Agentic AI and the human element: Building trustworthy automation in clinical trials
Clinical trials are full of moving parts and, despite recent technological advances, they still come with manual processes prone to error and inconsistency.
A recent study found that up to 40% of trials might be untrustworthy due to flaws in design, improper data collection, and other issues. Creating strong, trustworthy trials isn’t only about building the right forms and protocols. It’s also about connecting the dots from those protocols to real-world operations. The flooding of AI resources into the market presents both an opportunity and a challenge. As trials get more complex, personalised, and decentralised, AI can assist in identifying operational issues earlier in the process, before they escalate into risks that may delay or compromise studies. But without proper guardrails, like audit logs and coded constraints, agentic AI – which can operate autonomously – can add significant risk.
The truth is that biopharma is still catching up on AI and, while even small improvements can save weeks of work and help prevent trial failures, those gains only happen when the AI components involved in trial design are developed and validated with the same rigour as the underlying scientific methods.
Automating trial setup is about more than speed
The traditional drug development pipeline in the United States typically spans a decade or more, relying on weeks and years of human power to copy, paste, and double-check information and outputs, hoping for no errors. AI-enabled approaches are beginning to demonstrate meaningful efficiencies across multiple stages of drug discovery and clinical development, with the scale of time savings varying widely depending on disease area, study design, and application.
Automating trial setup isn’t only about saving time; it is also about increasing participant engagement and ensuring that nothing critical slips through the cracks. Early studies suggest AI-supported approaches may improve enrolment by 10–20%, though results vary by study design and patient population. Importantly, it has also shown potential to improve trial outcomes by supporting better design, data collection, and analysis.
While today’s large language models (LLMs) can already handle repetitive, time-consuming tasks and provide intelligent responses to queries, agentic AI takes things a step further. In some implementations, agentic systems can assist with stepwise planning, system interaction, and configuration tasks. For example, an agent may autonomously navigate a trial management system, identify relevant configuration and data points, implement necessary objects and processes, and validate output against the original requirements – all while you’re working on other critical tasks.
Agentic AI is not limited to early-stage support; it can also flag mismatches in protocol design, operational configuration, or data expectations before they impact study execution. For instance, it can detect when a trial’s protocol requires daily patient-reported outcome (ePRO) submissions, but the system is configured for weekly reminders and flags pre-screening questionnaires that eliminate most candidates before they reach formal screening. These kinds of mismatches are prominent in trials where clinical, data, and operations teams implement their sections independently, only to discover incompatibilities during site training or patient enrolment.
These systems offer a top-down operational view, enabling early detection of cross-functional inconsistencies. To get there, everyone involved must trust the system; and that trust is built on transparent audit trails, explainable decisions, and consistent validation against human review. Agentic AI can support the transition from siloed operations to integrated workflows by automating validation steps and reducing manual setup.
Why the best systems keep humans in the loop
Although agentic AI is built to act and make decisions adaptively without needing constant supervision, the truth is that the best agentic AI systems don’t just do things automatically. Instead, agents request human validation at critical decision points and maintain comprehensive audit logs to ensure full traceability. In trials, strong governance was applied to human designers and researchers long before AI came into the picture and has never been optional; AI makes it even more important.
With AI involved, every study needs strict permissions, audit logs, and a clear separation between what the agent can touch and what a human should change or review. That review isn’t just for the data and actions of the agent, but to evaluate automation boundaries, ensuring human expertise is augmented, not replaced. At all times, humans must have final sign-off during trials to ensure the integrity of the study and the safety of sensitive data. Real progress in speeding up trials starts with trust earned through transparency, consistent performance, and respect for clinical judgement.
Looking ahead: Real innovation is about balance
When it comes to making the best use of AI in trials, incremental improvements matter more than revolutionary claims. The most valuable changes are happening in trials where engineers and clinical teams are working together to build systems that are safe, traceable, and genuinely useful. Automation that goes too far or has too much autonomy can derail trials by creating risks instead of reducing them.
As AI takes on more operational roles in research, collaboration between engineers, clinicians, and regulatory experts will determine how rapidly and responsibly these systems evolve. With proper guardrails in place, agentic AI can evolve from a concept to a foundational tool in trial operations. Done well, agentic AI has the potential to improve trial efficiency, transparency, and coordination across stakeholders, from the researchers conducting trials to the patients who benefit from the results.
The future of clinical trials isn’t about replacing people. It’s about empowering them with smarter tools and ensuring human oversight at every critical decision point.
About the author
Ece Kalvaci is a software engineer at Lindus Health, with dual MSc degrees in Computer Science and Data Science. She develops ML pipelines for clinical trial data extraction and predictive modelling while also leading work on ML-powered tools for protocol design.
