Toward an Era of AI-Enabled Clinical Collaboration

May 19, 2025 | Eric Horvitz - Chief Scientific Officer, Microsoft

Returning to clinical medicine to complete my MD/PhD training at Stanford University, after finishing a PhD in AI, was full of eye-opening learnings about the real world of clinical decision making. I had spent years immersed in the theory and application of high-stakes decision-making under uncertainty, focusing on leveraging AI in time-critical medical contexts. Yet the lived reality of clinical care brought entirely new dimensions, with entering patients rooms and getting to know their detailed journeys in health and illness, engaging with families, navigating the fast-paced workflows of hospital floors, and witnessing the nuanced dynamics of care teams and their passion for delivering the best care possible.

One of the most formative and inspiring experiences was participating in tumor board meetings.

Across hospitals worldwide, tumor boards are foundational to multidisciplinary cancer care. These high-stakes, high-collaboration, resource-intensive gatherings bring together oncologists, pathologists, radiologists, surgeons, and other specialists to review complex patient cases. Experts from multiple specialties consider multimodal data patient records, imaging, pathology, lab results, genomics and align on a treatment plan. When patients present with rare, ambiguous, or complex conditions, there often isn t a clear evidence-based path. In those moments, decision making transcends individual expertise. Insights about the best steps forward emerge through a synthesis based on collective reasoning across domains.

The collegiality, intellectual rigor, and compassion in these discussions were inspirational. I found myself wondering: Could AI someday contribute meaningfully in this environment? Could AI tools help not just in narrow tasks, but as helpful connectors and assistants that could supercharge cross-specialty collaboration?

Conversation between Eric Horvitz & Shrey Jain, AI product lead, Microsoft Healthcare & Life Sciences.

At the time, tools at the frontier, Bayesian networks and supervised machine learning methods, were promising but limited in numerous ways. For example, they could not yet comprehensively analyze the large quantities of natural language captured in patient records, nor pull together patient findings, test results, and imagery that encode a patient s journey. They could not weave together patient-centric summaries of sets of relevant literature or perform inferences that consider the viewpoints drawn from multiple specialists. The vision of AI bridging specialties and contributing to complex care remained aspirational. Still, the idea stayed with me.

Polymathic capabilities inspire new AI possibilities

Fast forward to the last few years. With the emergence of large language generalist models like GPT-4 with polymathic abilities, strong progress with specialist multimodal models that combine language and imaging data to deliver expertise in pathology and radiology, and agent-based methods that capture multiple specialist roles, my long-held vision began to crystallize into tools that we could now build. Suddenly, AI tools that could help connect the dots across clinical domains, synthesize diverse information sources, and support collaborative, high-consequence decision-making felt within reach.

That spark from the past helped to ignite conversations, both inside Microsoft, as well as with colleagues working in biomedical research and clinical care at leading medical centers. We pursued the prospect of leveraging advances in AI technology, including generalist AI models, multimodal specialist models, and multiagent platforms for supporting tumor board preparation and real-time collaboration and decision making.

One of the first people I reached out to was Professor Sylvia Plevritis, a leading expert in cancer research and Chair of the Stanford Biomedical Data Science Program, to discuss possible projects and collaborations. The early conversations within Microsoft and with colleagues at Stanford soon grew into deeper work at Microsoft, at Stanford, and more recently in a fabulous partnership between Microsoft and Stanford Health Care.

Today, we shared an update on the work during Satya Nadella's keynote at our 2025 Build meeting, where we presented on Microsoft s Healthcare Agent Orchestrator.

From Concept to Clinic: A System in Deployment

The platform is now being deployed and piloted at Stanford Health Care. Together, we have set out to build something tangible to meet clinicians where they re at. I hope to soon see how the multiagent approach will be able to support preparation and real-time decision making at tumor board meetings. Stanford is on track to scale this system to support all 4,000 of its annual tumor board meetings.

I am hopeful that we'll see the methods extend tumor boards beyond its traditional boundaries into community hospitals and beyond that may not have access to the depth of multidisciplinary expertise available at teaching hospitals.

Imagine a clinician at a community hospital preparing for a tumor board case involving a rare sarcoma subtype. Rather than relying solely on local resources or referring the case out, the Orchestrator can provide real-time support: connecting with experts across the world, integrating contributions from specialist AI agents, synthesized insights from current literature, and structured analysis of multimodal data. The tools can amplify the capabilities of clinicians to support their patients.

We re also seeing momentum across the broader AI ecosystem. Paige, a leader in digital pathology, has developed a specialist agent, Alba, now available via the orchestrator to provide expertise in pathology analysis during tumor boards. This is a powerful example of how third-party agents can be integrated directly into clinical environments.  I'm also hopeful that making it easy to integrate the tools into widely available applications like Microsoft Teams will lower the barriers to integrating the capabilities into daily workflows, without requiring changes to infrastructure.

What Comes Next

For me, this marks a full-circle moment. From those early days sitting in awe as senior medical student at tumor board meetings, wondering if AI might one day play a role, to now seeing AI-powered specialist agents and their orchestration beginning to contribute to enhancing the timeliness and quality of real clinical decisions. I m filled with excitement and optimism.

We re still at the beginning. With continued collaboration among clinicians, researchers, engineers, and innovators, we have a chance not only to transform tumor boards, but to reimagine how medicine is practiced in high-stakes, complex scenarios.

The momentum is real and accelerating. As we expand deployments, integrate new types of agents, and incorporate feedback from the clinical frontlines, we're not just improving workflows. We're building a foundation for a new era of collaborative medicine, one where every patient, everywhere, can benefit from the collective intelligence of the world's best minds, amplified and made available via AI.

The best is yet to come.

More: Microsoft Healthcare & Life Science blog