Why Clinical Trial Failure Persists
Drug development is one of the most data-intensive endeavours in modern science, yet the industry continues to lose most of its candidates at precisely the Phase II and III stages, where investment is greatest, and the reasons are rarely surprising. Fragmented data, recruitment that stalls, and protocol assumptions that do not translate to how patients present in care settings have driven late-stage failure across the industry for decades. The response has largely been to treat this disruption as the cost of doing business, something to absorb and plan around rather than solve.
The data and analytics investments of the past decade have given research and development teams better visibility into where trials strain, but visibility has not changed the failure rate at the moments that matter most. Knowing where a trial is heading and being able to change course before it gets there are not the same thing. That gap has never really been closed. Agentic AI is built differently, less focused on flagging what is wrong and more on doing something about it before the damage sets in, offering a viable path forward to improve the efficiency of clinical trials and research.
The Gap Between How Trials Are Planned and How They Run
Most trial operating models are built on assumptions locked in at the outset. What teams hoped would be true about patient populations, site capabilities, and enrollment projections does not always match what real-world data would have told them. Sponsors design trials around expected treatment pathways, and when those pathways diverge from reality, amendments follow. Eligibility criteria that looked reasonable on paper quietly shrink the pool of eligible patients before a single one is screened.
Phase II and III are where those vulnerabilities surface most painfully. The investment is deepest, the timelines longest, and the consequences of a misstep are the most costly at these points. Yet these are also the stages most dependent on assumptions made long before the trial begins. According to researchers at Epistemic.AI, termination rates at these stages roughly doubled over the past decade, with strategic and business decisions now outpacing efficacy as the leading driver.
Recruitment tells a similar story. One study by researchers at the European Molecular Biology Laboratory's European Bioinformatics Institute analysing more than 28,000 stopped trials found insufficient enrollment was the single most common reason trials were halted, accounting for more than a third of all stoppages. In oncology specifically, the window to identify and engage an eligible patient can be a matter of days or even hours before a change in therapy eliminates their eligibility entirely.
Much of the information needed to identify and match patients never makes it into a structured format. Physician notes, unstructured data within electronic health records, and paper documentation all take immense effort to parse before anyone can act on them. That effort takes time that most trial and site research teams do not have. By the time the right data has been compiled and translated into something actionable, the operational window it was meant to inform has often already passed. By the time signals that precede a breakdown show up in formal reporting, the window for early intervention is usually already gone. Adding more people and more oversight layers has been the default fix, but it treats the symptoms. The underlying problem is that trials are dynamic systems being run on static plans, and that mismatch is where failure compounds.
How Agentic AI Changes the Equation
The distinction between analytical AI and agentic AI comes down to what happens after the insight. Analytical AI tells a team what is happening. Agentic AI helps coordinate what happens next, and that gap is where most of the value in clinical development has historically been left on the table. R&D teams have become better informed over the years, but humans have remained the bridge between insight and action. Without actionable insights and guided action, that bridge introduces exactly the kind of latency that turns manageable problems into costly ones. Agentic AI is designed to close that gap. It autonomously executes on next-best actions across complex, multi-stakeholder workflows, pulls from live data sources, continuously evaluates its own outputs, and escalates critical decision making, together with supporting recommendations and data, to human experts and their oversight at the key decision points that require it. Where it becomes particularly powerful is in how individual agents work together, each handling a discrete part of a complex problem and passing outputs to the next, orchestrating a response to questions that would otherwise take a team of analysts with varied backgrounds days or weeks to integrate, translate, and bring to a recommendation. In protocol design, this means treating feasibility as a continuous loop rather than a one-time gate. Eligibility assumptions are stress-tested against real-world patient data and site realities as new information becomes available, surfacing amendment risks before costs have compounded rather than after. Protocol changes that once required weeks of manual analysis to justify can be identified, modelled, and escalated for human review in a fraction of the time, keeping studies on track rather than forcing teams into reactive damage control. Site selection follows the same logic. Moving beyond historical performance data to integrate real-time patient and enrollment signals establishes a stronger site strategy from the outset. The feasibility, selection, and activation workflows that routinely create bottlenecks across legal, regulatory, and site liaison teams when managed manually can be automated, compressing timelines that have historically absorbed weeks to months of unnecessary delay. Sites that would have been flagged as underperforming only after enrollment targets slip can be identified earlier, giving teams the runway to intervene before the disruption to study timelines becomes unrecoverable. In operational trial monitoring, this shift ensures that signals of enrollment drift, protocol deviations, and data quality issues can be detected when they first appear in operational data rather than when they surface in formal reporting and review processes. Risk signals can be sorted by urgency, predicted likelihood and impact, and then routed to the right person in real time, empowering earlier intervention and action before risks have the chance to compound into negative timeline and budget impacts. Catching a site performance issue at week two rather than week eight can represent the difference between a minor course correction and a costly delay that ripples across the entire programme.
Deploying Agentic AI at the Stages Where It Matters Most
The highest-risk moments in clinical development are also the moments where the consequences of a poorly governed AI decision are greatest. Responsible deployment at Phase II and III is a different challenge than deploying AI in lower-stakes workflows. According to researchers at the Tufts Center for the Study of Drug Development, a single day of delay in a Phase III trial carries a direct cost of $55,716, with an additional approximately $500,000 per day in unrealised prescription sales. At that scale, the governance decisions made before deployment are just as consequential as the technology itself.
The goal is to give the experts making decisions new, data-driven insights that were previously out of reach by surfacing analyses they may not have known to ask for, integrating disparate data sources into a single coherent recommendation, and compressing the time between a question emerging and a decision-ready answer landing in the right hands. That does not mean removing humans from the equation. Human judgement still drives high stakes and critical decisions. Eligibility determinations, safety reviews, protocol change approvals — those stay with people regardless of what information the system itself has and recommendations it may provide. The value of agentic AI at these stages is in the quality and speed of analysis and insight that it can hand to the decision maker to enable better, faster decisions to be made.
Every recommendation, routing decision, and automated action needs a traceable record of the data and logic behind it. Not just for internal quality assurance, but because the regulatory submissions that follow depend on it and teams need to understand how recommendations and decisions were made when it counts. In good clinical practice-regulated environments that standard is not optional. An agentic system that cannot explain its reasoning is not just a governance risk, it is a liability that has the potential to surface at the worst possible moment.
Data quality underpins all of it. Poor data produces faster and more confident errors for the teams relying on agentic AI to guide their most critical decisions. Organisations that invest in establishing robust data quality requirements and high-quality governance around their data infrastructure before deploying agentic tools will see compounding returns. Those that treat data quality as an afterthought will find that the technology amplifies their problems rather than solving them. At the stages where getting it wrong is most expensive, fragmented or poorly governed data – and the agents and predictive models that it drives – is a liability that compounds at exactly the wrong moment.
Rethinking Risk in Clinical Development
Late-stage trial failure has been treated as an industry constant for decades, something built into the cost models and timelines of drug development rather than a problem with a solution. That assumption has persisted not because the failure points are unknown but because the tools available have never been capable of acting on them at the speed and scale that clinical development demands.
Agentic AI shifts that equation by addressing failure at its source. Rather than managing the fallout after signals have already compounded into something costly and hard to reverse, it coordinates a response while there is still time to change the outcome. The organisations moving earliest are already beginning to see what a structurally different approach to trial risk looks like in practice.
What becomes possible is a fundamentally different relationship with risk, one where the patterns that have historically predicted failure start working in R&D teams' favour rather than against them. The governance frameworks are catching up, the tools exist, and the pressure to move faster is only growing. As agentic capabilities continue to evolve and mature, the industry must integrate these capabilities into their development approaches, trading incremental, risk-averse change for the revolutionary improvements they make possible. Those that do will outpace their competition at a rate not seen before.
Read the full article — it's free
Register with Pharma Focus America to unlock expert insights, research articles and in-depth industry analysis.
