At a recent BIOLINK Fireside Chat, Katy Ghantous, VP of Data Platform & AI at Veridix AI, and Cesar Yazbeck, VP of Healthcare at Invisible Technologies, cut through the noise around AI in clinical research. Their message: the technology is ahead of the operating model, most of the value lies in infrastructure work, and the path forward requires discipline.

The Clinical Trial Machine Is Broken

Katy opened by mapping the end-to-end clinical trial lifecycle, from scientific question and protocol design through site selection, patient recruitment, study conduct, data analysis, and regulatory submission. The key insight: most trials fail not because of operational breakdown, but because of the very first decision: wrong endpoint, population, and biomarker. And once the protocol is locked, the entire trial becomes a rigid, linear machine with very little room to adapt. AI’s biggest promise therefore is in helping researchers make the right decisions before protocols become difficult to change.

Cesar framed the executive misunderstanding that pervades: most leaders ask whether the AI models are good enough. The real question is whether the operating model can absorb them. “The technology is actually way ahead of the operating model,” he said. “The rate-limiting step is not how good the models are, it’s do you have a good way to embed it within the world you’re operating in.”

Point Solutions Are Not Enough

Both Katy and Cesar were candid about the gap between ambition and reality. The market is full of point solutions, but point solutions deployed within heavily regulated, cross-functional workflows can rarely make an impact on time or cost in a meaningful way. As Katy put it, a 50% efficiency gain on a task that represents a small percentage of total trial cost is “a drop in the bucket.” The incentive for pharma to adopt isn’t as strong when the ROI is that diffuse.

Katy shared a lesson learned the hard way. Products that require users to adopt both a new tool and a new process at the same time rarely succeed. After seeing several such efforts struggle, she adopted a different philosophy: build solutions that fit naturally into existing workflows while laying the foundation for the future.

Where the Real Value Lives Right Now

Cesar proposed a four-layer framework for thinking about AI in clinical trials: decision intelligence, workflow intelligence, evidence intelligence, and participant intelligence. His conclusion: the near-term bets should be on workflow and participant intelligence, rather than the decision and evidence layers.

On the workflow side, Katy argued that data quality review and risk-based quality management are among the most immediately deployable applications of AI. These data-intensive processes are well suited for AI-assisted anomaly detection, risk prioritization, and review workflows. Similarly, AI is rapidly accelerating dataset preparation, transformations, and quality control activities, becoming the new norm across biostatistics and statistical programming. Authoring tools powered by LLMs are also maturing quickly, though in regulated environments they still require domain expertise, traceability, and human review to be trustworthy.

On the participant side, Cesar described a real-world deployment: using FHIR API integrations into site EHRs to generate “patient 360s” (longitudinal, pre-vetted patient profiles filtered against a trial’s specific eligibility criteria) then coupling those profiles with AI-powered voice agents for outbound screening. The result is a dramatically narrowed top of funnel, with human connection preserved for the enrollment conversation that actually builds trust.

The Data Foundation Comes Before Everything

Data infrastructure is the precondition for everything else. Katy’s maxim “data science is 90% data engineering” captured it well. Cesar added that the FDA had recently announced two real-time clinical trial pilots, in which data would be shared with the agency live rather than at submission. His proposed implication: the entire linear model of clinical development could eventually give way to something adaptive and continuous, but only if the underlying data is machine-readable, standardized, and centralized.

“Before I automate anything,” Cesar said, “if I can just create 360 visibility and standardization in my documentation process and my operations, then I can access all the data if I find an opportunity to automate, intelligently understand what I have, map out workflows, and build a strong feedback loop for regulators.”

The Regulatory and Organizational Reality

Genmab’s Hisham Hamadeh, joining from the audience, raised a question that resonated with the room: in an industry still running on faxes, PDFs, and analog handoffs, are we automating things we don’t actually control? Cesar’s response: the analogy is agriculture. The capability existed long before irrigation made it scalable, and the challenge in clinical trials is the same. You need the data pipes before the intelligence can flow.

Hisham also observed that the companies most likely to push the boundaries are smaller players with less to lose that are willing to showcase the art of the possible, and whose successes will eventually give larger organizations the blueprint to follow.

On regulatory ambiguity, Katy was direct: FDA guidance on AI in clinical trials remains largely principles-based. The expectations are clear: human oversight, bias controls, monitoring, traceability, and auditability, but the implementation details are not. As a result, every organization is effectively developing its own validation framework, operating procedures, and governance model while the industry converges on best practices.

The Three Waves of AI in Clinical Trials

Cesar closed with a vision of how the field will evolve. The first wave, already underway, is productivity: workflow automation and accelerated recruiting. The second wave is decision-making: AI stress-testing human assumptions before a trial launches, and optimizing protocols before a patient is enrolled. The third wave is evidence: synthetic control arms, digital twins, and prognostic covariate adjustment techniques like Procova that shrink required sample sizes while maintaining statistical rigor. That third wave remains largely gated by regulatory acceptance. Synthetic control arms have gained the most traction in rare disease and certain oncology settings, though broader adoption remains gated by regulatory expectations around evidence quality and comparability.

The Bottom Line

AI in clinical trials is real, but it is slower than the hype and harder than the pilots suggest. The near-term value is in data infrastructure, workflow automation, and smarter patient recruitment–  not autonomous agents or AI-designed protocols. More broadly, Cesar argued that the industry’s biggest opportunity may not be deploying AI itself: “Most people think the opportunity is deploying AI. I think the bigger opportunity is helping define what best-in-class clinical development looks like in an AI-native world.” The organizations that will benefit most are those that build the data foundation now, deploy solutions that fit existing workflows, and resist the temptation to leap to the aspirational before the foundation is in place. “There is change, we are seeing adoption,” said Katy. The impact is real. It may be slower than the hype, but the industry will get there, and do so responsibly.