Google Cloud launches Gemini for healthcare AI agents in preview

Clinicians spend nearly 28 hours a week on documentation and administration — and that number is the reason Google Cloud gives for launching Gemini for healthcare, a purpose-built agentic software platform for hospitals and clinics now available in preview.

The figure comes from a survey Google Cloud conducted with The Harris Poll. Two more numbers frame the problem the company wants to solve: a 2025 American Medical Association report puts physician burnout at 41.9%, and the American Hospital Association estimates that 25–35% of total US healthcare spending goes to administration. Gemini for healthcare is Google's bid to shrink those numbers with AI agents that handle both clinical and administrative workflows.

What does the platform actually do?

The system pulls together data that usually lives in separate silos: clinical, financial and administrative systems, electronic health records, billing data, medical guidelines, and information from the US government programs CMS and Medicare. Technically, it rests on four components — functional skills, connectors, a partner network and governance functions. On the governance side, customers aren't locked to Google's own models; they can also select models from Anthropic or OpenAI.

Google says the offering is designed to meet the requirements of HIPAA, the US health privacy law. It includes audit logs and source citations, and customer data is not used for model training.

One design decision stands out: all connections to established clinical systems are read-only, with no write-back. Integrations with Epic run through open programming interfaces, while Oracle Health connects via standardized FHIR interfaces. A NextGen connection is still in development.

Access requires a Gemini Enterprise Standard or Plus subscription, billed either through subscription models or usage-based pricing.

Who is already using it?

The first adopters read like a who's-who of US healthcare: Northwell Health, Highmark Health, HCA Healthcare, Optum, Blue Cross Blue Shield of Massachusetts, Hartford HealthCare and the Stanford School of Medicine. Medical guidelines come from the professional society ASCO.

One operator's metrics show what automation can deliver. At Highmark Health, 78% of digitally submitted outpatient prior authorizations are granted automatically — no human in the loop for more than three out of four requests.

How do partners plug in?

At launch, Google Cloud is wiring in specialized partners through standardized interfaces:

• Infinitus provides the plugin Infinitus Relay, which automates insurance verification and prior authorization checks. It is available to existing customers through the Google Cloud marketplace.
• Keragon, an integration specialist, says it has already automated workflows for more than 1,000 healthcare organizations. It connects the platform to over 400 applications in practice management, scheduling and billing.
• InterSystems, a software house founded in 1978, headquartered in Boston with 38 offices in 28 countries, is connecting its server to give AI agents structured data access. The company says it serves more than one billion health records worldwide.
• IMO Health is integrating a knowledge graph with deterministic medical terminology.
• UiPath, the automation firm, is offering joint customers a trial version for controlling agents at the enterprise level.
• Egen, a partner with more than 700 technology specialists in Canada, Europe, India and the US, supports implementations on data privacy and human oversight.

Why read-only matters

The decision to make every clinical-system connection read-only, with no write-back, is the quiet headline here. It means an agent can read a patient record in Epic, pull billing data, check a guideline from ASCO — but it cannot alter the record itself. For hospitals weighing AI adoption against liability and patient-safety concerns, that boundary lowers the barrier considerably.

The governance story works the same way. Audit logs and source citations make each agent's output traceable; the promise not to train models on customer data addresses the fear that a clinic's records could leak into a general-purpose system. And letting customers swap in Anthropic or OpenAI models under the same governance layer signals that Google is positioning this as infrastructure rather than a single-vendor lock-in.

What comes next?

With the NextGen integration still in development and the partner network expanding at launch, the preview version is clearly a first step rather than a finished product. But the read-only architecture, the HIPAA alignment and early metrics like Highmark's 78% auto-authorization rate suggest Google Cloud is aiming squarely at the 28 hours a week that clinicians currently lose to paperwork.