UiPath Fusion 2025 | Graham Sheldon, UiPath

Clips
More from UiPath Fusion 2025

Graham Sheldon

Chief Product Officer

UiPath

Agentic workflows take center stage in enterprise transformation

Agentic workflows are moving automation into a new era, shifting from rigid rule-based tasks to adaptive, goal-driven processes that respond to real-world complexity.This shift is more than just a technical upgrade — it’s redefining how enterprises think about efficiency, scalability and trust. Rather than settling for incremental improvements, organizations are embracing ambitious strategies that blend deterministic automation with adaptive, agent-driven intelligence

play_circle_outline Rebecca Knight and Dave Vellante host theCUBE's coverage of UiPath Fusion.

play_circle_outline UiPath’s Evolution: From RPA Tool to Comprehensive Automation Platform Through Strategic Acquisitions and Advanced Technologies

play_circle_outline Importance of maintaining human oversight in automation for compliance and trust.

play_circle_outline Introduction of Maestro, a control plane for orchestrating automation processes.

play_circle_outline Accelerating Agentic Capabilities: Rapid Deployment and Ensuring Quality Control in Automation Solutions

Info
Transcript

Graham Sheldon, UiPath

Graham Sheldon

Chief Product Officer UiPath

In this UiPath Fusion segment, UiPath CPO Graham Sheldon joins theCUBE’s Dave Vellante and Rebecca Knight to unpack UiPath’s next chapter: evolving from rule-based RPA (“act one”) to goal-directed, agentic automation (“act two”). Sheldon explains how Maestro functions as a control plane to model, visualize and orchestrate end-to-end processes – pairing deterministic software robots with non-deterministic AI agents and humans-in-the-loop. He details how Maestro not only executes workflows but also provides insight to identify bottlenecks and continuously optim... Read more

explore Keep Exploring

What is the context of the live event coverage involving UiPath Fusion and the guests present? add

What are the key developments and capabilities that have evolved in the RPA platform over the past 20 years? add

What is the latest development in the platform that helps customers manage and deploy various automation capabilities? add

What advancements in the platform are being used to help customers integrate and automate their business processes? add

What are the key challenges faced when trying to reduce time to value in implementing robotic process automation (RPA) solutions? add

bolt Powered by CUBE AI

Graham Sheldon, UiPath

search

Rebecca Knight

>> Hello, everyone, and welcome back to day two of theCUBE's live coverage of UiPath Fusion. I'm your host, Rebecca Knight, alongside Dave Vellante, my co-host and analyst. We have another jam-packed day with a terrific lineup of guests.

Dave Vellante

>> Yes, we do. Good night last night, pressing the flesh with some of the peeps here. It was excellent walking around the sessions, watching Brent in his gold lamé. So, yeah, excited for today.

Rebecca Knight

>> Well, indeed. With that, I would like to introduce our next guest. Welcome back to the show, Graham Sheldon, CPO of UiPath. Welcome-

Graham Sheldon

>> Thanks so much for having me.

Rebecca Knight

>> Yeah. No, we're excited to have you back. So, I want to begin with a little historical perspective because UiPath really was born as an RPA tool, and through acquisition and organic development, it has transformed into a platform company. Can you talk a little bit about where we are right now with the portfolio and that transformation and evolution?

Graham Sheldon

>> Yeah, absolutely. So, 20 years ago, Daniel started out with RPA and really established the category. And as you mentioned, through acquisition, through building our own technology, we've added a lot of other capabilities to that. Things like understanding documents in depth with AI and machine learning, things like computer vision over time. And we've evolved from simple tasks to full-on processes over that. And things like task mining and process mining we've added. Communications mining we've added. And it all comes together in that act one, it's a full platform, a suite that help you do end to end automation really well, and that was our act one. And this last year we've added act two. What that act one was all about was that deterministic work, where you have folks that are doing rule-based automation. And now, the real opportunity is combining that with the goal-based automation that agents are now capable of doing. And so, we're adding that into the platform in act two and orchestrating all that work together.

Dave Vellante

>> So, as a customer, I've got this menu of capabilities. How should I think about how I deploy those? Can I just add capabilities, subtract capabilities? Is there an AI advisor to help me do that? Because it's becoming a lot of companies who turn into platform companies, I can now have a lot of options. How are you helping customers sort through that?

Graham Sheldon

>> Yeah, so some of the most recent advancements in the platform include Maestro. And Maestro is like a control plane, if you will, for doing this orchestration end-to-end and using the right capabilities for the right task. Oftentimes, you're right, Dave, you don't have to go and reinvent the wheel again. You can reuse the existing task automations or things that are even outside of the UiPath platform and ecosystem in order to orchestrate an end-to-end business process. And what we do to try to help them is provide Maestro as a way to model what that end-to-end process looks like to be able to automate specific pieces of it either with robots for that deterministic work or agents for the non-deterministic work, and human in the loop for when people need to make those final decisions. But you don't have to lift and shift a whole bunch of things or reinvent anything in order to take advantage of that. Maestro helps you bring that together and meets you where you are.

Dave Vellante

>> Especially if you have software robots in place, you have doers and you're adding thinkers, that are agents.

Graham Sheldon

>> That's right.

Dave Vellante

>> Maestro orchestrates or helps orchestrate that or at least visualize it, and then your platform helps orchestrate it.

Graham Sheldon

>> Yeah, Maestro helps you visualize it, for sure, to help bring everyone together to understand what the business process actually is, but it also helps you orchestrate it and execute it, and then optimize it over time. And that's, I think, a really important thing about Maestro as an engine. It helps you use these tools in a way that help you view how that process is running, gain insight and understand where the bottlenecks are. And then, over time you can either add more agentic capabilities or even change the process as it's running, so that you can get better business results.

Dave Vellante

>> Are we at the point where it's self-learning? Where as the business changes, it can adapt to those changes and then feedback, "Hey, here's a better way to do it"? Are we there yet?

Graham Sheldon

>> Not quite, but-

Dave Vellante

>> But that's where we're headed, right?...

Graham Sheldon

>> that's clearly where we're headed.

Dave Vellante

>> Awesome.

Graham Sheldon

>> Right now, it's very clear that humans still need to be in the loop for making those types of changes because I mean the other part of this, David, this stuff is running on a trusted foundation. Our customers are highly-regulated industries, across many different geographies and they want to make sure that they can trust these systems to do work. And agents are not necessarily always the most trustworthy. There's a lot of well-documented things where they hallucinate or make things up and we can't afford that in these critical business processes.

Rebecca Knight

>> So, I wonder if you could walk us through a customer example where Maestro is acting, as you say, as the engine, adding and subtracting agentic capabilities. I know you can't pick among your favorite, but you are CPO, so you have a good grasp of how companies are using this.

Graham Sheldon

>> Yeah, so there's lots of ones. Some of our AI25 Award Winners are great examples of that. SunExpress Airlines is one that I just learned about yesterday. That's a really exciting one, where they're dealing with several different processes, but the one that was most interesting to me is they're actually planning out where their personnel have to be deployed to have the hotels. They're staying over at a certain place, they've got to coordinate where they can stay, how long they can stay, make sure that their policies are met in terms of furloughs and things like that, to make sure that they're well-rested. And there's a lot of unstructured data, requests that may come in, "Hey, I want to stay somewhere different," or, "Ooh, the flight got canceled or delayed." There's a lot of these adaptive changes that unstructured data and those adaptive changes where agents can really help to understand what's there extract key information? Figure out what the right next step is. But accessing the actual data and being able to take action to book a hotel or to enter something into the systems of record is something that the robots are doing because it's very reliable. You can't screw that up. You want to make sure that it's following the right rules. And so, that's a good example of where they've had some massive success in being able to automate that process end to end.

Dave Vellante

>> Yeah, I was a judge in the AI25 this year. I'm just looking at my ratings of SunExpress. It's quite interesting. I gave them fives and fours. And if I understand this correctly, if I recall, they had a number of operations that were constantly changing.

Graham Sheldon

>> Correct.

Dave Vellante

>> And that's where the traditional RPA struggles and agentic can handle that diversity. And crew hotels and group sales and things like that were problematic with the deterministic platforms, and so they're applying agentic for that-

Graham Sheldon

>> And Dave, you can sit down and ask a developer and try to get all of the decision trees right and write really long logic code, but you're not going to think of all the possibilities and the weather-related problems that might crop up, or someone asks for a group sale where they combine two different groups together, or whatever. Those are things where adaptive and judgment calls need to be made. Now, agents are not going to do all of that, but they can provide these recommendations to people, so that they can make those final judgments and react in real time.

Dave Vellante

>> And over time, won't the agents learn from those exceptions and the reasoning traces of humans? Will that get fed back to the agents? Again, I know I'm probably ahead of myself, but that is the vision of agentic and the real productivity drive.

Graham Sheldon

>> No, you nailed it there. A lot of our latest announcements this week have been around how we have built-in memory, both short-term and long-term, so that agents can actually adapt and learn those processes. When you see an exception happen and you treat it in a different way, you can have someone say, "Hey, we want to make sure that this happens," and the agent will remember. Then, on the modeling side of things, our agentic IXP announcements, we are actually building fine-tuned models that will get better and better over time based on the feedback that they get. So, actually, that's in the realm of today.

Dave Vellante

>> Yeah, I've seen some rudimentary examples, but what you're talking about is even more sophisticated and that's just going to escalate over time, right?

Rebecca Knight

>> But do you always foresee this human in the loop? Because you are making this case and this is where we are heading, but in terms of building that trust, where does the human fall? I'm curious, Graham.

Graham Sheldon

>> Yeah, so the short answer to the question is we don't know right now, but over time we think that some of the more mundane things, the places where agents have proven that they can meet a certain quality bar, there are many of our customers. I'm thinking of Fiserv and CSL, where indeed... I'll take CSL is an example.

Rebecca Knight

>> And they're a plasma donation?

Graham Sheldon

>> Exactly, they work with plasma donors and they help save lives. But in order for that plasma to be the highest quality and to avoid some of the problems that can occur if blood gets tainted, at the end of that process where you're looking at someone's medication history and maybe some of the allergies they might have to qualify or disqualify them for donating, you really want a physician there to make the final judgment to sign off on it for the compliance and audit trail that's there. So, I think that humans will be involved for the most critical decisions for the foreseeable future. And over time, yeah, we're going to get more comfortable with the quality bar that agents can do.

Dave Vellante

>> So, as you know, customers are getting inundated with the hype of agentic. They've got their CRM system, they've got their IT service management and overall service management system. They've got their HR system. All of the SaaS vendors are blanketing them with agentic marketing and products. So, can you explain the advantages of a vendor-agnostic approach? How you talk to customers about that and what their reaction is?

Graham Sheldon

>> Yeah, there's two angles to a vendor-agnostic approach, like the one that we've embraced. Being open and flexible means that, first of all, you can stay on top of the latest and greatest stuff. This world is changing so fast. I had a number of conversations this morning with customers who were like, "Oh, my goodness, Graham, I know you have to keep up with everyone, but there's so much for me to learn here." All the models, there's no one model to rule them all. There's no one framework to rule them all. There's no one tool to rule them all. You're going to want to stay on top of that. So, that's one thing that staying open and flexible allows you to do. And we bring all of the best models like GPT-5, which just got announced earlier this year, so people can build agents with that. The second thing is it allows you to use where the work is getting done today. You don't have to go and change everything, so it gives you that freedom. We meet you where you are. If you have critical workloads that you want to involve in an end-to-end process, you can bring that in through our orchestration capabilities in Maestro.

Rebecca Knight

>> So, you were talking about how you are hearing from customers that say, "Oh, my goodness, there's so much here. Where do I start?" This does feel very daunting for so many organizations, large and small. Yesterday, we had Lake Michigan Credit Union on. Can you talk a little bit about UiPath solutions and how it helps organizations like Lake Michigan Credit Union move faster?

Graham Sheldon

>> Absolutely. So, yeah, some of the pragmatic advice we give to customers is like, "Hey, we want to think big. Start with something that is really a critical business process for you." For Lake Michigan, that was about their loan origination process and quality controls. And for them, that was a daunting task. It took a long time. They've got members coming in asking for loans. Took them almost 50 days to process those. And they wanted to go after that end-end process because it was really critical for them, but they needed to help. They needed to know how best to do this. And so, what we did was we worked directly with them to build out this solution in such a way that we could take all of the domain expertise that they have about how their business runs and our expertise with how the technology works, and we were able to give them a prepackaged solution that was more turnkey in the past. Dave, we started off talking about how the platform has got a lot of different pieces to it. Well, that horizontal platform is like a bunch of Lego blocks, if you will. People want to have the Millennium Falcon. And so, for Lake Michigan, we were able to give them a bespoke solution where their end users could go and look at all the loans that are there, be able to interact with those loans, inspect with the agents were doing and make those final decisions. So, it was a much more turnkey solution for them. They didn't have a big dev shop.

Dave Vellante

>> So, simplicity works in this industry. At the same time, the whole move fast and break things mantra is being rethought. Move fast and don't break things is the new mantra. I remember when Snowflake first came out, their marketing was, "Zero to Snowflake in 90 minutes," which I thought was very clever. Now, of course, we're talking about spinning up data warehouses in the cloud and shoving a bunch of data in there and making analytic sense out of it. Not super risky. With agents, there's obviously other risks that you want to consider. There's certainly legal and compliance and other potential risks as you were talking about, of rogue agents. So, for customers, they want to go fast, they don't want to have to do long ERP-like projects that go over quarters or half years or even years. Is there a path for weeks of getting to agentic and a journey to get there, and avoiding that half-year, year multi-year projects? What's the realistic weeks-not-quarters plan to get to agentic and how can we-

Graham Sheldon

>> Yeah, so Rebecca asked about Lake Michigan Credit Union before, and I think that may be another good example where it's easy to get a POC in place. POCs are straightforward, but you talk about there are two key challenges that folks face when they're trying to reduce that time to value. One of them, as you mentioned, is governance and compliance and security. So, the advantage of working with UiPath is that all of the controls, roles-based access and control, the governance, the policies, all of that applies equally to robots that you might already have in place, as well as the agents uses the same control plane and the same controls we're adding to that. Things like agentic guardrails and PII in-flight masking, things like that. So, that really, really helps. The second big challenge is on quality. And so, that's where a lot of the newest agent builder capabilities around evals and being able to test these agents on an ongoing basis is super important. And with some of the things that Raghu talked about yesterday, where you're able to simulate and mock some of the responses that come back, that really reduces the time to value. The truth of the matter is it's still going to take you a few weeks to make sure that you get your first pilots into production and then collect real data, so that you can be sure that you're going to meet whatever quality thresholds it is, the 90% or the 95% that you were trying to do.

Dave Vellante

>> Yeah, weeks is good. I mean to get started from ground zero, no agentic to agentic, weeks not quarters is, I think, a real positive. Can you talk about the ecosystem a little bit? How you're tying in? Where you guys made some big announcements this week with some names, like OpenAI and I mentioned Snowflake, and Nvidia, and Google. So, some pretty big names there that you're cozying up to. How do you see the ecosystem? How is it different now in this era than it's been in the past and where is it going?

Graham Sheldon

>> Absolutely. So, you brought up earlier the Switzerland approach that Daniel likes to talk about, and being open and flexible and the benefits that it provides. On the AI ecosystem side of things, that applies to models and to frameworks and to some of the agents that are there. So, at the model area, you mentioned OpenAI, we're collaborating really closely with them on computer use. So, we're going to be putting out a benchmark for enterprise applications. There's a lot been done for the web and for consumer use, but not as much for enterprise applications. It just doesn't work as well as it needs to. And so, with our screenplay technology that we just announced, we're going to be incorporating the latest and greatest from all of the frontier providers to make sure that we have the best computer use models for that. So, our customers can take advantage of it for the things that still need that UI interaction to be successful. We also are exposing GPT-5 as a way for you to build agents to take advantage of OpenAI's latest and greatest stuff, and we're incorporating an integrating all of the automations that you build into ChatGPT enterprise, so ChatGPT users can use it via MCP. So, that's the big set of announcements around OpenAI. For the agents side of things, we're working with Nvidia. We have a connector now. So, you can use NIMS developed models directly as agents. And then, for Salesforce and Google and Microsoft, we have the ability to basically orchestrate all of the agents that are built in things like AI Foundry from Microsoft or Google's Vertex agents in a Maestro process.

Dave Vellante

>> And Snowflake with Cortex, would that be part of it or is that something different?

Graham Sheldon

>> Yeah, and Snowflake is actually where we're going even further than that. So, Snowflake and their new Cortex AI are going to be modeled as data agents. So, those data agents allow you to be even more expressive in Maestro, and they're a special kind of agent class that will be able to pass more rich data objects around and be able to create even more powerful types of automations.

Dave Vellante

>> Connecting the agent dots.

Rebecca Knight

>> Indeed. Indeed. Well, Graham, always a pleasure having you on the show. We really appreciate you coming by.

Graham Sheldon

>> Thanks for having me, Rebecca. Thanks, Dave.

Dave Vellante

>> Thank you, Graham.

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.