Image courtesy of Adobe Stock
Google Cloud and multiple healthcare partners announced production deployments of AI agents on Oct. 16, timed to the release of Google Cloud’s second annual ROI of AI in healthcare & life sciences report and the start of HLTH 2025, held October 19–22 in Las Vegas. The announcements include clinical-note summarization at scale at Hackensack Meridian Health, a public breast-cancer screening assistant from Color Health, a multi-agent prior-authorization platform from IKS Health, and a “self-driving” real-world evidence system from Castor. All run on Google Cloud; several cite the Gemini model family.
Agentic deployments gaining ground
Nearly three in four healthcare and life-sciences (HCLS) executives who have deployed generative AI report positive ROI on at least one use case, according to a gated Google Cloud survey of 305 HCLS leaders. Sixty-two percent say they’ve already moved use cases into production, and 80% reached production within six months. Those figures set the context for Oct. 16 announcements showing hospital-grade “agentic” systems moving from pilots to daily workflows: clinical-note summarization at Hackensack Meridian Health (HMH), a breast-cancer screening assistant from Color Health, and multi-agent operations platforms from IKS Health and Castor.
Google Cloud’s HCLS readout (n=305) reports that 74% of executives with production gen-AI see ROI on at least one use case. A total of 62% have moved use cases into production while 63% report revenue increases. In addition, 80% reached production in six months or less.
A separate cross-industry study this fall finds about half of enterprises are already using AI agents in production (52%), with four in 10 reporting 10 or more agents deployed (39%). That suggests the shift from experimental chatbots to task-specific workflow integrations is well underway.
Agents are emergent but quickly evolving
Agent outputs are gaining ground against specialty reference sets. A 2025 preprint reports a roughly 79% compression and up to 18.2% AUPRC improvement for heart-failure prediction when using distilled summaries over full notes. Peer-reviewed studies are emerging but still limited in prospective, EHR-embedded evaluations; recent work shows clinicians can partner with EHR-integrated models on hospital-course drafts (JAMA Network Open, 2025), while broader reviews flag thin evidence on safety at scale (JMIR, 2025).
McKinsey estimates 50–75% of manual PA steps are automatable. But earlier electronic PA sped decisions without reliably cutting provider workload or form-filling burden. Production agents should show payer-verified first-pass approvals, shorter appeal cycles, and fewer denials, quantified with synchronized agent logs and payer timestamps.
Clinicians spend about 28 hours per week on administrative work. Multi-agent deployments should calculate “control tower” headcount supervising workflows at target volumes, then compare all-in costs, including inference, orchestration and clinical review minutes, to pre-AI baselines. For screening-coordination agents, such as Color Health’s breast-cancer eligibility and scheduling assistant, report uptake and completion by age, race/ethnicity, and payer; no-show reductions; and time-to-mammogram, with auditable trails for adverse-event review.
