Crucial Data Solutions
The clinical trials industry has absorbed a decade of promises about AI-assisted study builds. Most of those promises delivered marginal efficiency gains on the margins of a process that remains expensive, slow, and heavily dependent on manual labor. Crucial Data Solutions is making a different argument. Its AI agent, Floyd, does not just assist a human study builder. It ingests a protocol, designs the case report forms, and constructs the full study inside the company’s eClinical platform, TrialKit, in under ten minutes. The human’s job shifts from builder to reviewer. But the more consequential claim Paul Grady is making is about what comes next: synthetic data as a regulatory-grade control arm, replacing the placebo group in oncology trials where patients in the control arm frequently do not survive the protocol. That is not a feature roadmap item. It is a structural argument about how clinical research gets done. Grady, who has been building clinical data systems for over 3 decades, sat down with Vanguard to explain where Floyd operates today, where the line between human and AI accountability currently sits, and why he believes the cost of clinical research is about to fall dramatically.
How has AI changed the actual mechanics of study development so far?
Paul Grady: There’s a real manual labor portion to study development, and that’s the first thing. Taking a protocol, designing a clinical research study from that protocol, and then building the actual study from that design is a very long process.
AI can help build a protocol. We don’t suggest that you just have AI build a protocol and go with it. You need to have your clinical experts working with AI to do that. But it can really speed up and improve the process. One of the biggest chores of building and designing clinical studies is taking a protocol and designing case report forms around it so that you’re collecting the data necessary to achieve the defined endpoints. On a big oncology study, that could be hundreds of case report forms, and that process could take months. When humans do it, it’s subject to error. Floyd does that in minutes. The longest I’ve seen Floyd take to design all the case report forms from any protocol is 10 minutes.
Where Floyd is different from a general-purpose AI is that once those case report forms and study schedule and eDiaries and unscheduled visits are all designed, Floyd understands TrialKit’s architecture and backend data model and will go and build the study within the eClinical platform. By doing that, you’re saving the study builder weeks, maybe months, depending on the size of the study. What Floyd does appears, at least in our experience,, to be better than any of the studies we’ve ever built the traditional way.
How does that process actually look on a real study build today?
Paul Grady: We don’t do any of the manual study builds anymore. It’s too expensive, too prone to error, and too time-consuming.
What happens now is you ingest a protocol, or you have Floyd build a protocol from a synopsis. Floyd then designs the study. All of this is done through what we call a prompt UI, a user interface where you’re entering English language instructions and Floyd is following them. One of those instructions is to go and build the study. It does it all in the background, and that takes anywhere from 2 to 10 minutes to build the entire study within TrialKit.
Once that’s done, the human still needs to review all of the case report forms. Floyd provides tools for that, including synthetic data. You can run synthetic data through, and Floyd will tell you exactly what data it generated and what’s in the database. It will tell you which edit checks should have fired and which ones did fire, and it creates what we call a study validation report. All of that happens within the TrialKit user interface.
I also want to say something about where prompt-based interfaces are going, because this is not just a TrialKit story. Right now, if you want to build a complex randomization scheme, you have to learn all the navigational steps: which menu to pull out, which option to use, how to complete the form, how to define which roles are blinded, how to upload the randomization schedule, how to verify it uploaded properly, how to decode it. That’s at least 10 to 15 steps. With Floyd, you just say, “I want to build a randomization scheme,” and Floyd asks you the relevant questions in plain English. Every software package in the world will work this way within five years.
“Study builds [in TrialKit] went from anywhere from two to six weeks to one to three to four days. The work Floyd is doing on that study build takes less than 10 minutes.”
How does the human still own the final decision, and what would have to change before that shifts?
Paul Grady: Contrary to popular belief, AI in clinical research is not going to eliminate jobs. All of the roles are still necessary within study development and study implementation. What Floyd does is handle a lot of the manual labor, and it can do that for just about any role. If you’re part of the research process, Floyd is going to make you much more productive.
Once the study is built, the study builder still needs to go in and test those case report forms. Floyd can set up a whole lot of statistical analysis, but it still needs to go to the biostatistician for final review. A lot of the labor of biostatistics is setting up all the models, and Floyd can do that. But Floyd doesn’t make that final decision. That’s still done by a human being. And that’s the way it is across the clinical research process.
Why did you build Floyd to operate in a closed session, and what does that mean for a sponsor concerned about data security?
Paul Grady: Floyd comes to the table with a tremendous amount of knowledge. Floyd itself has been trained in the clinical research process, and it has also been trained in TrialKit. TrialKit is entirely API-driven, so Floyd knows how to make every single API call. Floyd is literally doing the same thing any human being would do, just a whole lot faster.
TrialKit has what we call closed TrialKit AI sessions. Once a session has been initiated, typically by ingesting or creating a protocol, everything you do within that session Floyd learns from. So when Floyd creates a protocol and then a study, and we ask it to create 10 participants of synthetic data, Floyd takes that protocol and study within that session and understands exactly how to do all of that.
If you go to any other study in the system, and right now we’ve got about 12,000 studies in our system, and ask Floyd about the participants or the protocol in a different study, Floyd will say it has no idea what you’re talking about. This is an entirely closed environment. Floyd never learns from one study and applies that to any other study. The reason we do that is security. There has always been a wall between studies in systems like TrialKit, where access to one study does not mean access to any others. That same boundary applies to Floyd.
We may allow users to share sessions across studies at some point, but we haven’t done that yet. There is real concern about Floyd learning how to do things based on our technology and then applying that to somebody else’s work.
Where does synthetic data fit into the broader trial process, and how far do you expect it to go?
Paul Grady: The real cost of clinical research is not study development. That’s a small percentage of the total. The real cost is recruiting participants, keeping them involved, having your sites and investigators collect data across a variety of different technologies, aggregating all of that, and making the proper decisions for regulatory submission. The most expensive part is participant recruitment and retention. That’s where synthetic data will make the biggest difference. It’s not doing it today, but it will. There’s no question in my mind, at least.
Right now, the synthetic data Floyd creates is not ever submitted to the FDA or any other regulatory body. It’s really used to validate the study: did our protocol endpoints line up with the actual data, and is this study going to answer the questions we’re asking? That’s all part of a human process making a human decision.
I’ll be honest: the synthetic data TrialKit creates on its own is not the best synthetic data in the world. We haven’t applied our resources toward modeling it for different indications. However, there are companies that have, and they allow their synthetic data models to be connected to by TrialKit. One such company specializes in oncology. When that company’s AI is creating synthetic data, Floyd can access it and populate TrialKit with it. That synthetic data comes out of a very complex statistical modeling scheme that does its best to mimic what it believes the live data will represent. It’s almost scary how accurate it can be today.
I did a demonstration yesterday where I created 10 participants. There were 194 case report forms per participant, and within an hour Floyd had created nearly 2,000 completed forms. That would have taken a human being a month.
The bigger story, though, is what’s coming. If you look at an oncology study today, particularly for a catastrophic indication, the study is almost always randomized. In a two-arm study with 1,000 participants, 50% of them will not be getting the treatment. Oftentimes those participants do not survive the protocol. A number of companies are already petitioning the FDA to allow synthetic data to serve as the control arm, because it’s fairly inhumane to tell somebody they’re part of a clinical trial while knowing they’re not going to receive any treatment and probably won’t survive the protocol. I would say within a few years we’ll have synthetic data being the control arm of some of these studies. And that is really where the future lies.
Once we’re able to use synthetic data to accurately predict disease progression and treatment efficacy, the cost of clinical research will be cut dramatically. Because of that, we’ll be able to do so much more research, our treatments will get better so much faster, and you’ll see an enormous acceleration in healthcare benefits.
“It’s fairly inhumane to tell somebody, ‘Yes, you’re part of a clinical trial,’ but knowing that they’re not going to get any treatment and they’re probably not going to survive the protocol.”
Paul Grady is a co-founder of Crucial Data Solutions, the company behind the TrialKit eClinical platform and Floyd, its clinical AI agent.
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
