Google Cloud's HIMSS 2025 healthcare AI showcase lands inside existing clinical workflows

What Google actually announced at HIMSS 2025

Google Cloud used its HIMSS 2025 post to name specific customers building on two products: Vertex AI Search for healthcare and its agent tooling, including Google Agentspace. The framing separates the work into 'AI agents' and 'search with gen AI,' and each named organization maps to one of those buckets.

On agents, Basalt Health is launching agents that prepare patient charts and handle administrative tasks for medical assistants, and can flag care gaps like overdue mammograms or colonoscopies. Google positions Agentspace as the layer that connects agents to data sources, defines workflows, and manages performance — the plumbing beneath a customer-built agent like Basalt's.

On search, four organizations appear: Freenome (prioritizing colorectal cancer screening candidates from deidentified data), Counterpart Health (a Clover Health subsidiary searching across a patient's record from 100-plus integrated data sources), MEDITECH, and Suki. All four are described as building on Vertex AI Search for healthcare.

The interface decision hiding in the customer list

The frontend story is where these features surface. MEDITECH's example is the clearest: it integrated AI-powered search and summarization 'directly into its electronic health records system, Expanse,' so clinicians get summarized records, labs, and notes 'all within their familiar workflows.' There is no new app to open.

Counterpart Health follows the same logic. Its Counterpart Assistant software puts a generative search experience 'at the point of care,' synthesizing recent tests, hospital discharges, and medication adherence where the clinician is already working. Suki's assistant lets a doctor ask, in natural language, about colon cancer screening timing given family history and get a concise generated answer — folding medical reference lookup into an assistant that already handles ambient documentation and coding suggestions.

Google's own line explains why this matters: healthcare is 'data-rich and information-poor.' The bottleneck these partners are attacking is not storage but retrieval inside a moment of care. That reframes the frontend problem — the win is not a better search page, it is eliminating the context switch away from the record, the chart, or the assistant a clinician already trusts.

Agents versus search are different frontend commitments

The two categories imply different interaction models. Search, as described for MEDITECH and Suki, is pull-based: the clinician asks, the system returns a synthesized answer. Basalt's agents are push-based: they prepare charts ahead of visits and surface flagged risks without being asked. Google notes agents are 'still in early stages' and describes exploratory uses like remote monitoring and drug discovery.

That candor is worth registering. The search examples are shipping features embedded in production tools; the agent examples lean on words like 'researchers are exploring' and 'early stages,' with Basalt described as launching. For teams evaluating this, the maturity gap between a summarization panel inside an EHR and an autonomous chart-prep agent is real, and Google's own language marks it.

Basalt's description also spells out the frontend-adjacent obligations: operating in a secure Google Cloud environment, synthesizing structured and unstructured data, and 'maintaining ethical standards, ensuring transparency and addressing concerns such as data privacy and algorithmic bias.' Those are the constraints that shape how an autonomous agent can be shown to and trusted by a medical assistant.

The implication: distribution runs through the incumbent system, not a new tab

The through-line of this announcement is that Google Cloud is not asking clinicians to adopt a new destination. It is supplying search and agent infrastructure to the vendors — EHRs like MEDITECH, assistants like Suki, care software like Counterpart — that clinicians already open every day.

For applied teams building clinical frontends, the practical read is that the defensible surface is the existing workflow. A retrieval or agent capability that lives one click inside Expanse or an ambient assistant clears an adoption bar that a standalone tool never will. The named partners in this post all chose embedding over destination, and that choice — more than the underlying model — is what makes these features usable at the point of care.