Quick Buzz Feed
Clinical trial sites should establish robust data infrastructure before deploying AI tools, according to industry experts and recent guidance on sequencing implementation.
Despite approximately 75% of clinical trial sites utilizing eSource in their active studies, a recent panel discussion by CRIO and Clinical Leader revealed that many sites are still grappling with a more fundamental decision: whether to acquire off-the-shelf AI tools or develop custom solutions. Industry operators with practical AI experience caution that this 'buy vs. build' question is premature for the majority of sites, emphasizing the critical need to address foundational issues first.
Experts from Clinical Research Philadelphia, ALSA Research, and Velocity Clinical Research, moderated by CRIO Chief Innovation Officer Mike Wenger, consistently advised that sequencing of technology adoption is paramount. They recommend that sites still reliant on paper records must first transition to electronic systems, then establish a basic AI governance policy, and provide staff with enterprise Large Language Model (LLM) accounts. Skipping these foundational steps and moving directly to custom AI tooling often results in more rework and complications than benefits. For CRIO clients, leveraging existing backend data access through tools like Google BigQuery and Looker allows for direct layering of financial forecasting and custom dashboards onto operational data, eliminating redundant data entry—a necessary precursor to effective custom AI workflows.
Where AI has proven its value, its applications are intentionally narrow and always involve human oversight. A concrete example cited is the drafting and updating of investigator CVs: using an enterprise LLM to generate initial drafts, followed by human review and iterative refinement, has shown to be more efficient than attempting to fully automate the entire process. The panel highlighted that complex automations designed to ingest diverse data (documents, chat logs, system data) simultaneously and produce multiple outputs (emails, reports, dashboards) are difficult to trust and repair when issues arise. They specifically flagged inclusion and exclusion criteria as unsuitable for current AI tools due to the variability between studies and the high risk of inaccurate interpretations from inconsistent input.
The International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) E6(R3) guidelines, while not prescribing explicit AI governance, provide a risk-based framework that accommodates new technologies. This regulatory landscape places the responsibility on clinical trial sites to thoroughly document and justify their chosen tools. Consequently, deploying AI before a robust and clean data infrastructure is in place severely limits a site's ability to demonstrate that its AI-generated outputs are auditable. A crucial benchmark for compliance is whether a site can precisely identify the data inputs and human review steps for every AI-assisted output it produces, ensuring transparency and accountability.