Most clinical operations teams are not debating whether to use AI. They are stuck on a more uncomfortable problem: how to introduce it into studies that are already running, with data that cannot be rewound and regulators who will ask exactly what the system was permitted to do. A new analysis from Sitero puts a practical frame on that problem, and the most useful part is not the technology advice but the sequencing advice.
The adoption barrier, the analysis argues, is not technical skepticism. Organizations are actively evaluating AI but have largely confined it to non-critical workflows or limited pilots. What stops broader deployment is validation uncertainty and the absence of a governance structure, not doubts about whether the tools work. That distinction matters because it points to a fixable problem. Over 1,600 AI-enabled medical devices have already cleared FDA review, most relying on classical machine-learning approaches with established qualification paths. The gap between that track record and live GCP deployment is not capability. It is process infrastructure.
The most concrete guidance in the analysis concerns timing, and it is direct enough to act on. Introducing an anomaly detection tool into a study that has been running for months means the system starts on historical data it was never trained to expect, generates flags that reviewers must adjudicate on top of their normal load, and creates an irregular baseline in the trial record. The practical lesson from organizations that have already gone through this: pilot on new study starts. The data is clean, the team has room to absorb a learning curve, and nothing is added to an ongoing regulatory record mid-stream. The analysis also notes that timing cuts both ways. Deploy too early in enrollment and the model produces noisy output because there is not enough data. Deploy too late and months of records need reprocessing before the tool is useful. Neither failure is a technology problem.
On validation, the analysis draws a distinction that clinical operations teams will need to absorb before any inspector visit. Traditional computer system validation assumes deterministic logic: input A reliably produces output B. Large language models and probabilistic systems do not work that way, and model behavior can shift if a vendor updates parameters without surfacing that change to the customer. The analysis identifies three layers that validation documentation must cover: the orchestration layer defining what the AI is authorized to do and in what sequence; the human use layer, which it flags as frequently underweighted because well-formatted AI output masks data errors that reviewers then miss; and model performance, which matters only as one component of the full system. The governing concept for all three is context of use, documented with precision before deployment, not after the capability has quietly expanded from anomaly detection into query generation and site communication. ICH E6(R3), whose Annex 1 was adopted in January 2025, calls explicitly for proportionate oversight, and the analysis applies that principle directly: any output feeding into a regulatory submission requires documented human review at every step. Whether a sponsor can actually show that documentation is what an inspector will determine.
Source link: https://sitero.com/ai-in-live-clinical-studies-exploring-the-reality-of-implementation/
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.
