Healthcare AI Agents Stall at 3% Despite $440 Billion in Administrative Waste

Healthcare AI Agents Stall at 3% Despite $440 Billion in Administrative Waste

At HIMSS26 in Las Vegas last week, every major technology vendor showed up with healthcare AI agents. Epic debuted Agent Factory, a visual drag-and-drop builder for AI clinical workflows that run autonomously. Salesforce launched six Agentforce Health agents covering everything from referral routing to epidemiology. AWS released Amazon Connect Health. Google Cloud paraded Gemini-powered agents from CVS Health, Highmark, and Waystar. Oracle shipped an agent spanning 30 medical specialties.

The announcements landed in a market that, by all outward signals, should be ready for them. Eighty-five percent of Epic's customer base is already using some form of AI. Deloitte found that 61 percent of healthcare executives are actively building agentic AI solutions, and 85 percent plan to increase their investment over the next two to three years.

Then there is the number that reframes every booth demo and press release from the conference: only 3 percent of health systems have deployed an AI agent in a live clinical workflow.

That figure comes from research conducted by Microsoft and the Health Management Academy, published in the New England Journal of Medicine in January 2026. Forty-three percent of respondents said they are piloting or testing agentic AI. A full third reported no plans to explore it within the next one to two years. The gap between experimentation and production is not closing. It is calcifying.

The Vendor Sprint

HIMSS26 was not short on product. The vendor announcements were significant in both scope and ambition.

Epic's three named agents each address a specific operational pain point. Art, the clinical documentation assistant, saved more than 1,300 hours on discharge summaries in January 2026 alone, with a bedside nursing version launching in March and a home care version in April. Penny, deployed at over 240 organizations for revenue cycle management, helped top-performing customers achieve 20 percent fewer claim denials. Emmie, the patient-facing chatbot, handled more than 14,000 billing conversations at Rush University Medical Center and cut billing-related staff messages by 42 percent. Its AI patient scheduling capabilities achieved a 94 percent satisfaction rate.

Salesforce's six new agents under Agentforce Health target patient experience and operations. The referral routing agent reviews electronic health records (EHRs) to find specialists and schedule appointments, combining medical records AI with automated referral logic. The insurance claims agent provides patients 24/7 access to billing status and coverage details. An epidemiology agent helps public health agencies detect disease patterns. Amit Khanna, Salesforce's SVP and GM for Health and Life Sciences, framed the thesis plainly: "How can we give time back to clinicians and reduce time in paperwork?"

Google's partners brought production numbers. Highmark Health's Sidekick assistant grew from one million to over six million prompts in a single year, delivering an estimated $27.9 million in AI-enabled value for 2025. Waystar's AltitudeAI prevented more than $15 billion in denied claims and reduced appeal workflow time by 90 percent. Chris Schremser, Waystar's CTO, pointed to the structural driver behind those numbers: $440 billion in annual administrative waste across the U.S. healthcare system.

The rest of the exhibitor floor followed the same pattern. FinThrive showed agentic AI executing across 50-plus revenue cycle use cases, recovering nearly one million dollars in underpayments within three months for early adopters. Innovaccer's Flow Capture autonomously codes approximately 80 percent of clinical encounters. XiFin debuted an autonomous appeals agent that independently reviews denials, retrieves documentation, drafts appeal letters, and submits them to payors. VSee launched what it called the world's first fully autonomous telehealth robot, a step toward full telehealth automation, using LiDAR and infrared to navigate hospital corridors for virtual rounding.

Where the 97 Percent Gets Stuck

Healthcare AI agents clearly work in controlled settings. Several are producing measurable results. Yet the industry cannot get them into production at anything resembling scale.

The Microsoft and Health Management Academy research identifies a disconnect between executive confidence and deployment reality. Sixty percent of respondents agreed that agentic AI will meaningfully improve the provider-patient experience. Seventy-seven percent expect it to improve backend productivity. The organizations expressing the most optimism are largely the same ones without agents running in clinical settings.

Deloitte's survey sharpens the divide. Among early adopters, predominantly large organizations with more than five billion dollars in annual revenue, 82 percent prefer multi-agent architectures and 59 percent expect cost savings above 20 percent. Among smaller organizations in the 500 million to five billion dollar range, 87 percent are in wait-and-see mode, and 92 percent prefer single-purpose point solutions. The organizations with the least capacity to build governance infrastructure are the ones most likely to need it before they can deploy anything.

Regulation has not kept pace. The FDA has approved more than 1,300 AI medical devices since 1995. Agentic AI, meaning systems that reason, decide, and act autonomously, does not fit cleanly into those existing categories. The security implications compound the regulatory challenge. The agency loosened oversight rules for certain clinical decision support tools in January 2026, and a full regulatory framework for autonomous agents remains under development.

Tina Joros, VP of Policy and Innovation at Veradigm, described the situation at HIMSS26: "It remains a very complex environment, with few guardrails for the use of AI in healthcare, and still a lot of work to do to really create an environment that is going to produce safe, reliable artificial intelligence for clinical use."

State-level regulation adds another layer of fragmentation. Few states have a single body responsible for AI oversight in healthcare. Rules are scattered across insurance departments, attorneys general, and health agencies, creating compliance puzzles that are especially difficult for systems operating across state lines.

Danny Sama, VP of Digital Platform at Northwestern Medicine, raised the liability question underneath all of it: "At some point, some agent is going to make a mistake, and there will be a malpractice lawsuit, and, who's liable?"

The Infrastructure That Does Not Exist Yet

What separates the 3 percent from the 97 percent is not access to a capable AI model. The models exist. The agents exist. The operational scaffolding that makes deployment defensible does not, at least not in standardized form.

Production-grade healthcare automation demands audit trails documenting every autonomous decision, validation protocols that extend beyond pre-market review into continuous post-market monitoring, cross-platform governance spanning EHRs, claims systems, scheduling platforms, and patient portals, and scoped authorization ensuring agents act within boundaries a health system can defend. The FDA's predetermined change control plans were not designed for systems that adapt on their own.

MUSC Health is completing 40 percent of prior authorizations without human involvement. Sentara Health has reclaimed thousands of nursing hours through virtual nursing with ambient documentation. These are real deployments producing real results. They also represent organizations that built or assembled governance scaffolding themselves, an approach that does not scale across more than 6,000 hospitals, most of which lack the technical staff or budget to custom-build agent oversight.

athenahealth's HIMSS26 announcement pointed toward part of the answer: a Model Context Protocol (MCP) server allowing authorized AI agents to securely access patient data within athenaOne. Singulr AI launched Agent Pulse for real-time runtime governance of autonomous agents. AgentPMT operates in the same structural territory, providing audit trails, payment rails, and cross-platform governance that let autonomous agents operate within accountable boundaries regardless of which model or vendor built them. These are early signals of an emerging infrastructure category, one built on the recognition that deployment failures trace back to governance, and governance requires infrastructure that exists independently of any single agent or model.

RecovryAI's FDA breakthrough device designation, the first for a generative AI chatbot, illustrates both the opportunity and the gap. The company built an LLM-powered chatbot that monitors patients for 30 days after joint replacement surgery through twice-daily check-ins. The FDA designation validates the clinical approach. Scaling it across surgical specialties, health systems, and patient populations will require governance infrastructure that no single product company can provide alone.

What Moves the 3 Percent

The 440 billion dollars in administrative waste represents prior authorizations that take days instead of minutes, claims denied for documentation errors, nursing hours consumed by charting, and patient messages that sit unanswered. The agents demonstrated at HIMSS26 address each of these directly.

The deployment stall is an infrastructure gap, not a capability gap. Health systems will not build their own audit trails, payment rails, and cross-platform governance from scratch, not at the pace vendors are shipping agents. Whether the 3 percent figure moves depends on who provides that infrastructure and how quickly it matures.

Dr. Haider Warraich, an ARPA-H program manager developing clinical AI agents for cardiovascular care, framed the operational standard: "A system that has no human oversight would be unacceptable. The goal of the supervisory agent is, how do you create a system that can optimize human oversight, especially in the absence of ground truth?"

The healthcare industry does not lack AI agents. It lacks the infrastructure to run them responsibly. HIMSS26 proved the first part beyond any remaining doubt. The second part is where the real work starts.

Sources

• AI Agents Are Rapidly Spreading in Health Care, but Validation Is Lacking — STAT News
• Salesforce Releases Six New AI Agents for Healthcare — Healthcare Brew
• AWS Amazon Connect Health AI Agent Platform — TechCrunch
• AI Is Moving at Lightning Speed. Can Regulation Keep Up? — Healthcare Dive
• FDA Breakthrough Designation for Generative AI Chatbot RecovryAI — STAT News
• Assessing Healthcare’s Agentic AI Readiness — Microsoft Industry Blog
• Health Care Leans Into Agentic AI — Deloitte Insights
• Google Cloud to Showcase Gemini-Powered AI Agents at HIMSS26 — Healthcare Finance News
• HIMSS26: Epic to Highlight No-Code Agent Factory and Other AI Advances — Healthcare IT News
• HIMSS26 Pre-Day Recap: How Agentic AI Is Taking Over Healthcare IT — HIT Consultant