At a private conference steps from the BIO International convention in San Diego, 88 CEOs and senior executives from pharma, biotech, and AI spent hours debating a single question: what does AI mean for the future of global health care? The gathering occurred as AI agents already outperform human scientists in molecule design and clinical trial simulation, and the FDA signaled it will not be the sole regulator of these tools.
"It's not going to just take one person, one group, or one entity. It has to be all of us moving together," said Capt. Dianne Paraoan, acting director of the Office of Medical Policy at the Food and Drug Administration, who flew in from Washington for the event. The discussions took place at the Marriott Marquis, a block from the packed convention floor, with participants including Nvidia, venture capitalists, and AI engineers.
AI agents are running experiments and outpacing human scientists
During Nvidia's private panel, executives listed dozens of companies using AI agents to test molecules and run experiments. Some of those agents are outperforming human scientists in the lab. The agentic models can create new molecules, predict drug trial success, and simulate experiments in a fraction of the time.
San Francisco-based Chai Discovery is working to ship its agentic models now, inking over $1 billion in big pharma deals this year alone. The company built a model that can design molecules and antibodies for drug discovery. "A year ago, finding a viable antibody meant screening a thousand molecules to get one that worked. Today that success rate has jumped to 10% to 15%, two orders of magnitude in 12 months," said Josh Meier, CEO and co-founder of Chai. This is a concrete example of AI Agents & Automation reshaping how labs work.
These models save time, not money. Compute costs are rising. "When you let them run for days, you can actually accomplish months of intellectual work," said Andrew White, CTO of Edison Scientific, which created an AI scientist called Cosmo. Cosmo has not yet been deployed in a physical lab.
FDA launches new frameworks while stepping back from some oversight
On Monday, the Department of Health and Human Services and the FDA announced "Operation Trailblazer," a program designed to speed drug development from early investigation through late-stage clinical trials. The agency is also loosening some guidelines. Earlier this year, the FDA decided that recommendation-based AI tools do not need premarket review when used with clinician oversight.
At the same time, the FDA published a seven-step risk-based system for determining whether an AI model is reliable enough to inform high-stakes regulatory decisions. The framework is "useful" to developers, said Samuel Aronson, chief AI officer at AIwithCare, but much of the oversight remains with the companies themselves. Paraoan was direct: "FDA isn't going to regulate what we haven't regulated." Because AI lab work does not directly touch patients, the agency has no technical jurisdiction. The new program reflects the growing complexity of integrating AI for Healthcare into regulated environments.
Almost every AI health care company at the conference maintained that safety was their "No. 1 priority," including Hippocratic AI, which uses AI agents to educate and enroll patients in clinical trials. Paraoan said the guardrails are something companies need to think about to make sure "the public is comfortable with the tools." Some audience members later questioned the legitimacy of self-given safety benchmarks. "It's easy to pass a test that you write," one attendee said.
During the Nvidia panel, executives noted that the new agentic life science toolkit is available on GitHub for anyone with ChatGPT or Claude to begin experimenting with. "If it becomes a concern, we will, of course, abide by whatever the concern is, but at this point, being able to sequence a genome and get a clinical variant in the matter of minutes versus days is very good for humanity," a Nvidia representative said.
Why this matters for healthcare professionals
AI agents are already designing molecules and predicting trial outcomes with growing speed, and the FDA is not positioned to regulate the computational work directly. For healthcare professionals, this means that tools entering clinical settings may carry varying levels of external validation. Understanding how these models are built, tested, and monitored-and what safety benchmarks they actually meet-will become a core competency as AI-driven drug discovery and patient-facing applications expand.
Your membership also unlocks:
