The next healthcare AI fight will center on work, not words. Jeremy Hudgeons, a senior sales specialist at GDT, points to a real operating mess: multiple EHRs, communication systems, authentication methods, and local variations in how the same patient interaction gets handled.
A chatbot can answer a question. An agentic AI system can pursue a goal with limited supervision, collect current information through APIs or databases, reason through a multi-step task, and act on external systems.
The difference sounds technical, but executives should translate it into plain business terms. Google Cloud describes AI agents as systems that use reasoning, planning, memory, and autonomy to complete tasks on a user’s behalf. In a hospital, that could mean understanding why a patient called, checking the record, routing the request, scheduling the appointment, sending a reminder, and escalating when the case exceeds defined boundaries.
Patient access offers the cleanest starting point because it combines volume, cost, frustration, and measurable outcomes. Hospitals already operate in a payment environment where patient experience affects value-based incentives, efficiency measures, and quality programs. That makes scheduling, prescription refill support, inbound call routing, proactive outreach, and aftercare follow-up more than back-office conveniences. They shape how patients feel about care, how quickly staff can respond, and how much labor the organization burns on repeatable contacts.
That does not mean every scheduling desk should turn into an unsupervised automation project. The smarter path starts with a high-volume workflow, measures the current cost per interaction, defines the handoff rules, and tests whether automation lowers cost without damaging the patient relationship.
Hudgeons frames the business case around simple math: reduce time per interaction, reduce avoidable interactions, raise consistency, and measure actual output against predicted ROI. That discipline keeps healthcare automation from becoming another expensive pilot, especially as federal policy keeps pushing payers and providers toward more interoperable prior authorization and data exchange processes.
Clinical management raises the stakes. The FDA says clinical AI in software as a medical device can transform care by learning from health data, while also requiring careful lifecycle management. In practical terms, that means leaders need to separate administrative agents from clinical decision-support tools. An agent that updates a refill workflow creates one risk profile. An AI system that analyzes clinical notes, surfaces patient history, or influences a care decision creates another. Hudgeons’s point deserves emphasis: the physician still makes the decision. AI should make the relevant information easier to see, faster to retrieve, and simpler to act on.
The foundation layer decides whether any of this works. The Office of the National Coordinator says health data interoperability helps clinicians deliver safer, more effective, and more patient-centered care while giving patients and caregivers better access to electronic health information. Agentic systems depend on exactly that kind of foundation. Without trusted identity, secure data transport, clean system connections, and role-based access, the agent will either fail to complete the work or complete it in a way compliance teams cannot tolerate.
Governance therefore belongs at the front of the project, not at the end. The NIST AI governance framework gives organizations a structure for managing risks to people, organizations, and society. In healthcare, that structure has to become operational: define what the agent may access, what it may change, which tasks require human approval, when it must escalate, how logs get reviewed, and how performance gets validated. If those rules remain abstract, autonomy turns into exposure.
The ethical layer requires the same discipline. The World Health Organization says healthcare AI should place ethics and human rights at the center of design, deployment, and use. WHO guidance on large multi-modal models also warns that broad health care use has been predicted, while wide task performance remains unproven.
That principle becomes concrete when a patient shares sensitive information with an AI voice agent at 2 a.m. The patient does not care whether the system technically counts as a chatbot, agent, or orchestration layer. The patient cares whether it understands the request, protects the information, routes the need properly, and involves a human when the stakes rise.
Security controls must match the new access pattern. HHS explains that the HIPAA Security Rule requires administrative, physical, and technical safeguards to protect electronic protected health information. Agentic AI makes that obligation harder because a useful agent often needs broader access than a narrow tool. It may touch scheduling, records, call-center software, identity systems, and analytics. The more systems it can reach, the more carefully leaders must control credentials, permissions, audit trails, data retention, and incident response.
Physicians will accept this faster when leaders describe it as augmentation rather than replacement. The American Medical Association uses digital health language that emphasizes AI as assistive technology enhancing human intelligence, and its 2026 physician survey found broad professional use of AI alongside concerns about privacy and the patient-physician relationship. That combination should guide deployment.
Doctors and nurses do not need another screen that creates more work. They need agents that remove administrative drag, summarize reliable information, and leave judgment where it belongs.
The investment case should begin with the workflow, not the model. Hudgeons is right that much of the cost sits in the foundation: governance, advisory structure, cloud strategy, compute, bandwidth, secure transport, and integration with the applications where work actually happens. Leaders who skip that foundation may still spend heavily on pilots, model testing, and vendor demos, but they will struggle to show durable ROI. Leaders who connect strategy to workflow value can build a stronger case for responsible AI adoption by showing exactly which cost, delay, or error they intend to reduce.
The uncomfortable conclusion is that agentic AI will expose the operational truth inside healthcare organizations. Clean workflows, clear rules, interoperable systems, and disciplined measurement will create leverage. Messy processes, unclear ownership, fragmented data, and vague ROI claims will create risk.
The technology will matter, but the operating model will matter more. Healthcare leaders should stop asking whether an agent can sound human and start asking whether it can complete the work safely, measurably, and under control.
