Using AI to Responsibly Accelerate Clinical Trials Under Regulatory Scrutiny

Ryan Muise — Host, XTalks
Today's talk is entitled "Using AI to Responsibly Accelerate Clinical Trials Under Regulatory Scrutiny." My name is Ryan Muise, and I'll be your XTalks host for today. Today's webinar will run for approximately sixty minutes, and this presentation includes a Q&A session with our speakers. The webinar is designed to be interactive — webinars work best when you're involved. Please feel free to submit your comments and questions for our speakers throughout the presentation using the questions chat box, located in the control panel on the right-hand side of your screen. If you require any assistance along the way, you can contact me at any time by sending a message using the same chat panel.

Ryan Muise — Host, XTalks
I'd also like to welcome our attendees viewing from our LinkedIn live event — please submit your questions via the comments tab. All participants are in listen-only mode, and the event will be recorded and made available for streaming to those who have registered on xtalks.com. I'd like to thank Suvoda, who developed the content for this presentation. Suvoda is a global clinical trial technology company with a market-leading real-time software platform that empowers sponsors and CROs to make confident decisions, and sites and patients to take calm, controlled action. Suvoda delivers interconnected, action-driven software solutions and industry-leading services and support so that even in the most time-sensitive, mission-critical moments, life-changing studies keep moving forward.

Ryan Muise — Host, XTalks
Now I'd like to introduce our speakers. Priyanka Sharma is Senior Vice President, Software Engineering at Suvoda, responsible for all eClinical and eFinancial product development, testing, and site reliability initiatives across the company. Prior to joining Suvoda, Priyanka oversaw conceptualization, architecture, resource management, requirement analysis, strategic planning, development, and product launches for multiple business portfolios at IQVIA.

Ryan Muise — Host, XTalks
Jadon Sargeant serves as Senior Product Owner for Suvoda RTSM and Sofia, the AI assistant on the Suvoda platform. He champions innovation across mission-critical trial technologies, emphasizing reliability, teachability, scalability, and data security. With a master's in healthcare systems engineering, Jadon blends technical rigor with deep domain expertise and product leadership to deliver better trial experiences for sponsors, sites, and patients.

Ryan Muise — Host, XTalks
And Andrew McVeigh is Suvoda's Chief Architect, responsible for product software architecture and design. He also leads the AI initiatives across the company, focusing on agentic and RAG support for integrating Suvoda's clinical and financial data sources and streamlining processes. Andrew has extensive experience in software architecture, having served as chief architect at internet-scale companies such as LiveRamp and Hulu. Without further ado, I'll hand the presentation over to our speakers. You may begin when you're ready.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Ryan. I am Priyanka Sharma, and we are excited to have you all join us. I have this wonderful opportunity to moderate this session with my esteemed colleagues, Andrew McVeigh and Jadon Sargeant. Today we're going to be exploring the topic of AI in clinical trial technology. I know this is a topic we could talk about for hours, but today we'll keep it to around forty to forty-five minutes of discussion followed by questions.

Priyanka Sharma — SVP, Software Engineering, Suvoda
AI in the clinical trial industry is changing from a theoretical discussion to an operational reality. The stakes are all-time high, and there is no margin for errors. We're going to explore some hard questions today: What do responsible guardrails mean for our industry? What is the real value AI can bring? And what are customers actually telling us? Keeping the nuances of our industry in mind, let's first talk about why our industry has historically been very careful in adopting new technology. I'm going to ask Andrew McVeigh for his thoughts on this first.

Andrew McVeigh — Chief Architect, Suvoda
Thanks, Priyanka. It's a really good question. All of these AI technologies — and in particular LLMs, which have become so prevalent in agents — they're game changers. They really are. They're going to have profound effects on pretty much every industry. But they have certain characteristics that we have to be very mindful of when we're designing, because we're in a highly regulated and absolutely critical industry.

Andrew McVeigh — Chief Architect, Suvoda
The first responsibility is to the patients and the clinical staff. Whenever we design around these AI technologies, we make sure we keep the human in the loop for every critical decision. We're also very aware that the underlying LLM technology, and agentic technology in particular, can be non-deterministic, and we've experimented extensively with these before incorporating them into products. These form the backbone of our care to patients — making sure we design products very carefully around these characteristics. For instance, Sofia — although it's a conversational interface — at its core has a deterministic engine. We've been very careful so that hallucinations can't get through. These are very challenging design choices, because we want to make sure that at all points we have the human and the human experts in the loop. Sofia has a three-layer architecture which ensures that.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew. Quite insightful. Our industry is built upon trust and responsibility. However, what we're finding now is an accelerated rate of AI adoption — AI is at work for both non-clinical and clinical processes. What do you think is driving the industry to be more open and push for faster adoption of AI, Jadon?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
Good question. I think there are a few things really driving this. The first — and Andrew touched on this a little — is the transformational nature of AI in general. This is a technology unlike anything we've seen before. The closest thing is probably the invention of the internet. It's going to fundamentally change the way we both work and live our daily lives. We're seeing the real power of AI, and it's almost impossible to ignore.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
It's also a little different from some of the other "next big things" we've tracked. Direct-to-patient clinical trials from 2020 and beyond, or from the more technical side, data containerization — these are important trends, but they were siloed into specific parts of the industry. AI spans that gap and is everywhere: on the process side, the technical side, and in our everyday lives. A lot of technologies are limited to a professional setting, but AI is not. People use it to plan vacations, home renovations, and get their everyday lives in order. Part of what's driving adoption isn't just the usefulness we see, but also our familiarity with it. We're all experimenting with it all the time, learning its strengths, weaknesses, and use cases both professionally and personally.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
Beyond its power and our familiarity with it, we're also getting really good regulatory feedback now. It's taken a little while, but in January 2026, we finally got joint guidance from the FDA and EMA — ten principles covering best practices around PHI management, mandatory reporting, and the audit trails needed when using AI in this setting. Now that we have that codified, it's been a lot easier to start implementing AI products, because there's a lot more comfort in the industry about what will generally be accepted. I think 2025 was really the year of AI pilots — everyone was experimenting and trying out new tools. And now in 2026, it's the year of implementation. Everyone has run those pilots and is saying: now is the time to put these things into action and see what real benefits we can get at broad general availability.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
We've also seen a huge increase in capability, especially in the agent space, in the last year or two. Two or three years ago we were talking about context window sizes of 2K tokens. Now we're talking about a million — I saw something recently where they expanded to twelve million. There's a huge potential for efficiency in clinical trials: taking the burden off site users, off patients, and automating study setup in the first place. There's a huge amount of potential for automation.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Well said. As the years have gone by, the tech has matured and the pilots have shown results, so trust is being built. It feels like now is the time to look into real use cases and see how AI can be used as a productivity layer. No more experimentation phase — time to get to work.

Andrew McVeigh — Chief Architect, Suvoda
And we know enough about its characteristics now. If you're using something like ChatGPT as a general chat interface, it pushes any hallucinations up to you and you can deal with them — you know the document, so you can correct it. This is where we have to be very careful designing products, because we don't want to push that uncertainty up to users. That's where we always have the human expert in the loop and use a deterministic core where needed. On the subject of capabilities — I've used ChatGPT in the last six months to navigate a very complex legal situation with my home insurer. We had a claim, and we navigated it successfully using ChatGPT. The capabilities of these agents are just incredibly good. There's a massive potential to revolutionize how we work, make us more effective, and remove a lot of the daily grind.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew. With the complexity of our industry, it's upon us to use this technology wisely and do something good for human society. It's great to see AI adoption accelerating across organizations. That leads me to the next question: How would you describe Suvoda's overall philosophy towards AI? Any thoughts on the guardrails or design principles you apply when deciding where and how to use AI? I'll pass it to Andrew first, and then Jadon, please feel free to add.

Andrew McVeigh — Chief Architect, Suvoda
We have a principle of human in the loop. This is designed to augment people, not replace them. The human, by being above this and using it as an intelligent assistant, gets a lot of benefit because they can delegate tasks to it. In the case of Sofia, we have a deterministic core. There's a conversational system, but what it's doing is mapping onto a large library of questions it knows how to answer deterministically. We think at all points about the level of abstraction and the level of detail we're allowing the LLM to access.

Andrew McVeigh — Chief Architect, Suvoda
In theory, we've got thousands of functions inside our RTSM product. RTSM — randomized trial supply management — is our drug logistics system for clinical trials, and the evolution of our earlier award-winning IRT product. We could have just let the agent loose on all those thousands of functions, and we don't, because it needs to be taught like a subject matter expert. That's essentially where a lot of our effort has gone: having that authoring loop where subject matter experts like Jadon can teach and script the system so it has deep contextual knowledge. Then we lock it in a deterministic way — that's the inner loop. And then we have the outer conversational loop, which allows it to remember preferences and surface insight on different questions. But essentially, when it goes to answer a question, it's doing something it knows how to do and can guarantee the result.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
One of the big takeaways is that if you just unleash an LLM on a big trove of data, it'll make interesting connections and you can get a lot of good information. But it doesn't have those years of experience that your trained staff have. What we really want to do is combine the breadth of knowledge we've accumulated over decades in the clinical trial space with something that can synthesize it instantaneously. We don't want it to start from scratch and learn on the fly. We have very specific protocols our support staff, protocol designers, and developers follow. If there's a specific challenge to investigate, we can have Sofia follow that script all the way from start to conclusion — much faster than a human can. The knowledge comes from the humans; the speedy execution comes from the AI.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew and Jadon. When I step back and think about Suvoda's overall philosophy towards AI, two words come to mind: trial wisely. We've had long discussions with Andrew and Jadon to figure out what trial wisely means when adopting AI. Does it mean continuing to find reliability within the inner and outer loop concepts Andrew described? Does it mean keeping it safe and secure — ensuring rules and permissions are enforced and no data is shared with LLMs? Or does it mean creating efficiencies in our workflows to ensure ease of use for our users, sites, sponsors, and patients? The answer is all of these things. And our three pillars around the AI approach — which were the foundation for Sofia — are reliability, teachability, and security.

Andrew McVeigh — Chief Architect, Suvoda
Zero-day retention — that's bread and butter for us. We don't train any models on user data or sponsor data. We're extremely rigorous on the controls we put around the AI end. Sofia is built on a bedrock of permissions so that you can't access data you're not meant to see. You don't want someone in a blinded role having access to unblinded data — that would be disastrous. Those permissions are built right into the core of the system.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
We started from the most complicated and most risky case and made sure we could validate that first. The number one thing we can't do is unwind a trial — we can't give data to a role that shouldn't be able to see it. That's absolutely critical for our RTSM systems, especially when integrating AI. The reason it can work for us is because we already have data protections and role-level rules in our system. If you're a site user who's blinded, there's a certain set of data points you're allowed to see, and data points you're not. When you ask Sofia a question, Sofia can only see the data points you already have access to.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
If your role changes, or if there's a different study where those data points are slightly different, Sofia's access changes accordingly. If you were to ask Sofia something that would unblind you, it will tell you it can't answer that question — and the important thing is, it's not that it knows those data points and is told not to tell you. It genuinely can't access them. We have a lot of hard stop barriers in place. This applies at the blinding layer, site association, and depot association. Even if you had access to multiple studies, Sofia would only access one study at a time, and you could never access another sponsor's data. Everything is built on top of this rock-hard foundation of data security.

Andrew McVeigh — Chief Architect, Suvoda
It has been quite interesting watching Sofia evolve. What I found really interesting is how much it can introspect on its own abilities. I'll go to Sofia and say I need to do something, and it'll say, "Well, actually, how about we do this first and then this?" It chains together fairly complex questions in a very powerful way. What's your view on the conversational layer versus the deterministic core, Jadon?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
We wanted the deterministic core so that when it executes something, we know it's doing it correctly. If you want to look up a particular piece of data or get a link to a page, we know it will do that correctly — it's never going to give you a wrong link. In terms of the conversational aspect: we have this knowledge base on the back end with all these prebuilt queries and questions, and we want to make sure users can easily access those even if they don't know exactly what's back there or the right combination of words to type. The LLM can direct them, ask a follow-up — "clarify what site you're talking about" or "did you mean this context or that context?" It knows what it can do and directs you toward results it can actually pull.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
We're basically trying to bake our cake and eat it too — we want that conversational aspect that's innate to these technologies, but we don't want the hallucinations and classic mistakes that LLMs inherently make. We want to harness the power and limit the downsides. The two-layer approach is how we do that.

Andrew McVeigh — Chief Architect, Suvoda
You've actually got a third layer, haven't you? You've got that authoring layer where a subject matter expert — who is not necessarily a programmer — can go in and describe a whole set of business processes they want to automate. It will then store that away in a deterministic fashion to be invoked as part of the conversation. Can you talk to that?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
What we want to do is take the training we already do for any new hire learning how to navigate, operate, or troubleshoot the system — and instead of training them every single time, we train Sofia once. For a query like "why did a drug shipment not send" or "how would I release this series of lots" — we tell Sofia: here's what you do, in this order, check this configuration, then do this. You teach it once, and now it just knows that forever. It also lets the expert verify the steps: we teach it what order to go through, and it can list those out.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
What the LLM does very well is interpret the user's intent. The user will ask something — it could be a bit vague — and the LLM will get to the heart of what they really need to know, then reference its knowledge base to figure out the results. The authoring step is also iterative: we look at what information Sofia still needs, find the expert in our company who knows how to handle it best, and have Sofia learn directly from them. We also monitor what questions are currently being asked of Sofia and use that as our guide. Someone's asked about something we hadn't considered? Sometimes it's something we can absolutely train Sofia on — we find whoever knows how to do it, train Sofia, and we've expanded its capabilities forever. Other times it's outside scope: "How do they forecast drug several tiers ahead of the RTSM system?" — that's not something we want Sofia to answer. Our system can be conservative: "I'm sorry, I can't answer that question — how about we explore these topics instead?" That applies to things outside the clinical trial scope for Suvoda, and also off-topic questions like the weather or a local sports game.

Andrew McVeigh — Chief Architect, Suvoda
Can you give us an example of a complex query you might issue through Sofia that would otherwise take a long time to handle manually, even as an expert?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
Let's start with "why did my shipment not send." Someone might be on a predictive resupply schedule, expecting a shipment, and instead gets a failed drug order alert. These alerts can be a little opaque because there can be planning considerations at play. You might have to email Suvoda and ask why the shipment failed, and there's an investigation: Is there enough drug at the depot? Were those units released? Is there an expiration issue? An address issue? There's a whole list of things to research. Instead of having a support staff agent go through that process, you just type into Sofia: "Why did the shipment fail?" And Sofia goes through all of those checks almost instantly — as instructed by an expert, in a deterministic way.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Very smart approach — and it sits right on top of Suvoda's patented no-code, low-code platform, which helps position it to execute rapidly. I love the concept of authoring, Jadon. It took us some time to figure it out, but it's all there, and it unleashes the teachability part we were all very particular about as we expand it to many other portfolios.

Andrew McVeigh — Chief Architect, Suvoda
That's a really interesting point. We're providing Sofia with a high-level vocabulary — under the covers, it's essentially the hundreds of questions it can answer and the actions it can do. With agentic RTSM — our way of configuring the system using agents — we're working with the vocabulary of our low-code platform: deterministic configuration steps and extension steps. A lot of agentic approaches say "we have thousands of functions, let's just give the agent everything." In our case, that doesn't work. It's like going to an intern or even a PhD graduate and saying, "Here are thousands of functions — we're not going to teach you anything about RTSM or the complex clinical domain. Just go for it." That naive approach just doesn't work.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew. That's a great segue. Understanding the approach we've put in for Sofia and what we're thinking for agentic AI in Suvoda — what have we learned from our customers and users about expectations, concerns, or must-haves for AI? And where do we see it providing the most value for our sponsors, sites, and CROs? Jadon, any thoughts?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
The first thing is the changing attitudes. It used to be that when we approached customers and said we're thinking about using AI, they'd say, "You better not use AI on our stuff — we don't want any data going near that at all." We'd offer reassurances around zero-day retention and non-training models, but it was an uphill battle. Now the first question folks ask is, "Can you use AI? How can you help us? How can you accelerate our trial using these tools?" Attitudes have changed a lot, and folks are specifically requesting these kinds of tools.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
In terms of the feedback we're getting: we've started out with a conservative approach, focusing on things LLMs are very good at initially, then slowly expanding into things that used to be human-only. Examples of things LLMs are innately very good at: documentation summaries, being available 24/7, and multilingual translations. The very first thing we built with Sofia was the ability to summarize information. You can ask a general question, and it gives you the answers and references the exact spot it got them from — the exact guide, the exact page number. Rather than a site user having to pull up a training document they read months ago, they just ask Sofia a general question and get the exact set of steps they should follow, plus the page number if they want to verify for themselves.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
We still have a lot of discussions around security measures. Even though folks are very excited about AI, they want to make sure it's being used correctly. We always get: "Is my data being saved anywhere?" No — we're not saving any data at all. It's all siloed individually, and we're not training any models. Anyone who uses ChatGPT or Claude in their everyday life might know that as you use it more and more, it adapts to you — it learns your tendencies and tailors responses specifically to you. Sofia does not do that, because we don't let it keep any data about our users. You can have a whole conversation, talk to Sofia for hours, and when you click New Conversation, it won't have any of that context. Every instance is a brand-new instance. That's critical in this industry.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
In terms of other feedback — we constantly get requests for new features. Right now, Sofia is essentially responsive: you ask Sofia something, and it responds, gives you links, makes charts. But what folks want is something proactive — a proactive agent that works for them. Things like: "Let me know when ten subjects have randomized," "Run an audit on my data and look for any irregularities — flag those for review." We're moving toward a proactive agent that can streamline your work and take away the busy work. Rather than having to log in to check something, Sofia sends you what you need to know when it happens. And because most folks are already fairly experienced with AI, the feedback is very informed — they understand the strengths and limitations. It's a fascinating time to be getting feedback from people who are already experts in the technology we just launched.

Andrew McVeigh — Chief Architect, Suvoda
Adding on to this — the agents are getting far, far better at following instructions. For our agentic RTSM, we can describe very complex scenarios, and it's far better now than it was even a year ago — by a factor of two or three times. The capabilities are only increasing. But what we've also seen is that our core design philosophies have been very robust and durable: keep the human in the loop, and build on a deterministic core. Those two things are really important for ensuring our responsibility to clients and patients — that they're building on something trusted.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew. To give everyone a perspective, RTSM is the next evolution of Suvoda IRT. Andrew, I'm glad you started to talk about agentic RTSM as a concrete example of Suvoda's philosophy in action. Let's explore further — what have you learned on the journey of developing the RTSM solution?

Andrew McVeigh — Chief Architect, Suvoda
These agents have taken a quantum leap forward in the latter part of last year. Before that we had applied them, but we hadn't had such consistent and spectacular results. What it means is the human in the loop will have fewer corrections to make and see less variability. We spend a lot of time optimizing prompts and building a prompt authoring loop using approaches like the genetic Pareto method — a way of getting better results than reinforcement learning on top of these models. Essentially, we're teaching the system how to configure IRT and how to extend it, which is a massive job. We've got a system that can cope with even the most complex trials and workflows. Seeing the agents being able to handle this across multiple models and coding assistants is something that is seriously possible now in a way it wasn't before.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew. As we develop agentic RTSM and Sofia, I also want to call out that we're learning a lot from our early adopter program and finding where the systems provide the most value and where guardrails need to be tightest. Our solutions are not being built in a silo — they're being built with the industry, on real sponsors, real studies, real feedback. That keeps us honest and helps us understand where we need to go deep and where we can proceed in a more staggered form. Looking ahead, Andrew — what does responsible innovation with AI mean for Suvoda over the next two to three years?

Andrew McVeigh — Chief Architect, Suvoda
As capabilities increase, we double down on robustness — we don't want to lose that. That's core to our philosophy: human in the loop. But we can automate more and more processes, essentially enabling people to have superpowers. Clinical staff don't want to be sitting inside an RTSM system moving around menus. They want to be doing their job. We're going to expand Sofia across all of our product range. And with data exploration, we may be a lot more free-form — working off a much more fluid vocabulary where that's appropriate.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
That's true. We're always trying to balance that deterministic core with the conversational aspect — that's the key push and pull. It's easy to just unleash an LLM on a lot of data. The hard part is the constraint. We want to make sure we're always working from a place where we know the constraints are still valid, and then expand capabilities safely. We never want to do anything that breaks our core concepts. We can always make it more powerful — it's harder to rein it in if you've untethered it more than you wanted to. The classic tech mantra is "move fast and break things." Ours is: be safe, be controlled, have security, and make incremental progress. We don't want to break anything in the clinical trial space.

Andrew McVeigh — Chief Architect, Suvoda
And this comes to the point — no one is going to vibe-code an RTSM system. That's just a given. So it's up to us to work out how to responsibly incorporate these agentic capabilities, and we're on this every day. If people take away one thing from what I'm saying: we're building on layers that are already validated and proven, and we're giving AI access to those in a responsible way — which also means allowing them to see the thought process. One of the things you'll see in Sofia is that when you ask a question, you can then ask, "How did you answer that question?" You can go through the chain of thought and reasoning. That's quite important. As a non-expert user of RTSM — even as the chief architect — I sometimes look at these queries and I'm amazed by how complex they are. It's just good to be able to see how it was done. It's not a black box.

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
For complicated things, you can ask Sofia to do it, and it can do it. For simple things, Sofia can do it for you and then tell you how it was done — so you learn. You can give someone a fish, or you can teach them how to fish. We want Sofia to be both: an aid for getting information, and a tool for learning how to better use our RTSM systems on your own.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Thank you, Andrew and Jadon. The key theme remains the balance between innovation, regulation, and responsibility — which we have followed in all our products. With AI as well, we are adding that same layer, making sure our agentic RTSM and Sofia continue to balance this responsibility, and we continue to expand AI across many products and personas in the Suvoda platform. When I think about what success looks like for AI at Suvoda, I think of reduced friction, better experience, and sustained trust. Technology like AI is a means to that — for our patients, our sites, and our sponsors. We started today by saying AI has moved from a theoretical reality to an operational one. I hope the thoughts shared today provided some guidance and make that tangible.

Andrew McVeigh — Chief Architect, Suvoda
Priyanka, you phrased it really well — it has moved from experimentation to operationalization, and this is not a fad. It's not going anywhere. It's just going to get more and more intense. I did computational linguistics at the start of my career, and I didn't think we'd see anything as sophisticated. The onus is on us to use it responsibly and carefully, while at the same time revealing the full power to users. This is the thing we think about daily.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Exciting times — times to use this in a manner that could make a big impact. More to come on what else is being built in our future sessions as well. Ryan, back to you.

Ryan Muise — Host, XTalks
Thank you very much, Priyanka, and thank you all for this insightful discussion. As we move into the Q&A session, I'd like to direct the audience's attention to the handouts module, where you'll find additional documentation related to today's presentation. The very first question wants to know if you could discuss the utility of your platform systems for RTSM and other operational aspects of clinical trials. Jadon?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
The Suvoda platform is the base for all of our different products: RTSM, eCOA, eConsent, and a whole range of financial products. Right now, Sofia is available specifically for RTSM, but we're in the process of building it out for other products — the plan is to have Sofia available across the entire platform as soon as we can. In terms of other AI initiatives, Andrew has talked about agentic RTSM — essentially the ability to build a trial using AI to reduce the timelines of that build. Our dual approach is both an external-facing AI assistant where you can talk to it directly and have it pull data and do requests on your behalf, and internal tools that will make our build times much faster and more efficient.

Andrew McVeigh — Chief Architect, Suvoda
How much of this helps the new user, the middle user, and the expert user?

Jadon Sargeant — Senior Product Owner, RTSM & Sofia, Suvoda
It certainly helps the new user the most — they're the ones asking system navigation questions: Where do I find this? How do I access this? What does this mean? Low-hanging fruit for an LLM. The more experienced user asking more complex queries can also be helped. For those, it may be something where they ask once and now they know how to do it. I use LLMs to streamline things that are tedious or to teach me how to do something. If I already know how to do something, I typically wouldn't have it do it on my behalf. But even then, it might take a lot of clicks — and what we're really targeting with the agentic direction is reducing click bloat. Maybe you have to activate a bunch of sites and make sure the depots are active and all the settings are set. Rather than doing that by hand, Sofia walks you through it and asks for the relevant information. We're tackling from a lot of angles: the teaching aspects, streamlining data acquisition, and making things that inherently take a long time go much faster through AI.

Ryan Muise — Host, XTalks
A lot of vendors are talking about AI in clinical trials right now. What is different about Suvoda's approach, and where are you headed?

Andrew McVeigh — Chief Architect, Suvoda
We're building an extremely dependable core. We have this no-code, low-code platform with all these deterministic primitives that raise the level of abstraction to a very high level. On the Sofia side, we're building on a very powerful deterministic core that is teachable. We're not in the wild west — we're building from a very responsible place. We've got a patent pending on both the Sofia side and the agentic side. We believe these are serious IP solutions to very complex problems.

Priyanka Sharma — SVP, Software Engineering, Suvoda
Adding to what Andrew said: we are demonstrating execution by building live study builds of real studies, not just slide presentations. We're also extending this across our platform. Agentic RTSM is currently in an early adopter phase, but we're looking at expanding it to eCOA, payments, and travel in future. And everything is grounded in the voice of our customer and our early adopter program. We're working with customers to identify the highest-value use cases and working through change management and training — because it's not just the technology that needs to be adopted. There's a lot more in the form of training and user sessions. We want adoption built in as a buy-in as part of our rollout itself, not as an add-on.

Andrew McVeigh — Chief Architect, Suvoda
Someone asked: for agentic RTSM, is a digital protocol required or preferred? What agentic RTSM has allowed us to do is distill the input documents into what we call a study definition package. One of those is an information-gathering document — about fifteen to twenty pages — where you answer a collection of questions about the protocol. We don't drop the protocol in directly. The design consultant sits down with the client and answers a set of questions. In more complex cases, we have an additional study information document we feed in. Rather than a three-hundred-page URS, the URS now essentially becomes an output of the process rather than the input. This separates the "what" from the "how." The agent standardizes the "how" and gives us a much tighter feedback loop for learnings at the core phase. Within an hour or two of dropping that in, you can be into an actual demonstration of the study.

Andrew McVeigh — Chief Architect, Suvoda
Someone's asking: is there a tool within Suvoda that can actively take over the job of site monitoring? Sofia currently handles complex queries and actions, and monitoring is on the road map. There are limits — we're going to stay within our core competencies: RTSM, eCOA, eConsent. We're not planning to replace the entire role of a monitor from the operations side. We want to help empower it. If you want to be notified when something occurs, or sent an email when reconciliation happens for specific subjects — that's fine. What we don't want to do is collect lab data we haven't previously collected just so Sofia can vet it. Our approach is to streamline and empower within our existing systems. Suvoda is not going to replace entire roles on the clinical operations side that previously didn't interact with Suvoda. If there's anything you're doing in our system, you can use Sofia to access it, summarize it, output it. But we're not an EDC system, so there will be limits to the types of data we have.

Andrew McVeigh — Chief Architect, Suvoda
To answer quickly: is agentic RTSM being used across more than RTSM? Yes — it's being used for eCOA as well. eCOA is a modular approach that can plug into the platform for visit schedules and configuration. We also have an agent that can take a PDF instrument and turn it into an eCOA study questionnaire. We're automating those aspects in a modular fashion across our platform portfolio.

Andrew McVeigh — Chief Architect, Suvoda
Someone asked how Suvoda gets permission for AI data processing. We have enterprise agreements with large providers — we run OpenAI models on Azure and use Amazon AWS Bedrock for cloud models. We have zero-day retention policies, so they don't retain any of our data. We get permission from the sponsor when required. We don't fine-tune models with customer data — it doesn't get into the learning side and doesn't leak into the parameters. We don't hold data longer than necessary with the large frontier model. We have a lot of guardrails in place to make sure we don't leak data.

Ryan Muise — Host, XTalks
Thank you very much. We have reached the end of our time. If we couldn't attend to your questions, the team at Suvoda will follow up with you. You will be receiving a follow-up email from XTalks with access to the recorded archive for this event. Please join me in thanking our speakers for their wonderful time here today. We hope you found the webinar informative. Have a great day, everyone, and thank you for coming.