AI Agents in Clinical Trials Require Human Review Before Data Commits

A phase 3 protocol in 2020 collected an average of 3.56 million data points. By 2025 that number had climbed to 5.96 million, a 67% rise in five years and more than six times the 2012 baseline of 929,203, according to joint research from TransCelerate BioPharma and the Tufts Center for the Study of Drug Development. That trajectory is exactly what is pushing AI agents from demo to deployment across clinical operations. The problem is that the volume of data now arriving in trials outpaces what human monitors can process, which is what makes autonomous agents attractive. The problem is also that those same agents make errors confidently, which is what makes unsupervised autonomy dangerous.

The same TransCelerate/Tufts work found that roughly one-third of all procedures and data points collected in trials are non-core or non-essential. Agents trained on bloated protocols inherit that bloat. An agent processing 6 million data points across endpoints, wearables, and decentralized assessments can surface patterns faster than any CRA team, but a misconfigured rule or a hallucinated query string propagates across every site simultaneously. The failure mode for autonomous systems is not slow and local; it is fast and global. That asymmetry is what the industry has not fully priced in yet.