Article
SEI Convened Healthcare Leaders to Discuss Where AI Agents Actually Deliver
The takeaway: well defined tasks and human accountability help AI agents deliver results.
Agentic AI is already delivering in pockets, mostly behind the scenes. The challenge now is to scale what works and decide how much autonomy each job can safely carry.
Panel discussion at SEI’s healthcare AI event in New York.
SEI’s NYC office gathered over 50 leaders in the healthcare space across private and public hospital systems, medical device companies, clinicians, healthcare providers, payers, technology leaders, and medical device innovators to explore what’s working today, and what it will take to build trust, drive adoption, and scale AI successfully.
The evening was moderated by Dennis Barger of SEI. Dr. John Halamka of Mayo Clinic Platform delivered the keynote, followed by a panel with Dr. Deepak L. Bhatt, Director of Mount Sinai Fuster Heart Hospital and Dean for Clinical Research and Industry Partnerships at the Icahn School of Medicine, Ammu Menon, Chief Data and AI Officer at NYC Health + Hospitals, James Green, CEO of Cognome, and Charlie Whelan, Senior Director of Consulting, Medical Devices at GlobalData.
Healthcare AI today thrives in operational workflows
Adoption is moving fastest in imaging and diagnostics, remote monitoring, and workflow support. Those tools still depend on connected data, devices, governance, and integration into daily work.
Operational agents take on bounded work
Panelists described the clearest current examples in operational settings where the job is narrow, repeatable, and easier to supervise:
Checks around care: Mayo Clinic Platform uses agents to support discharge planning, readmission risk, advance directive checks, and procedure readiness.
Less paperwork: Mayo summarizes about 12 million pages of outside records a year. More than 80% of its clinicians use ambient notes, which were reported to reduce documentation burden by 38%.
More outreach capacity: A 24/7 Medicaid agent handles work that would otherwise require 300 to 400 counselors. Automated pre-procedure calls have also drawn strong patient satisfaction.
Clinical decision support keeps a clinician accountable
The clinical examples were strongest when the signal was measurable and a clinician retained the decision. An ECG screening tool identified heart disease two to three years before it appeared; stroke decision support can help speed treatment. Panelists also discussed finding insights from similar patients. The tool informs the work, while the clinician decides what to do.
Some areas remain difficult. Attendees said current AI still falls short in coding and claims, reading clinical literature, and tracking outcomes. Dr. Deepak Bhatt noted that no tool had yet made their day dramatically easier.
SEI’s Dennis Barger opens the evening’s discussion with James Green of Cognome and Dr. Deepak Bhatt of Mount Sinai.
How to Move Forward with Agentic AI
Make the value case before scaling
Returns are easiest to explain when an agent takes responsibility for a defined job with a known cost. A project manager agent, for example, was built to cover a role the organization could not fill. It is harder to demonstrate value when AI only assists someone with an existing job. In some cases, a better workflow may solve the problem without AI.
The panel suggested judging each project by whether it improves at least one of three outcomes: patient health, the patient experience, or the working lives of doctors and nurses.
Costs need the same scrutiny as benefits. Agents can add platform fees and per-use charges. Mayo used three billion tokens in one month, and panelists said vendors are often not transparent about usage costs. Revenue cycle AI can create value while also adding to total healthcare costs.
Buy when a vendor has demonstrated market acceptance. Build only when the organization has the computing capacity and expertise to operate the system. Either way, start with a real problem and a clear measure of success.
Why pilots stall before they scale
Organizations often discover four to ten times more AI in use than they expected, and about half of it is abandoned. The panel identified six conditions that help pilots move into lasting use:
Plan the end state. Decide how the pilot will scale or end before it begins.
Keep it clinically relevant. A prediction should be capable of changing a clinical decision.
Fit the workflow. Put the tool where staff already work.
Have capacity to act. If AI identifies more patients, make sure a care team can respond.
Choose the right partner. Compare vendors carefully; many offer similar tools.
Plan for updates. Models change, so build in retesting and a way to retire them.
Healthcare’s complexity can surprise AI specialists, and many devices were not designed to share data. A strong pilot accounts for those limits from the start.
Set the right level of autonomy
Panelists drew a clear distinction between routine operational work and clinical decisions. A bounded operational agent can act independently when a manager remains accountable. Clinical decisions stay with clinicians, with AI providing support rather than taking over.
Routine refills may be a next step. An attendee cited a Utah pilot for renewing routine prescriptions, and a panelist supported the idea. The discussion also noted that clinicians can spend about half of each visit on administration. Agents may give some of that time back if they remove real work instead of adding another system to manage.
Keep people involved in the transition
Staff concerns about job displacement need direct attention. Nurses at the event noted that electronic health records had sometimes added work rather than reducing it. The strongest case for adoption is an agent that takes meaningful work off clinicians and staff while making the process safer or easier for patients.
Give every agent an owner
At Mayo, the review of an agent takes about two weeks. The process includes checking how it was built and how it performs across patient groups, validating it on local patients, matching review to its risk, and monitoring it after launch. Panelists emphasized that every agent needs a named owner. Physicians’ concerns about liability make a clear review process essential.
Performance also has to be checked in the setting where a tool will be used. A model that worked in Rochester performed worse on New York’s more diverse patient population. An agent’s ownership can involve an open model, cloud provider software, and tools developed by the health system. The panel noted that internally built tools may be treated differently for FDA review; organizations should assess the applicable requirements for each use.
Looking ahead: One possible direction is a single system where clinicians work in one place and that system calls the appropriate AI tools behind the scenes.
Attendees discussed where AI can relieve operational pressure and where it still needs work.
What healthcare leaders can do in the next year
Panelists encouraged organizations to focus on value rather than adopting technology for its own sake. They advised leaders to start with problems raised by frontline staff, strengthen data foundations and governance, and target the clinical labor shortage.
The most relevant use cases will depend on the organization:
Health systems: Documentation, revenue cycle, clinical decision support, and patient engagement.
Medical device companies: Monitoring, imaging and diagnostics, and clinical decision support.
AI startups and investors: Documentation, revenue cycle, connecting systems, and data.
Payers: Revenue cycle, care management, and patient engagement.
Academic and research organizations: Clinical research, decision support, imaging, and diagnostics.
Public sector and nonprofits: Patient engagement, benefits and eligibility, and connecting systems and data.
How SEI can help healthcare organizations scale what works
The gap between AI activity and value is rarely solved by technology alone. SEI’s approach is to help organizations choose the right use cases, prepare their data and oversight, and measure whether an AI system holds up in practice.
Plan: Benchmark against peers, prioritize use cases that can carry real workload, and assess data and AI readiness.
Build the foundation: Integrate data across EMRs and enterprise systems, structure content, and establish oversight and ownership.
Deliver and scale: Validate accuracy and safety, measure outcomes and cost, and scale only what performs well.
Scale the AI that proves it deserves to scale.
Start with the problem, make ownership clear, and measure what changes for patients, care teams, and the organization.
Our Accolades
