AI agents are no longer just a promise — they’ve become a reality inside clinical trials.
And the reason is pretty straightforward: the volume of data has grown so much that no human team can keep up on its own. To put it in perspective, a Phase 3 protocol collected an average of 929,203 data points back in 2012. By 2020, that number had climbed to 3.56 million, and by 2025, it’s approaching nearly 6 million — a 67% jump in just five years and more than six times the figure from 2012, according to joint research from TransCelerate BioPharma and the Tufts Center for the Study of Drug Development.
That staggering growth opened the door for autonomous tools to take over significant parts of clinical operations. But here’s the point a lot of people still overlook: speed without oversight becomes risk. An agent that makes a mistake doesn’t make it slowly or in just one place. It makes it fast, at scale, and across every research site simultaneously. This asymmetry between slow, local failure and fast, global failure is exactly what the industry hasn’t fully accounted for yet. That’s why human review hasn’t left the picture — it’s just changed shape. 👇
What AI agents are actually doing in clinical trials today
When we talk about AI agents inside clinical trials, we’re not talking about a chatbot answering questions or a system that generates pretty reports in PDF. We’re talking about systems that process millions of data points in real time, cross-reference information from endpoints, wearables, and decentralized assessments, flag inconsistencies in medical records, signal protocol deviations, and recognize patterns far faster than any CRA team ever could. This is real autonomy, applied in an environment where mistakes have direct consequences for patients’ lives.
One important detail from the TransCelerate and Tufts study is that roughly one-third of all procedures and data points collected in clinical trials aren’t essential or central to the study’s objective. This creates a chain reaction problem: agents trained on bloated protocols inherit that bloat. In other words, if the model learns from data full of excess, it reproduces that excess — and worse, it can spread configuration errors or poorly formulated queries across every site at once. A miscalibrated rule or a machine-generated query propagates instantly, and that’s where the real danger lives.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Adoption ramped up significantly once it became clear that traditional monitoring models simply couldn’t keep pace anymore. The old model — built around periodic in-person visits to research sites — was designed for a world where data arrived on paper and was organized manually. With the full digitization of clinical records and the arrival of wearables, remote monitoring devices, and integrated sensors, the volume of information exploded in a way that no in-person review could absorb with the necessary frequency. AI agents filled that gap with an efficiency that impresses anyone who’s worked in clinical research for more than a decade.
What few people realize is that this efficiency comes paired with a new and fairly complex responsibility. When an autonomous system analyzes thousands of patient records per hour and makes decisions based on patterns it taught itself, the traceability of those decisions needs to be flawless. Who’s accountable for an error? How do you audit a process that happened in milliseconds, inside a model layer that isn’t always transparent? These are real questions, and they’re at the center of regulatory discussions happening right now in the United States, Europe, and Brazil.
Why human review is still non-negotiable
Human review isn’t a brake on progress. In practice, it’s what keeps progress from turning into chaos. And this becomes much clearer when you understand how AI agents fail. They don’t fail the way a human fails. A tired researcher misrecords a single entry. An agent with a training bias systematically gets it wrong across every record that shares that specific characteristic — with no variation, no obvious red flag, and often for a long stretch before anyone notices the pattern. In clinical trials, where data integrity is the backbone of any regulatory approval, this kind of silent error is the most dangerous of all.
Making matters worse, models tend to be wrong with confidence. They don’t signal uncertainty when they deliver an incorrect answer, and that apparent certainty can fool even experienced teams. This is precisely the behavior that makes unsupervised autonomy so risky. That’s why platforms like Castor EDC already embed mandatory human review within AI-extracted data workflows, ensuring nothing gets recorded in the clinical database before a qualified person signs off on it. This architecture reflects exactly where the practical ceiling of the technology sits today: the agent handles extraction, flagging, and pattern recognition, while the human gives the final go-ahead before a record becomes permanent.
What’s changing isn’t the need for human review, but the point in the process where it kicks in and the type of skill it demands. Reviewing an agent’s outputs at scale calls for a professional profile that’s different from traditional monitoring: less travel to sites, more alert triage, and a lot more judgment about whether the agent’s flag represents a genuine signal or just a training artifact. Sponsors building these workflows right now are effectively redefining what it means to be a clinical data reviewer in a study generating 6 million data points.
There’s also an ethical dimension that can’t be ignored. Clinical trials exist because society needs solid evidence before making a medication or procedure available to everyone. This chain of trust depends on human beings being actively involved in the most sensitive decisions of the process — not as figureheads who rubber-stamp machine-generated reports, but as intellectual agents who understand what the system did, why it did it, and what the implications are. Human review is, therefore, part of the trust architecture that makes a clinical trial’s results valid in the eyes of regulators, physicians, patients, and society as a whole.
Autonomy with control: the model that’s emerging
The discussion around autonomy for AI agents in clinical settings is heading in a very specific direction: it’s no longer about deciding whether agents should or shouldn’t have autonomy, but about precisely mapping which tasks can be fully delegated, which need active supervision, and which should never leave human control. This mapping — informally called clinical automation levels within the industry — is being built empirically, based on the hits and misses observed over the past few years of real-world implementation.
In practice, the most mature models being adopted today run on an escalation logic. The agent operates with full autonomy on low-risk, highly repetitive tasks like duplicate checks, standardized field completion, and consistency verification between forms. When it spots something outside expected parameters, it routes it to a human review queue with an associated priority level. The human reviewer no longer needs to wade through an ocean of data to find what matters. They receive what matters — with context, with the agent’s reasoning laid out, and with available action options. This reduces cognitive load, increases response speed, and keeps the human exactly where they need to be: in control of the decisions that carry weight.
The metric worth watching here is how this review capacity can scale relative to protocol complexity, because the data volume curve simply hasn’t stopped climbing. As studies get more complex and data points keep multiplying, the pressure on the human review layer will increase right along with it. Those who structure this capacity intelligently now will have an enormous operational advantage in the years ahead.
What this model requires, though, is an extremely well-organized data infrastructure and a documentation culture that many organizations are still building. An agent can only scale correctly if the data it receives is reliable, if the flagging criteria have been properly calibrated, and if the human teams have been trained to interpret alerts critically — rather than just automatically rubber-stamping whatever the system suggests. When any of those elements breaks down, the promise of intelligent autonomy turns into an operational blind spot that can take months to detect. That’s why investment in data governance and human upskilling remains just as urgent as investment in the AI tools themselves. 🤝
What regulators are saying about all of this
The FDA, the EMA, and Brazil’s ANVISA are all — each at their own pace — developing specific guidelines for the use of AI agents in clinical trials. The FDA’s framework for real-world data in regulatory submissions, for example, makes it clear that the agency expects traceability and human accountability at the decision point. In practice, this turns fully autonomous recording functions into a regulatory risk, regardless of the system’s technical capability. The point of convergence across all three agencies is pretty clear: system autonomy is tolerated and even encouraged at operational stages, but decisions affecting participant safety and data integrity need documented human traceability.
This means that even when an agent makes a technically correct decision, if there’s no auditable record showing that a qualified human validated that decision in the right context, the data can be challenged during a regulatory inspection — which could compromise the entire trial. This regulatory stance has very solid logic behind it. The approval system for drugs and therapies was built on the principle that every step of the process can be verified, questioned, and reproduced.
The introduction of AI systems with a high degree of autonomy challenges this principle because it creates black boxes inside a process that needs to be transparent. The answer that’s emerging isn’t to ban the agents, but to require that they operate with explainability — meaning every relevant decision can be described in language that a human expert can verify and validate. This is pushing the AI industry toward developing more interpretable models, which, interestingly enough, is generating benefits that go well beyond clinical trials.
In Brazil, ANVISA is still in a more exploratory phase of this debate, but internal discussions already acknowledge that the arrival of AI agents in trials conducted in the country is a matter of when, not if. And regulatory preparation needs to happen before the technology arrives at scale — not after. This alignment between technological innovation and regulatory maturity is one of the biggest challenges clinical research will face over the next five years, and how that balance is built will define the pace at which new treatments reach patients. 🧬
