Why Clinical Trials Need an Execution Architecture Before AI Can Deliver: Q&A with Abraham Gutman, AG Mednet

The clinical trial industry has spent more than two decades perfecting the collection and storage of data, but the infrastructure needed to orchestrate what happens next—the decision workflows, the handoffs, the moment-to-moment execution of a protocol across teams and organizations—has never been fully built.

Following the recent SCOPE X conference, Applied Clinical Trials caught up with Abraham Gutman, founder and CEO of AG Mednet, about the operational architecture gap that AI cannot fill on its own, why the analogy of an orchestra without a conductor captures something the industry has been slow to name, and where enthusiasm for agentic artificial intelligence (AI) is outrunning what clinical trials can actually support.

ACT: What does operational architecture for AI actually mean in clinical trials, and why has it taken this long for the field to start defining it?

Gutman: In clinical trials we've had systems for many, many years. Roughly 1997 we saw EDC begin to exist—when Phase Forward, Paul Bleicher, and others began to put forms in computers. What that did is digitize the collection and storage of data in a way that eliminated distance and compressed time. From that EDC renaissance we saw other types of data capture systems—eCOA, ePRO, all the normal suspects. Over the last 20-some years we have very much perfected the art of capturing data, and that's super critical. That is the enabling foundation for today's clinical trials. Without it, it's inconceivable that the level of complexity could have been managed without that type of computational capability.

That being said, data that is captured doesn't do anything. At the conference I used a slide showing an orchestra on stage—but not the players. It was the sheet music, the heavy instruments. And the question was: is that a symphony? Well, no, it is not a symphony. Even though we have captured the data, we have all the sheet music there for every instrument. What constitutes a clinical trial is what you do with the data—how you enable decision making. And decision making requires not only the players in the different sections of that orchestra, but also a conductor. Even though the sheet music will tell every player exactly what they need to do—think of the collection of sheet music as the protocol—the conductor is still paying attention to how things are going. Somebody may be playing a little too soft, somebody may be going a little too fast or too slow, and the conductor can enable the music to actually happen according to how the protocol goes.

In clinical trials, decisions occur because different parts of the organization do certain things and pass the piece to the next one, and then they do something, and then they pass it on, and sometimes something needs to be pulled back. That is exactly where people start using trackers—Excel, spreadsheets, SharePoint, email. Those people are essentially operating the orchestra without a conductor. You don't know at all times the who, what, when, where, how, and why of everything going on. In order to do that, you need something that orchestrates, that manages and ensures that during the handoffs nothing falls through the floor—that even between teams, across teams, across organizations, everything plays according to how the protocol called for.

That's the part that comes afterwards. After you have captured and collected the data, you still need execution. And it is that execution architecture that has not come into our industry.

You asked why it has taken so long. One part of the answer is everything in our industry takes very long. Not six or seven years ago, when we moved our operation from a data center onto AWS, I told the auditors at the time and the auditor said absolutely not—you need to be in control of your machines. Today that sounds completely ridiculous. We're very conservative, and in many respects for very good reasons. Regulations keep participants safe and make science be science. But regulations should not be an excuse for eliminating progress.

What I'm talking about in terms of this architecture is process management. It's different from data capture. Data capture produces workflows that allow you to ensure data is captured and completed correctly. The workflows I'm talking about are decision workflows—the things that define how decisions are going to be made in the clinical trial. Those live side by side with the capture systems that hold the data.

ACT: How does the right infrastructure change the role of human experts—and what does it look like when they're owning decisions rather than managing process?

Gutman: We all hear how trials have become incredibly expensive, and that cost doesn't seem to be coming down anytime soon. Why? Because we're highly regulated, and we have thrown people at the problem of ensuring that every bit of SOP for every particular trial is met. Meet a group of project managers and trial coordinators, and it's the definition of hair on fire—everybody is very, very scared. One bit of PHI escapes and the billion-dollar project for the greatest blockbuster drug is now in danger. So how do you handle that? You throw more brains at the problem.

But many of those people are solving overhead problems that up until now were very hard to computerize or automate. When we look at the areas of AI that are more objective—areas where there is significantly less opportunity for hallucination—PHI redaction is a good example. There's a level of objectivity that allows you to look at whether you did the right thing or you didn't. Was a name there? A social security number? A patient ID? A telephone number? These go beyond simple search and replace.

There are a lot of elements like this where a site needs to provide a set of documents, somebody needs to see whether those documents contain the information necessary for downstream decision makers, so you have people reading those documents, and then people QA-ing the first person to make sure they didn't miss anything. That type of work is very different from the people looking at the statistics of the clinical trial and thinking about whether the trial should continue, whether some arm should be stopped or accelerated because the drug is working or not working. Those are decision-level things.

So I separate those. There are certain elements that AI enables—often rote activities, but not rote in a simple sense. These things require a degree of reasoning, but not decision making. When you are able to use AI for those activities, you eliminate as many of them as possible, reduce the number of people you need, and the people that remain have a lot more clarity and ability to make decisions on data they know is clean and contains what they need.

But how do you deliver this capability? You need a road. Imagine the United States without a road system—the interstates going north and south, east and west, around cities. That defines how traffic goes, how commerce takes place. Now imagine it without that. How do you know where to transship something from this 18-wheeler to that one? How do you do it, where do you do it?

The road system is what enables the exits and entrances, and it's what enables you to attach more services or less services—which is another way of saying attach more AI or less AI at different points. That is what manages the process.

ACT: What are the key takeaways you want attendees to walk away with from your recent session at SCOPE X?

Gutman: The understanding that you need an architecture in order to deliver the developments and the innovation that AI is bringing. People are developing AI capabilities, many of which are not only interesting but super useful. But these things living tetherless in space, without understanding where they should come in—that's the problem.

The example I gave in the talk: you have an orchestra, and they're going to play a concerto with a soloist—a violinist playing along with the orchestra. Think of that violinist as an AI. The violinist has their back to the orchestra and is facing the public. They have absolutely no idea what the orchestra is doing because they're not looking at them. They're listening, but they don't know exactly when they need to come in and play. What I want people to understand is that you need a conductor.

Developing AI and leaving it tetherless in space is just not going to bring all the benefit it was developed to provide. But if you have a conductor—if you have exits and entrances into this highway where you know you can connect these services—that piece of infrastructure allows you to bring AI's power and expertise at the right point to the right people, and then allow the road to continue on for the next person to do what they need to do.

There's also something I saw at the conference that was a head scratcher for me. The flavor of the day is agents—agentic AI, which by the way is super important and something that will work in the types of AI we're talking about. But there were some folks talking about armies of agents that can run entire clinical trials—essentially one person managing these agents from time to time while the agents do absolutely everything. That struck me as disingenuous. A clinical trial is a way of debugging an illness—debugging a body that has some error that needs to be corrected. We know you can use agentic AI to do code development, but this is not code development. These are humans. Things vary from time to time, and minor adjustments that require human judgment are very critical.

Architectures matter. A way to manage processes is the entry point for AI to truly deliver on its promise in life sciences.