Google Cloud, Dexcom and Recursion see AI agents shifting from demo to practical lab applications

AI agents are shifting from tech demos to practical business applications across life sciences organizations, according to speakers at Google Cloud’s life-sciences roundtable held today.  “Think of these as AI assistants that are going to handle complex tasks on their own, but instead of just answering the questions, they can actually take action,” said Shweta Maniar, global director, healthcare & life sciences at Google Cloud. Dexcom said its Stelo over-the-counter glucose sensor uses generative models to turn sleep, diet and activity data into personalized health insights.

Meanwhile Recursion Pharmaceuticals reported a 50% cut in inference costs after moving its drug-discovery foundation models to Google tensor processing units (TPUs). Google executives added that Agent Space, unveiled in December 2024, has become one of the fastest-growing products in the company’s enterprise lineup. A blog post on the general traction noted “tremendous interest” in the platform. In other words, more organizations across industries are starting to treat AI agents as business infrastructure rather than experimental tools.

AI agents are fundamentally changing the way that we are going to operate. —Shweta Maniar

Inching toward industrializing science

With McKinsey pegging life-sciences AI upside at $60 billion to $110 billion annually, the steady-but-sure path from demo to day-job is looking less like caution and more like common sense. The panel urged a “crawl-walk-run” rollout: tackle safe, retrospective analyses first, then inch toward real-time control as trust builds. “As we bring AI into the picture, we’re taking a very thoughtful step wise approach, starting with retrospective insights and then moving eventually to real time coaching, progressively building confidence and trust as the technology matures,” said Dexcom CTO Girish Naganathan. Recursion CTO Ben Mabey called its agents a “co-pilot. ”He elaborated that while today’s systems assist, the “cutting edge” involves “thinking reasoning models” potentially “replacing more and more of what the scientists have to do” with agents that “can do it autonomously.” This, Mabey suggested, would allow human scientists to “save the really hard problems for humans,” ultimately enabling Recursion to “turn it [drug discovery] into an industrialized problem.” Similarly, Naganathan added that generative tools let engineers focus on “creative brainstorming and problem-solving” and leave the rote work to machines.

Orchestrating multi-agent workflows

One of the next frontier for science could involve connecting individual AI assistants into coordinated workflows that span entire organizations. Google Cloud announced it was the first provider to create an “agent-to-agent interoperability protocol” that lets AI systems collaborate regardless of their underlying technology. On that front, the organization is working with more than 50 partners including Salesforce, ServiceNow and SAP. Microsoft recently joined the effort. For life sciences, this could mean seamless handoffs from discovery agents identifying drug targets to clinical-trial agents optimizing patient recruitment to manufacturing agents monitoring production quality. “Think about how much time could be saved with multi-agent collaboration across regulatory document generation and clinical trials,” said Lillian McNealus, director of outbound product management for cloud AI at Google Cloud. Yet such orchestration requires robust governance, she noted. Early generative AI adoption created “a fragmented approach within a lot of organizations” with “different tech scattered across” departments. The solution, McNealus explained, lies in centralized platforms that give enterprises “access to innovation” while maintaining “well controlled and governed” oversight, a balance that’s proving key as AI agents move from departmental experiments to enterprise-wide infrastructure.

Turning a data flood into an advantage

Life sciences companies may be uniquely positioned to capitalize on AI agents because of the sheer scale and complexity of data they generate. “Our approach at Recursion is built on three pillars: data, compute and people,” explained Mabey. “We don’t just generate data that’s focused on a particular program. We do fit-for-purpose data sets where we span the entire human genome and a vast chemical library.” Rather than using “a flashlight pointing in on one particular area,” Recursion has “a whole area of the map kind of lit up” for AI to analyze. This data advantage becomes key because, as Mabey noted, “AI multimodal models can look at lots of things and ingest things at a scale that humans can’t do.” That capability, in turn, can enable companies to surface “novel insights” from biological complexity that would overwhelm human researchers.

As Mabey explained, the company has deployed Agent Space to streamline its IT department, where staff ‘are no longer looking up all the information across all the systems, but rather, they’re just communicating with the agent’ to resolve tickets.  Meanwhile, Dexcom is embedding AI directly into patient-facing products like Stelo, connecting lifestyle data to glucose outcomes. This dual approach, using AI to optimize internal workflows while creating AI-powered products, suggests that aggressive early-adopter life sciences companies are beginning to fundamentally rewire how they operate. As Naganathan put it, the technology brings “engineering focus and product development focus on creative and brainstorming and problem solving.”

Data quality, compliance set the ceiling

Both CTOs said the shiny agent ideas collapse if the data stack is shaky or non-compliant. “We don’t just generate data for one program,” Mabey said. ” We build fit-for-purpose sets that cover the entire human genome and a large chemical library. That lights up the whole map for the AI.” Dexcom’s Naganathan added the regulatory guardrails: “We’re working very closely with the FDA to follow their generative-AI framework,” taking “extraordinary precautions” on privacy and security. In short, the smartest agent is still only as good as the data pipeline, and the audit trail, behind it. In January, the agency proposed a framework to “advance credibility of AI models used for drug and biological product submissions.” Later in the year, it noted a plan to phase out animal testing in drug development while mandating agency-wide use of genAI by June 30.

With AI agents increasingly capable of tackling complex autonomous tasks, the life sciences industry is gearing up to “save the really hard problems for humans,” as Mabey put it.