Clinical Finance AI Agent: Get the "why" behind the numbers, in seconds

‍Jeff Jahnke: All right, we're four minutes past the hour. If anybody's late, they can catch up, so let's get started. Welcome, everyone. I'm really excited for today's webinar. We've been hard at work here at Condor taking R&D finance to the next level - to a place that, having been at Condor for six years, was just a figment of my imagination when I joined back in 2020.

But we are here now. We're in the AI era of Condor. We recently released an AI agent for clinical finance, and that is the purpose of today's webinar: to show you what we've been working on, how we've advanced clinical finance, and how you can understand not just what the numbers are, but get the why behind the numbers instantly.

Quick introductions. My name is Jeff Jahnke, and I'm Director of Customer Operations here at Condor. I see a lot of familiar names among the attendees - thank you all for joining. I'm joined today by Nim Fox, our VP of Product. Nim has been central to developing this new capability in Condor and has been eager to show it off, so I'm really excited to give him the opportunity to show you this functionality.

For those of you who might not be familiar with Condor, let me set the table a little. Condor is the financial intelligence platform purpose-built for life sciences R&D. We partner with the Big Four accounting firms, and we're trusted by leading pharma organizations such as Acadia, BridgeBio, Madrigal, Stemline, and many others. Our customers have seen 90% forecast accuracy, 70% faster month-end close, and up to 30% budget savings on change orders. And not to set ourselves up here, but I think these numbers are just going to go up as we introduce the AI capabilities we're so excited to show you today.

So let's talk about Condor's AI platform and how it works. First and foremost, the platform is meant to deploy easily without changing your tech stack. We're not replacing any of your core systems - your EDC, ERP, corporate FP&A tools, or CTMS. What we do is unify that data. Condor sits on a layer of its own and connects to all of those data sources, so you have a single source of truth that spans all of your operational and financial systems.

Then, within Condor, you can automate your workflows. A team of AI agents automates your forecasting, close, accruals, change order management, and vendor management. This lets you run your processes faster and more cost-effectively, and it gives you insights you haven't been able to get with more manual processes such as spreadsheets or other disaggregated systems.

One thing I really want to emphasize is that we at Condor are big proponents of human in the loop. Our agents surface a lot of exciting insights and can read your data and make sense of it on top of our R&D knowledge graph, but they always require human approval. We're not taking humans out of the equation whatsoever. This is meant to augment your staff and your processes.

And last but not least, with those automated workflows, you can get the why behind the numbers in seconds. One of our customers has been playing around with our agent and asked it some challenging questions. The agent came back with an answer, and the customer said, "Wow. I tried to dig up this answer myself, and it took me two full days of digging through systems to understand why our study enrollment was not what I expected and what the timeline impact would be. Your agent answered it in less than five seconds, and we came up with the same answer." For us, that was proof that this is a really valuable asset that can bring a lot of value to the market as a whole.

As I said, you get the why behind the numbers in seconds. Condor AI Insights surfaces patterns, flags risks, bridges context across functions, and delivers the why behind the numbers.

We've had a lot of webinars recently on practical AI applications. We've talked with folks at conferences, with our customers, and with people across the industry, and three main patterns have come up in every one of these conversations.

First, as I alluded to on the last slide, when budget, forecast, and actuals diverge, R&D teams can spend hours or days reconciling data across their ERP, CTMS, EDC, and IRT, then plugging those numbers into spreadsheets and manually crunching them to understand the why. That's a massive pain point. Understanding what the variance is takes time, but understanding why there's a variance is where the real value is unlocked - and it's also the part that really takes time.

Second, lagging decisions have real consequences, like continuing to fund an underperforming site or discovering a change order problem after negotiations are already underway. Teams are very busy with their day-to-day jobs, and teasing out patterns and nuances in data is time-consuming. From what we're hearing, having a tool that streamlines that exercise so you can make more meaningful decisions is paramount.

And third: how can AI help? That's why we're here today.

So let's talk about how we get the why behind every number instantly. First, data is scattered across systems. As you all know from working in clinical finance, R&D finance, and clinical operations, there are systems specialized for very specific functions or silos within your organization. Your CTMS tracks a lot of clinical information. Your EDC is great at tracking patient enrollment and the visits that have occurred at sites, along with all of the underlying visit information. Your IRT is good at tracking drug supply, which patients are enrolled, and what their statuses are.

Then you've got your ERP and FP&A systems, which are good at tracking the dollars, and your procurement systems tracking POs. But bringing all of that together into one cohesive picture is incredibly difficult, and that's why Condor exists.

Our solution is the Clinical Finance AI Agent. First, it's powered by Condor's proprietary R&D knowledge graph. A biopharma R&D knowledge graph is essentially a layer that sits on top of your data and understands all of the data relationships. It understands what a patient is, what a patient visit is, what a CRO is, what a lab is, what a blood draw is. It's basically a massive knowledge map that interconnects all the different variables of a clinical trial, so the agent can make sense of all your data and pick out meaningful patterns and insights for you.

The agent sits on top of that knowledge graph and deterministic math - and by that I mean math that is set in stone - and puts AI reasoning on top. It can explain what's driving a variance, run what-if scenario planning, and build the resulting model directly in Condor. The beautiful thing is that you get the answer in seconds, not hours. We're really excited to show you that today, along with several examples of the types of questions you can ask.

A couple of items before we dig into the demo. I want to be crystal clear about how Condor works.

One: we do not mix the numbers and the interpretation. Condor's calculation engine, its financial engine, is deterministic, meaning the logic is set in stone. We don't ask the AI to calculate the numbers for us, because that can lead to hallucinations and the other risks you often hear about with AI. In Condor, the AI sits on top of the rock-solid foundation of a deterministic calculation engine, reads the results that are calculated automatically, and presents you with insights, information, and answers. It's the why, not the what. The what is deterministic. The AI does not create the numbers, and I cannot be more clear about that. How AI interacts with your numbers is one of the biggest questions we get, because in our industry, accrual calculations, forecasts, and the numbers themselves are incredibly critical. We are not leaving those numbers up to AI to calculate.

Two: humans are always in control. The AI does not take action on its own. It drafts, and you decide. For example, if you ask it, "What would happen if my enrollment slowed down 30%?", it will explain the cost impact based on the deterministic calculation engine. But it won't automatically update your corporate forecast in your FP&A tool until you say, "This scenario makes sense to me. Please update my corporate forecast."

Three: security is built in, not bolted on. The AI models don't have direct access to your data. Nothing is stored, reused, or used to train the models, and everything the AI performs with your permission is logged in an audit log, so every output is traceable and can be audited.

Nim Fox: Maybe I can add something on that note, because I think it's a really important point Jeff is making.

Jeff Jahnke: Yeah, that would be great.

Nim Fox: People ask us whether their data is being used to train the AI, and we say no, it's not used for training at all. The fundamental reason is that we don't need to train it. Training AI on your data essentially means reverse engineering from your data what the relationships should be, so the system remembers them next time. But we spent many years building our knowledge graph, and a knowledge graph is much more than just stored relationships. Because the knowledge of how everything connects is already there, we don't need any training data. Your data can simply be shown by the AI agent in the way that works for you, with the additional context. You don't need to train it, and you don't need to feed it.

The other part is that a robust knowledge graph, besides showing all the relationships and giving the models context, lets us avoid the hallucination problem. Hallucinations start when a large language model stops understanding the data but is eager to fill the gap with something. When you take that out of the equation completely and rely solely on our own calculation engine - which you can see on the platform and change on your own, playing with different parameters - we can focus on getting you insights faster and helping you get your job done faster, without you worrying about AI making up numbers or your data being used for training.

Jeff Jahnke: That's a great explanation, Nim. Thank you for jumping in. Enough slides - I think we're all here because we want to see the agent. So let's see it in action. Nim, take it away.

[The live demo is not included in this recording. To see the Clinical Finance AI Agent in action, book a personalized demo at condorsoftware.com/book-a-demo.]

Jeff Jahnke: All right, a few questions that aren't specific to the scenarios the agent ran. One is from Sumeet, and Nim, I'm going to need your support on this one: "Can you clarify what you mean when you say your AI models don't have direct access to our data? Aren't you connecting to our data sources via MCP and accessing them?" I think the answer is nuanced, but Nim, take it away.

Nim Fox: It's a great question. We are already connecting to your data in Condor today: your EDC data, your CTMS if you have one, your data lake, the recently released investigator grants reconciliation suite, CRO-reported payment data. The agent has access to all of that. What our AI model doesn't do is use MCP to go into your source system and start changing things there. We gate it, so it doesn't have direct access to the source system, because that's your source of truth. Now, if you save a scenario in Condor, have a bidirectional connection, and push that data through a user action in our UI, that is supported. But we gate that access - we don't put a Condor agent in your EDC system. That's what we mean when we say it doesn't have direct access. It does have access to all of the data you have in Condor today.

Jeff Jahnke: That's a really important distinction. It's not that Condor isn't accessing your data. It's that the agents are walled off inside the confines of Condor. If data lives outside of Condor, a Condor agent can't access it. Once it's in Condor, the agent can access it, read it, and apply the knowledge graph's understanding of the data to surface insights.

Lucas asks, "What kind of third-party data is used, and are you expanding to answer questions comparing numbers against market pricing?" Nim, are you able to speak to the third-party data?

Nim Fox: Happy to. Let me take a step back. We don't just show third-party data as a pass-through, for lack of a better term, because that doesn't give you a constructive answer. It might as well be a black box saying, "This number shows that." It doesn't take into consideration your performance and your context. The way we look at it is: what additional inputs can we feed into the model to give you an answer that you can also trace back?

In this example, you saw data pulled from ClinicalTrials.gov, which is a public source. We're also working on partnerships with private data sources, but on their own those just give you blank numbers. What we do is take that as an input and map it against your performance and your existing data. Then we can tell you very clearly - say we're talking about enrollment rate - this is your actual enrollment rate, this is your forecasted rate, and these are comparable trials and why they match: phase, indication, rate, site count, patient count. We show you the comparisons that are relevant to the question you're asking.

And yes, we're continuing to expand this, including toward fair market value. What we want to avoid is showing you a number and saying, "Trust us, it's a black box." We want it in the context of all of your data and site performance, and that can only be achieved with software like Condor that understands how your trial is structured and has the knowledge graph to show you that information in that light.

Jeff Jahnke: Awesome. We've got time for a few more questions. Marlon asks, "Does the AI agent come standard with a Condor implementation, or is it an additional service?" I'd encourage you to talk with your Condor engagement manager or business development rep. Marlon, we've worked together - happy to talk with you about getting it switched on for your instance. We're really excited to share this with as many sponsors as possible, both new and existing.

Lucas also asks, "What's the timeline to implement this AI solution in our system? One month? Two months?" If you have a trial that's already up and running, there's essentially no implementation time. It's a matter of switching it on, because the agent traverses your existing data set. If your clinical trial is already implemented, we'd run a training for you on how to use it, but it really is just flipping a switch and then getting comfortable with what kinds of questions to ask and how to review the answers.

If you're a brand-new customer, we run a full implementation process to make sure your data is ingested into Condor correctly and your clinical trials are configured. Once that's complete and you're trained on Condor, the agent gets turned on - at that point it's also just a flip of a switch. For anyone interested in a more detailed demo of the entire Condor platform, our implementation process, our services, and our features, we're happy to walk through all of that on a formal demo call.

All right, we still have a few minutes if there are any other questions. Going once...

Nim Fox: I think there are a couple more in the open questions. I can take them. Claire asks, "Can the AI give a more detailed summary of the forecast - for example, under investigator fees, can it break out components such as site fees, visits, and invoiceables?" Yes, absolutely. Reach out to your Condor engagement manager, and they'll show you how to use it with your instance and data. Because we're short on time and have a lot of questions, we can't go into every scenario, but yes, you can drill down much further using our agent. The data is there - it's about understanding it.

The next question is from Kelly: "Can you demonstrate something like what the cost would be if we halted enrollment at only 100 patients?" The simple answer is yes, you can just do that. But I think it's a great question because, while we keep emphasizing that the agent runs on top of your data, it doesn't just answer clarifying questions about your data. As with the patient-shift modeling we showed, you can throw whatever you want at it. It does that by understanding your trial data, using the actual calculations and the actual contract information you provide, and then projecting forward. We see a lot of sponsors plugging potential CRO rate cards into change orders to see how that models, or asking what happens if they stop. You can do all of this on the fly, as many times as you want, on the platform.

Jeff Jahnke: We also have a question: "Can Condor be used for a full, comprehensive forecast, or only for the portion related to clinical trials?" It depends on what you mean by a full, comprehensive forecast. Right now, the primary use case for Condor is R&D forecasting for clinical trials. We're not doing G&A forecasting, labor forecasting, or anything related to non-clinical activities. We're primarily focused on R&D finance. You can forecast your clinical costs and certain things like some CMC vendors, but for the most part it's focused on the clinical space. It's not a general horizontal platform that handles every aspect of a corporate forecast.

All right, I think we got through all of them. Going once...

Joe Roualdes: Hey Jeff, there's one more question, about generic LLMs: "What's the benefit of using the AI agent versus Claude or ChatGPT?"

Jeff Jahnke: That's a really good question. We've touched on it, but I don't think we've highlighted it quite this way: the primary benefit is the proprietary knowledge graph. Our LLM sits on top of a foundation of knowledge that deeply understands every aspect of clinical trials and all the relationships within the data. To Nim's point earlier, generic LLMs are eager to fill the vacuum when they don't understand your data or the deep context of your clinical trial, and they'll give you an answer that seems credible. What I've seen is that generic LLMs are actually pretty good at getting you the what, whereas our specialized agent really excels at giving you the why - and the what, for that matter. Nim, any other insights you want to add?

Nim Fox: Absolutely. I'll give a cheeky response and a full response. The cheeky response: let's both build a budget plan, one using Condor and one using a generic LLM, and see who does better.

The actual answer is that large language models are a great tool. They're a great statistical tool for guessing what the next word will be. That's extremely powerful for many things, including clinical trials, and that's how we use them. But in many industries, not just clinical trials, there's a realization that they're only one tool in the toolbox. Think about having Excel with all the formulas available, but without the knowledge of how to operate it, how to build a model, or how the data interacts. Without that, it's just a tool - a very strong one.

That's exactly where the knowledge graph comes in. It does two things that drive everything we do at Condor. The first is providing clinical trial context, and that is massive. It's what people retain in their minds over decades of work experience: how it all connects and how the knowledge maps together. Large language models can't do that on their own, which is why, when you ask them, they start drifting. When you ask Condor a simple question, it knows to answer in context, because it is fine-tuned to do that.

The second is that, without that context and without knowing how to traverse the data, large language models in their basic form aim to fulfill your request rather than tell you they don't know. That's where you start to see hallucinations, model drift, or questions that don't get answered. Combining them with a knowledge graph lets each part focus on what it's strong at. Large language models are strong at natural language comprehension, computational power, and the presentation layer. But everything underneath needs to be deterministic, otherwise it will fail you.

This is built into the core of how we do things. With or without AI, our math will always be deterministic. Combining the two gives you data you can rely on and agent answers you can rely on and use in production, without double-checking the agent's work all the time - so you can focus on decision-making. Less legwork, less correlating data, more decision-making, faster forecasting, and a faster understanding of what is actually happening in your trial, so you can focus your energy there. That's the extended answer. We don't do this instead of LLMs. We do it with LLMs, with the knowledge graph, and with Condor's proprietary models.

Jeff Jahnke: If anyone wants to see those differences and benefits in action, we've done previous webinars on exactly this, diving into specific topics like investigator grants reconciliation and change order analysis. They're all available on our website at condorsoftware.com. If you want to see how our agent compares to generic LLMs, that's a really good resource, and we also share practical tips for improving the output you get from generic LLMs. So it's not just a comparison - we also want to help you level up your skills with generic LLMs. I'd definitely encourage you all to check it out.

All right, I really appreciate everyone's time today. Amazing questions - thank you all for the engagement and for the feedback in the survey earlier. Nim, thank you so much for the demo. I know it's been a lot of work for the product and engineering teams, with great feedback from our customers as we've iterated on this. We're really excited to bring it out and get all of you using it. If you're not an existing customer, head to condorsoftware.com and book a demo. If you are an existing customer, reach out to your engagement manager and we'll get this turned on for you so you can benefit from it immediately. Thank you all for your time.

Nim Fox: Thank you.

‍