Phase-3 clinical trials are generating unprecedented data volumes, a trend that fuels the vision of an autonomous clinical trial while keeping human supervision at the core.
Data explosion and early AI adoption
A 2020 study showed an average phase-3 protocol collected about 3.56 million data points. By 2025 that rose to roughly 5.96 million, a 67% increase and 6.4 times the 2012 average of 929,203, according to research by TransCelerate BioPharma and the Tufts Center for the Study of Drug Development. The analysis flagged that nearly one-third of procedures were non-core, suggesting room for tighter discipline.
“I don’t think sponsors are as worried about collecting too much data,” said Venu Mallarapu, chief transformation and AI officer at eClinical Solutions. He noted that AI and machine-learning models now make it easier to ingest and analyze larger datasets, while wearables and sensors promise even higher volumes.
AI’s impact is most visible in the data-processing layer. eClinical Solutions reports that an average client runs three to five defined AI use cases, mainly for drafting or triaging reports. Most agents remain assistive; a human still validates the final action. The most advanced deployments embed scoped agents within reengineered workflows to maximize efficiency.
Human oversight remains essential
Clinical-research staff face mounting burnout as trial complexity climbs. “Sites are understaffed and overwhelmed,” said Janice Chang, CEO of TransCelerate BioPharma, citing administrative burdens such as training, paperwork and contracts.
Agents are easing that load by handling routine data-management tasks. Pamela Tenaerts, chief medical officer at Medable, explained that a monitoring agent can surface discrepancies—like a new medication entry without a matching adverse-event report—allowing a CRA to focus on the issue rather than hunting across multiple systems.
While agents accelerate processing, they do not replace judgment. Ken Getz, executive director of the Tufts Center and lead author of the data-collection study, warned that capacity-driven data collection can become a self-reinforcing cycle.
Regulatory guidance already calls for uncluttered study designs. ICH E8(R1) advises that critical-to-quality factors stay clear of extensive secondary objectives. Yet sponsors often retain data “just in case,” creating a tension between extensive collection and efficient analysis.
In practice, agents act as “propose-and-dispose” tools. Mallarapu emphasized that no agent takes action without human approval, and every step, trigger, input, version, output, decision timestamp, is logged for auditability. This traceability satisfies the stringent transparency required in clinical trials.
eClinical’s Data Advisor illustrates the “swarm” model: a single interface talks to the user while multiple agents operate behind the scenes, handling tasks such as SDTM mapping. As automation improves one bottleneck, another emerges, often at the site level, where staff must absorb faster query cycles.
Industry adoption and measured impact
A May 2026 METR survey of three hundred forty-nine technical workers reported median self-reported productivity gains roughly doubling each year. Although self-assessment may overstate actual efficiency, the trend shows a broader confidence in AI-augmented workflows across regulated environments.
Data architecture and workflow integration
Underlying the surge in automated processing is a shift toward unified data platforms. Vendors note that spreadsheets, while familiar, often generate multiple copies of the same dataset, complicating version control and auditability. To address this, organizations are increasingly adopting cloud-based lakehouse solutions such as Snowflake and Databricks, which provide a single source of truth for disparate clinical data streams.
These platforms enable agents to query raw records directly, then pipe curated slices into familiar tools like Excel for reviewer inspection.
