Tammy Hawes is a healthcare tech executive who founded Virsys12 (acquired by Healthstream) and now serves as VP, Payer Solutions.
For the past two years, most of healthcare’s AI conversation has centered on the co-pilot—a smart assistant that drafts a note, summarizes a chart or answers a question while a human stays firmly in the driver’s seat. That framing is already starting to feel dated. Agentic AI—systems that don’t just suggest but act—is beginning to move into healthcare’s administrative core, and it will change how operations run long before it changes what happens at the bedside.
I’ve spent nearly four decades in healthcare technology—from writing software to sitting in the CIO and CTO chair to building and selling a company. That vantage point has taught me to be skeptical of hype cycles. However, the shift from AI that advises to AI that executes isn’t hype. It’s an architectural change, and leaders who understand where it lands first will be positioned to capture the value while managing the risk.
What ‘Agentic’ Actually Means
It’s worth separating generative from agentic because the distinction drives the risk profile.
A generative model produces content on request; ask it to draft an appeal letter, and it drafts one. An agentic system pursues a goal across multiple steps—retrieving data, calling APIs, applying business rules, making decisions and handing off exceptions—with limited human prompting along the way.
In practical terms, the instruction changes from “draft this appeal” to “work this case to resolution, and escalate only what you can’t.” That’s a meaningful leap in both capability and accountability.
Why Operations Goes First, Not The Exam Room
Healthcare’s administrative machinery is the natural first home for agentic AI, as it’s where the burden is heaviest and the clinical risk is lowest.
In a March 2024 McKinsey survey, 72% of respondents said their healthcare organizations—payers, providers and health-services companies alike—were already pursuing or implementing GenAI, with administrative efficiency among the highest-value use cases. Physicians, meanwhile, consistently point to cutting administrative burden as AI’s most valuable near-term contribution.
Back-office work—eligibility checks, claims triage, prior-authorization intake, provider and member communications, appeals—is rule-bound, high-volume and repetitive. It requires an agent to move a well-defined process forward without a person touching every keystroke. That’s exactly the kind of work where autonomy earns its keep and where a wrong step is recoverable rather than dangerous.
The Payer Opportunity
Payers sit on large volumes of structured data and run workflows governed by explicit rules—the ideal substrate for agents. Utilization management, claims adjudication support, member service and appeals are all candidates for agents that can triage a request, gather the supporting information, draft a response and route only the genuine exceptions to a human.
McKinsey has argued that this is a moment for payers to capture AI value now, and I agree—provided the deployment is disciplined. A well-scoped agent can compress cycle times that used to be measured in days into ones measured in minutes.
Where It Breaks If You’re Not Careful
Autonomy without accountability is the fastest way to lose trust in this technology. An agent that acts rather than merely suggests has to leave an audit trail—what it did, why and on what data. The most important shifts to get right:
• Human oversight has to move, not disappear. The human role changes from doing the work to reviewing and approving what the agent proposes to do as well as defining the thresholds at which the agent must stop and ask. Teams need to be reskilled toward judgment and exception-handling, not eliminated.
• Data quality becomes non-negotiable. An agent amplifies whatever data it’s given, faster than a person would. Clean, reconciled data is the precondition for letting a system act on your behalf.
• Governance and compliance can’t be retrofitted. HIPAA obligations, state regulations and the simple question of who owns an agent’s decision all have to be answered before scale, not after. Agents should inherit the data-access permissions of the user on whose behalf they act so that security is enforced by design rather than by policy alone.
• Bias travels with the model. As with any AI, agents learn from historical data that carries historical bias. In healthcare, where decisions affect access and cost, that risk has to be actively monitored, not assumed away.
How Leaders Should Prepare
The organizations that get this right will be the ones that deploy agents where the work is bounded, the data is clean and the accountability is clear.
Start with processes that are high-volume, rule-based and reversible—places where a mistake can be caught and corrected. Instrument everything the agent does so decisions are traceable. Define escalation thresholds up front. Build the governance framework before you scale, not in response to a problem. Invest in your people’s move from execution to oversight because that transition is where most implementations succeed or stall.
Agentic AI will redraw the line between what people decide and what systems do—and it will draw that line whether or not leaders are paying attention. The organizations that draw it deliberately, starting in the back office where the stakes are manageable and the returns are real, will set the pace for everyone else. The time to draw it is now.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?