, like, how do you know? It's fully integrated and aligned that everything is still 100% auditable and compliant. This standard, it's changing, it's evolving, and we're adapting as an industry and learning, but it provides new challenges.
yeah, a a absolutely. And as we get closer to more and newer challenges, even if they fit in the old paradigms, uh, or or lead us to problems that will never go away, that arms race, uh, I think we're. It's gonna be very, very important to have insight like this, like we've had for the past two episodes.
actually, Matthew, one more monolith Yes, please, bad for agentic, exactly, monoliths, monoliths yeah,
are bad for agentic. Now, what and do you mean by monolith? it's it's Um, you know, I mean, two
like, it's just this huge 1,001, black box, you know, large program that really doesn't have the right exposed endpoints, it doesn't have that sort of right granularity of entitlements. You know, when we used to build these really big, complex enterprise systems that are fractured, siloed, don't integrate and harmonize well. Very interesting, you know, you know, I've been doing, you know, working in the data, the data, s uh, data industry for a long time. You know, you you use terms like linking and concording records, or harmo- harmonizing data, or taxonomy, or or the semantics.
Like, you know, you have a record over here, I have a record over here. They need to talk to each other. How do you map these things, right? LLMs can help solve that interestingly because they they they can provide semantics. But you used to hear about Sparkle and RDF, and, you know, uh, they never really got the job done. In terms of semantically describing the data, harmonizing the data, providing that right level of data governance, LLMs themselves actually can potentially start to chip away at that and provide sort of better outcomes.
yeah. And which leads us to our classic problem. We'll have to have you come back to to talk about it. But uh, how do you pitch solving AI with AI? And I think we're getting better at that. But but, uh, I think, that's just like anything another one. Yes, else, yeah, yeah,
very interesting.
for and just, and just like the arms race, we'll need to, that'll be an evolving problem. We'll need to get better at our pitch as we see more black boxes and more impenetrable technologies because they're getting more complex as we go. Alex, really, really appreciate you being with us these episodes. Thanks so much.
great conversation. Have a great week.
Wrapping up today's episode, I think there are three key takeaways from our conversation with Alex. First, agentic AI only works in enterprise healthcare when the underlying infrastructure is mature enough to handle dramatically high task volume and system interaction. Second, meaningful performance gains come from domain adapted reasoning, where trained professionals help decompose tasks and shape dynamic chains of thought that make agents reliable inside regulated workflows. And finally, despite rapid advances, core operational realities remain. Enterprises must manage model drift, maintain compliance, and ensure systems are modular rather than monolithic, so agents can safely execute across interconnected workflows.
In this episode, we cover agentic AI in healthcare, to go deeper on AI topics and learn how to evaluate AI vendors by assessing leadership expertise and why funding benchmarks can signal product maturity and stability. Download our free PDF report five ways to select the right AI vendor at emerj.com slash aiv1. That's M E R J dot com slash aiv and the number one. For further executive level analysis and to join our network of leaders delivering workflow impact with AI, visit emerj.com.
On behalf of the team at emerge, we'll see you on the next episode.