AI Agents in Clinical Trials Need Human Oversight Before Data Commits

Data volume drives AI adoption—and risk

Phase 3 clinical trials now collect an average of 5.96 million data points per protocol, up 67% from 3.56 million in 2020 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 exponential growth is pushing AI agents from pilot projects into production across clinical operations.

The appeal is straightforward: autonomous agents can surface patterns across endpoints, wearables, and decentralized assessments faster than any clinical research associate team can manually review. The risk is equally clear. When an agent makes an error—whether from a misconfigured rule or a hallucinated query string—that mistake propagates across every site simultaneously. The failure mode for autonomous systems is not slow and local; it is fast and global.

Why it matters

Sponsors face a fundamental tension between operational necessity and regulatory accountability. AI agents can handle data volumes that overwhelm human monitors, but the FDA's framework for real-world data in regulatory submissions requires traceability and human accountability at the point of decision. How the industry resolves this tension will determine whether AI accelerates or complicates trial execution at scale.

Mandatory review becomes the operational standard

The emerging model treats agent supervision as a defined role rather than an afterthought. Platforms like Castor EDC now build mandatory human review into AI-extracted data workflows before any record commits to the clinical database. This architecture reflects where the practical ceiling sits today: agents handle extraction, flagging, and pattern recognition, while a human signs off before the record becomes final.

That design choice also addresses a regulatory reality. Fully autonomous commit functions represent a liability regardless of technical capability, because current frameworks expect human accountability at decision points.

Protocol bloat compounds the challenge

The same TransCelerate/Tufts research 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. When an AI system processes 6 million data points, it cannot distinguish between essential endpoints and unnecessary collection—it treats all inputs as equally significant.

What changes operationally is not whether humans remain in the loop but what they do there. Reviewing agent outputs at scale demands different skills than traditional monitoring: less site travel, more query triage, more judgment about whether a flagged anomaly reflects a genuine signal or a training artifact. Sponsors building these workflows now are effectively defining what a clinical data reviewer does in a trial generating millions of data points.

The metric worth tracking is how review capacity scales relative to protocol complexity. Data volume has not flattened, which means the gap between what agents can process and what humans can validate will continue to widen unless operational models adapt.

These details were first reported by eClinical Solutions.