UiPath Fusion 2025 | Joseph Gannuscio, OSF Healthcare & John Sharpe, UiPath

Clips
More from UiPath Fusion 2025

Joseph Gannuscio

Director, Advanced Analytics Development

OSF HealthCare

John Sharpe

Sr. Technical Account Manager

UiPath

From RPA to intelligent agents: Healthcare automation takes on patient outcomes

Healthcare automation is no longer a matter of hypothetical efficiency gains — it is becoming a defining metric of patient care. Hospitals strained by rising demand are turning to intelligent systems that can interpret nuance, escalate decisions and act in real time instead of waiting for overburdened staff. The pressure for better outcomes is overhauling how health systems work with technology vendors. OSF Healthcare System — a nonprofit Catholic provider operating 17 hospitals across Illinois — has joined forces with UiPath Inc. to go beyond rules-based automation and into agent-based orchestration

play_circle_outline Introduction of guests Joe Gannuscio and John Sharpe at UiPath Fusion 2025.

play_circle_outline Empowering Healthcare Automation: John's Impact at UiPath and Strategies for Overcoming Challenges in Large Enterprises

play_circle_outline Key use cases for agents include optimizing medication refill fax processing.

Info
Transcript

Joseph Gannuscio, OSF Healthcare & John Sharpe, UiPath

Joseph Gannuscio

Director, Advanced Analytics Development OSF HealthCare

John Sharpe

Sr. Technical Account Manager UiPath

In this UiPath Fusion 2025 interview, theCUBE’s Dave Vellante and Rebecca Knight speak with Joe Gannuscio, director of advanced analytics development at OSF HealthCare, and John Sharpe, senior technical account manager at UiPath. The discussion traces OSF’s automation journey from early RPA wins during COVID to today’s agentic era – where LLM-powered agents augment brittle workflows and improve patient outcomes. Gannuscio shares how OSF evaluates agentic use cases, keeps humans-in-the-loop with UiPath Action Center and iterates quickly with stakeholder feedba... Read more

explore Keep Exploring

What is the purpose of the segment being discussed and who are the guests? add

What strategies can be employed to ensure that organizations like OSF Healthcare are addressing the right problems with automation? add

What is an example of how automation is improving the handling of medication refill faxes in a hospital system? add

bolt Powered by CUBE AI

Joseph Gannuscio, OSF Healthcare & John Sharpe, UiPath

search

Rebecca Knight

>> Good afternoon everyone, and welcome back to theCUBE's live coverage of UiPath Fusion 2025. I'm your host, Rebecca Knight, alongside Dave Vellante, my co-host and analyst. I would like to welcome two guests to this segment we have Joe Gannuscio, Director, Advanced Analytics Development at OSF HealthCare. Welcome, Joe.

>> .

Rebecca Knight

>> And next to you is John Sharpe, senior technical account manager at UiPath. Thank you so much for coming on, John.

John Sharpe

>> Great to be here.

Rebecca Knight

>> Yeah, so let's set the stage for our viewers a little bit. I'm going to start with you, Joe. Tell us a little bit about OSF Healthcare.

>> Sure. So OSF Healthcare is a non-profit Catholic healthcare institution based in Peoria, Illinois, which is Central Illinois if you're not familiar with it. We have 17 now hospitals throughout the state and really have been using UiPath for the last five years to help address efficiency improvements and ultimately improved patient care.

Rebecca Knight

>> Okay, and John, tell us a little bit about your role at UiPath.

John Sharpe

>> As a senior technical account manager I work with customers, large enterprise customers and help them grow their programs. And I'm finding that we back into a lot of that growth that comes from discovering what the business units care about the most.

Dave Vellante

>> And you're in from Hope Idaho, is that right?

John Sharpe

>> Hope, Idaho. A town of 200.

Dave Vellante

>> Wow.

Rebecca Knight

>> .

Dave Vellante

>> I didn't know that about you, John.

John Sharpe

>> On the lake.

Dave Vellante

>> I love Idaho.

John Sharpe

>> Yes. Oh, it's great.

Dave Vellante

>> A great place.

John Sharpe

>> Yeah.

Dave Vellante

>> Fantastic. All right, well let's get into it. So maybe you could start by sharing your journey in agentic, where did it start? Was it really scary at first, like going into a dark tunnel and then was there light in there? How did you deal with all that?

Rebecca Knight

>> No, I mean really agentic's been the natural evolution of our automation journey. So within OSF, we started with UiPath back in 2020, helped us get through some of the COVID challenges we faced, and we've just been building those blocks ever since. And so when we were here at FORWARD last year and heard about agents, light bulbs just started going off. And so we were immediately working with John, working with UiPath, how can we get our hands on this? Because we saw the potential. So it definitely wasn't big and scary, it was, how can we lean into this and really help push further leveraging the technology?

Dave Vellante

>> So what have you learned so far?

John Sharpe

>> It's kind of echoing what they said at the keynote this morning, proof of concepts are easy, implementation's hard. Because when you get into it, getting that first iteration out is really quick, but building these robust, reliable systems with something that's inherently non-deterministic can be a challenge, especially in the healthcare field where we have to be especially careful that the data that we're feeding into these, ultimately as patient lives, so making sure that we're putting the appropriate guardrails in, putting in the appropriate validations and making sure that we've got robust handoff mechanisms.

Rebecca Knight

>> So John, I want to come to you now because we heard on the main stage today, Daniel Dines talking about how companies need to think, big organizations need to think big, be ambitious, and as we just heard Joe say, the proofs of concepts are easy, but it does get a lot harder. These problems are big hairy problems. How do you work with organizations like OSF Healthcare to make sure that they are in fact a, trying to solve the right problems with automation?

John Sharpe

>> Well, I think one of the things that Joe said is, and you alluded to is that if the proof of concept, proof of technology doesn't get into production, the value meter only runs in production. And I'll quote Dez from Kaiser as the one who taught me that, it's a great thing to say. But you've got to look for not only what will get you into production, but big problems. And even if the problem is big and the barrier is AI council and all of these obstacles that can get in the way, I like to encourage customers to go for those big solution kit level problems that everybody knows, that relate to the CFO's metrics, get those out in front of people. If you get the business sponsor keyed in on that, everyone will come out of the woodwork and help you solve. Everybody wants to be part of a winning story, so that's where the magic happens is getting the right people to articulate the problem and solve for it.

Dave Vellante

>> John, when you go from high school sports to college sports, the coaches, they beat everything out of you that you learned in high school, they've got to retrain your brain. You have a lot of experience in RPA. Was that experience similar in that when you go to a dig agentic you got to kind of forget everything you learned or are you building on top of that because it's sort of automation related, but one's deterministic, one's probabilistic. How would you describe that transition?

John Sharpe

>> One thing I know, I'm very self-aware about what my strengths are and I'm a words guy and I'm a live guy. And so before programming, I know words and words to me are like math. And so I've taken to LLMs from ChatGPT 3.5 on, and there hasn't been a day that's gone by. And so by solving my own pain, that's what we tell our COEs to do is solve your own pain, if you can do that, you can articulate the problem in words that's like all of it. If you can articulate the problem, solving is easy. So I think yes, there is a transition, but if it goes back to, can you articulate the problem? If you can't, AI can help you solve it. I use AI to solve a lot. Go ahead.

Dave Vellante

>> Piggybacking on that a little bit though, they're actually complementary. Some of the first agentic use cases we got into production we're taking messy hairy RPA logic and putting an agent in place of that because the agent can work through things in ways that a flowchart can't. And we've found that we've been able to really increase throughput on some of those RPA processes by infusing them with agents.

Dave Vellante

>> Okay, so you had the-

Rebecca Knight

>> So-

Dave Vellante

>> Sorry, Rebecca, so you had the process nailed with RPA, but then you've just enhanced it and optimized the code, if you will?

Dave Vellante

>> Yeah, it was never quite hitting its true potential because there were too many edge cases and the agent can really help. When you've got the LLM, it can work through that variability much better than you can through a rigid flowchart.

Rebecca Knight

>> So what are the processes that you are using the agents for and how do the agents orchestrate people and robots differently than other systems?

Rebecca Knight

>> Yeah, it's really interesting. So we've got a... One of the ones we have in production that I'll talk about is around processing our medication refill faxes. This is the one I was just alluding to where there was really messy hairy logic because there's a ton of variability, every pharmacy's form looks different, everything's... It just drove brittleness into the automation. With the agent, we're able to handle all of that very effectively. And then the human piece of it, UiPath just gives you the great tools with action center, we can just hand it off, escalate it, get it tackled, and then it picks right back up and continues. Having those tools really make us able to comfortably and confidently build these processes and keep the humans in the loop where they should be.

Dave Vellante

>> So paint a picture, you've got 17 hospitals in the network, is that right? So lots of referrals, millions of patients. How are you serving those constituents or even how do you envision serving them differently in this agentic era?

Dave Vellante

>> Sure. I think it's really an amplification. We can do so much more now than we could a couple years ago. Some of the use cases we're looking at like the denials proof of concept that we worked on or some of the referrals management stuff we're about to start on, you couldn't even approach this problem two years ago without an LLM in place. And so really to me, it's such a great enabler. It empowers us to be able to build more, solve new problems in ways that were previously impossible with this technology. It just really helps us expand our impact.

Dave Vellante

>> Do you have to put new infrastructure rails in place to be able to take advantage of that increased, that optimized performance and procedural?

>> I mean, it definitely puts a premium on your data structures. Where's the data coming from? Because context is king when it comes to building agentic systems. Garbage in, garbage out still holds true. And if you don't have that context, I mean we've all experienced it, ChatGPT will connect the dots itself and it'll just make stuff up and it'll be confident in that and in the healthcare field, that's especially dangerous. So yeah, making sure that we've got our data lined up, we've got good context, we're engineering it for the problem we're trying to solve is really the critical addition that we've seen as we've moved into this.

Dave Vellante

>> So is that fundamentally a sequential process? Start with the data, make sure you've got the data at least harmonized and organized in a way, or at least narrowed it down enough so you can apply the AI, right?

Rebecca Knight

>> Yeah, what we've found so far is it's iterative. You start with what you think you're going to need, do evals, get feedback, find, "Oh, we missed this, we need more data." And you keep refining that. And then the key is really getting that feedback loop as tight and quick as possible so that you can rapidly cook through this.

Dave Vellante

>> You know how when you talk to the big training foundation model companies, they do training and they do post-training and they do reinforcement, learning with humans, and it's quite a complex process. Is there a mini version of that now in the enterprise? Are humans improving your localized models?

>> John, you might be able to better take a crack at that. You've got a broader stretch of the industry.

Dave Vellante

>> Great, thank you.

John Sharpe

>> One thing we are doing at UiPath is using feedback loops, one that come out of action center where the human in the loop process reinforces what we want to happen. So we've been doing that for a long time with machine learning, intelligent document processing, but then in addition, you can train LLMs, you can't train a large language model. You can give it additional information and reinforce the path that you want, and so that's how we affect the outcome with feedback loops is through adding that information to a database and then using that as collateral or an artifact to influence the decision.

Dave Vellante

>> And it's not just thumbs up or thumbs down, it's actual sort of context .

John Sharpe

>> It's kind of like, yeah, what was the context, what was the question, what was the outcome, and what did the person ultimately decide can influence in the future, yeah.

Rebecca Knight

>> I want to ask about, I mean, as you said, Joe, these are people's lives at stake and this is really sensitive private information, their medical histories, their medical data, what kind of medications they take. How did you make sure that the clinical stakeholders were comfortable with this?

>> Yeah, it was a process, but it really goes all the way back through our whole journey, building that trust with them. So we meet regularly and can incorporate our compliance, privacy, all of those folks in all of our builds, both before we start and then as we're progressing just to make sure that we are checking all the boxes they need to see us check. So it's really been the journey over that whole course. I mean the first automation we launched, I think it took us almost four months to get through all the approvals, but we've been able to work with them and understand, what do you need to see? Making sure we're building in a way that's accounting for all of those and then keeping them engaged throughout the build helps us make sure that we're doing that in the appropriate manner.

Dave Vellante

>> So how do you get those approvals and validating things like edge cases for HIPAA and PII?

>> Yeah, it's definitely something that depends on the build you're working on. So depending on the project, we may take samples, have it be reviewed, we may have compliance watch the whole thing, run it all the way through before we go live. It really just kind of depends. I don't know, John, you have anything you've seen from other?

John Sharpe

>> Yeah, one thing you find is that people that work in compliance tend to be more detail oriented by their nature and they tend to be very thorough. Another thing that you can layer on top of that though is use an LLM to embellish, like here's some examples of things we want to test and it will generate valid and you of course guide this process of generating what should be checked. Like, "Here's what we want to do, give me some more examples. Here's three things we want, give us nine, give us 27."

Dave Vellante

>> And we all have experienced LLMs. You ask a question and they ask the same question and you get a different answer. It's not radically different always. Or if you ask it to edit a picture, you know what I'm talking about. So when people talk about guardrails, is that the type of guardrail that you put in place? Can you give us an example of how you actually apply that or maybe how UiPath helps with that?

Dave Vellante

>> Sure. So really what we've found is it's all in the system prompt. As you're engineering that system prompt, you need to make sure that you're spelling out specifically what it needs. And where UiPath really helps with that is the scoring it gives you in Agent Builder. We lean heavily on that to help make sure that we're getting appropriate guardrails put in place, making sure that our prompts hit the standards that UiPath's found are needed for that. So we lean heavy into using the tools that are available in Agent Builder and have found they've been super helpful.

Dave Vellante

>> Did you have a follow-up on that?

John Sharpe

>> Yeah, so we also have a feature in Agent Builder, which is called Guardrails. And so here's an example of something that can happen. Say that you have emails coming into a customer service email box and someone is able to come from company XYZ and say, "What are my invoices? Can you give me the status of this invoice?" Well, what if a bad actor at ZXY comes in and says, "I want to know all the XYZ." You want to be able to have guardrails so that you can't have what's essentially prompt injection across. And this is where you've seen in the recent news, UnitedHealthcare had a refund happen. This happens all the time. And so that's what UiPath guardrails are for. So part of it is prompt engineering, and then additionally, especially with external facing inputs, we want to make sure that we add those guardrails, which are essentially large language model calls.

Dave Vellante

>> Yeah. Give me a paint by numbers, how to build an agent. Where would you start? What advice would you give to a novice?

Rebecca Knight

>> I mean, from my perspective, and then we can get yours, John, start with, like John said, what is the problem you're trying to solve? Because if you're not anchored to that, you're not going to be successful. You also need to know who do you need to get input from. And then the way we've been coming at it is get a version in the hands of the stakeholders quickly, not the final version but a version so they can start reacting to it and giving you that feedback. While you're doing that, that helps you then frame up, how do you pull together the context you need, how do you get your prompt rights. Having that rapid feedback loop, it's important in software development in general, but we've found at least with hi agentic builds, that it's all the more important. And the tooling we have really lets you get something in the hands of the business quickly. As long as they understand and you understand that's not the final result, that's the first stepping stone, but then that can really guide your development as you go.

Rebecca Knight

>> Can you talk a little bit about the role of trust in this process? Because I'm interested in how you said your advice would be to get even an unfinished product into the hands of stakeholders quickly, even if it's not perfect. On one sense that's risky because they're going to say, "This is horrid. We still need so much done to make this real." On the other hand, it maybe can help build the trust that you hear them, you're listening to them. Yeah.

Dave Vellante

>> I mean, that's been my experience. You've got to lean into your relationships with the people you're working with, but that's how you get the best products. And once you get a couple of those under your belt, the whole organization starts to get on board. That's been our journey really with automation from the beginning. We started in revenue cycle because we had one very engaged person. We leaned in with them, and then before you know it, we've got requests coming in from across the ministry, just every area, "Oh, I want a piece of the automation. I want a bot. I want a bot." Then they're naming their bots, and so it's just a snowball effect that's really a... But it really all starts with building good relationships with your key stakeholders.

Dave Vellante

>> What's your favorite bot name?

John Sharpe

>> Teddy.

Dave Vellante

>> Teddy.

Rebecca Knight

>> Aw, Teddy bot. Good bot. So what's next? What's next for this relationship between OSF and UiPath? What's next for OSF in this journey?

Dave Vellante

>> I want to learn as much as I can from you guys and really just I've appreciated the collaboration we've got. I feel like every time John and I talk, I'm learning new stuff or coming away with ideas that I can take back to my teams and then use to really amplify our program.

John Sharpe

>> Something we've been talking about is IXP, which has been a popular topic today. That's prompt-based model creation for intelligent document processing. It's really powerful because you can use documents in all variations that they come in, imagine invoices or receipts from different restaurants. You can infer, just like a human would, which the word infer just means to deduct or deduce what we're looking for. So that's why IXP is so powerful is it allows you to infer the information that you want even if you haven't seen it before, you can get it.

Dave Vellante

>> And you guys bought a company a few years ago called Re:infer.

John Sharpe

>> Re:infer, absolutely.

Dave Vellante

>> that name.

Rebecca Knight

>> Yeah, I'm also a words person. Words are like math to me as well.

John Sharpe

>> I love words, yes.

Rebecca Knight

>> Yes, yes. Yes. Excellent. Thank you so much Joe and John, a pleasure having you on the show.

Dave Vellante

>> Same, thank you so much.

Dave Vellante

>> .

John Sharpe

>> Thanks for the opportunity.

Rebecca Knight

>> I'm Rebecca Knight, for Dave Vellante, stay tuned for more of theCUBE's live coverage of UiPath Fusion. You're watching theCUBE, the leader in enterprise tech news and analysis.