Contributed Commentary by Dr. Dr. Gleb Tsipursky, CEO, Disaster Avoidance Experts 

August 28, 2026 | Clinical research is getting faster. The FDA’s Operation TrialBlazer is explicitly aimed at accelerating and modernizing clinical development, while the agency has also been building toward real-time clinical trials. Earlier this year, Clinical Research News covered FDA proof-of-concept work in which trial signals were transmitted to the agency in real time, and the FDA said it intended to complete selections for a broader pilot in August. 

Faster signal flow is a meaningful advance. It can reduce the delay between what happens in a trial and what regulators or sponsors can see. But it also changes the human work around the data. When information arrives continuously, teams have less time to reconstruct why a signal was accepted, discounted, escalated, or acted on. 

That is why I think real-time trials need a human exception log. 

I do not mean a second database full of routine entries. The log should capture only consequential moments when an automated rule, analytical system, or AI-assisted process produces something that requires human interpretation, correction, escalation, or override. 

A useful entry can stay short. Record the signal or recommendation, the context that mattered, the named human reviewer, the action taken, and the reason. If a later event shows that the initial judgment was wrong, add what changed. 

This sounds administrative until something unusual happens. Then it becomes institutional memory. 

Imagine an automated system flags a safety signal that appears urgent, but the investigator recognizes a data-quality problem tied to one site. The decision to delay escalation may be correct. Months later, however, someone reviewing the record should be able to see why the system’s output did not control the decision. 

Now reverse the case. Suppose an automated process treats a pattern as noise, while a clinician notices that the patients involved share a feature the model did not weigh properly. If the clinician escalates the issue, that exception should be preserved, too. Otherwise, the organization may fix the immediate problem and lose the lesson that could improve the next protocol, rule, or model. 

The FDA’s Guiding Principles of Good AI Practice in Drug Development point in the same direction. They emphasize human-centered design, a clear context of use, data governance and documentation, risk-based performance assessment, and life-cycle management. An exception log turns those ideas into a small operational habit. 

It also helps separate oversight from hindsight. 

In many organizations, “human in the loop” becomes a label rather than a workflow. A person technically has authority to intervene, but nobody defines what intervention should look like, what should be recorded, or how recurring exceptions feed back into system design. Real-time clinical trials make that weakness more consequential because decisions and signals can accumulate quickly. 

A good exception log creates three feedback loops. The first is immediate. It gives the current team a compact record of why a high-stakes automated output did or did not drive action. The second is managerial. Trial leaders can review recurring exceptions and identify where training, thresholds, data quality, or escalation rules need adjustment. The third is developmental. Sponsors and technology teams can use aggregated exception patterns to improve the system itself without treating every human override as a failure of automation. 

The log should also have boundaries. Teams should not record every ordinary confirmation, because that would bury the important events in administrative noise. They should define a small set of triggers in advance: a human override of an automated recommendation, a signal escalated despite a low system score, a high-severity alert that a reviewer closes, a material data-quality concern, or a decision where professional context changes the automated interpretation. 

Those triggers can differ by protocol and technology. The point is to make the threshold explicit before a difficult case appears. That gives reviewers a shared expectation and makes later analysis more useful. 

Governance should stay lightweight. A trial leader can review exception patterns at an existing quality or safety meeting rather than creating a parallel committee. The review should ask whether the same exception is recurring, whether a threshold needs adjustment, whether a site needs support, and whether a lesson belongs in the next version of the workflow. That turns documentation into improvement rather than paperwork. 

That last point matters. The goal should not be to minimize human intervention at all costs. A falling override rate can mean the system is improving, but it can also mean people have stopped challenging it. The better question is whether the combination of automation and professional judgment produces more reliable decisions. 

This principle applies beyond AI. FDA’s current work on digital health technologies reflects a broader move toward continuous and technology-mediated evidence. As more trial data arrive through sensors, platforms, and automated analysis, teams will need disciplined ways to preserve context around the exceptions that matter. 

Clinical research has spent decades building systems for traceability, accountability, and protocol discipline. Real-time trials should extend that tradition rather than bypass it. 

The strongest version of a real-time clinical trial will not be the one with the fewest human interventions. It will be the one where investigators, sponsors, and regulators can understand which automated signals mattered, where expert judgment changed the path, and what the organization learned from those moments. 

Speed is valuable. Traceable judgment is what keeps speed from becoming opacity. 

Dr. Gleb Tsipursky helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of the future-of-work consultancy Disaster Avoidance Experts. Gleb earned his doctorate in the history of behavioral science at the University of North Carolina at Chapel Hill in 2011, his master's degree at Harvard University in 2004, and his bachelor's degree at New York University in 2002. He can be reached at Gleb@disasteravoidanceexperts.com.