Health systems are moving to build reliable, well-governed data foundations before they scale agentic artificial intelligence, betting that clean data and clear oversight will determine whether the technology delivers real returns. Providers see use cases ranging from boosting staff productivity to reducing data fragmentation, but leaders say those goals depend on the groundwork beneath the tools.
“AI can definitely increase efficiencies, but the key is getting it to increase productivity for the organization,” says Christian Lindmark, vice president and CTO at Stanford Health Care and the Stanford University School of Medicine. “That’s where governance really comes into play.”
What Makes Agentic AI Different
Agentic AI refers to systems that pursue goals with some autonomy. Unlike a chatbot that answers a single prompt, an AI agent can plan steps, pull information from multiple sources and trigger actions across software.
In a hospital setting, that could mean an agent that gathers information for a prior authorization request, drafts follow-up tasks for care teams, or reconciles records scattered across departments. Each of those tasks depends on the agent reaching accurate, current data.
That dependence explains the focus on foundations. An autonomous system that acts on incomplete or conflicting information can compound errors faster than a human reviewer can catch them.
Background: A Fragmented Data Landscape
Healthcare data has long been scattered. Clinical notes, imaging, lab results, claims and operational information often sit in separate systems, each with its own formats and access rules.
Mergers and acquisitions have added to the problem, leaving many large health systems running multiple electronic health records and legacy applications side by side. Interoperability standards such as HL7 FHIR have made it easier to exchange information, but adoption and implementation vary widely.
Agentic AI is both a promising answer to that fragmentation and a technology that exposes it. Agents can stitch together information across systems, but only if those systems are connected and the underlying data is trustworthy.
Governance Moves to the Center
Lindmark’s comments point to a theme gaining traction among health IT leaders: efficiency gains on paper do not automatically become organizational productivity. A tool that saves one team minutes may shift work to another team or create new review burdens.
Governance is how organizations decide which use cases to pursue, who is accountable for outcomes, and how results are measured. It typically covers data quality standards, access controls, model validation and ongoing monitoring.
For agentic systems, governance also has to define limits. Health systems must decide what an agent may do on its own, what requires human approval and how its actions are logged for audit.
Building the Foundation
Organizations preparing for agentic AI generally focus on a handful of priorities:
Data quality and consistency. Teams standardize definitions, clean duplicate records and resolve conflicts between systems so agents draw on a single, reliable version of the truth.
Integration and access. Providers use standards-based interfaces and centralized data platforms to give AI tools controlled access to information across the enterprise, rather than building one-off connections.
Security and privacy. Because agents can touch protected health information across workflows, systems must enforce HIPAA requirements and the principle of least-privilege access for software as well as people.
Oversight and measurement. Leaders define success metrics up front, such as time saved or reduced administrative backlog, and track whether the gains reach the organization rather than a single task.
Implications for Providers and Vendors
For health system executives, the message is that data readiness is a prerequisite, not an afterthought. Organizations that invest early in integration and governance may find it easier to deploy new AI capabilities as they mature.
Those that skip the groundwork risk pilots that perform well in isolation but stall when they meet messy, real-world data. They may also face compliance and patient safety exposure if autonomous tools act without clear guardrails.
Technology vendors face pressure as well. Health systems increasingly want AI products that fit into existing governance structures, support audit trails and work with standardized data rather than requiring bespoke integration.
Clinicians and administrative staff also have a stake. Productivity gains from agentic AI will depend on whether tools reduce burdens such as documentation and paperwork or simply reshuffle them.
What to Watch Next
Expect health systems to formalize AI governance committees and tie AI investments more closely to measurable operational outcomes. Watch for how organizations define the boundaries of autonomy for AI agents, particularly in workflows that touch clinical decisions.
Regulators and standards bodies are also likely to sharpen their focus on transparency and accountability for AI in healthcare. The health systems that pair sound data infrastructure with disciplined oversight will be best placed to turn agentic AI from a promising pilot into everyday productivity.
Frequently Asked Questions
Why can agentic AI make mistakes worse than a simple chatbot in a hospital?
Agentic AI plans steps, pulls from several sources and triggers actions across software on its own. If it works from incomplete or conflicting data, errors can spread through workflows faster than a human reviewer can catch them, unlike a chatbot that only answers one prompt.
Why doesn't a time-saving AI tool always raise organizational productivity?
Efficiency gains on paper can simply move work elsewhere. A tool that saves one team minutes may create new review tasks for another team. Governance addresses this by setting success metrics up front and checking whether gains reach the whole organization, not just one task.
What limits should governance set for an AI agent?
Health systems must decide what an agent may do independently and what needs human approval. They also need to log its actions for audit. These limits sit alongside the usual governance elements: data quality standards, access controls, model validation and ongoing monitoring.
Does adopting HL7 FHIR solve the data fragmentation problem for agentic AI?
Not entirely. FHIR makes exchanging information easier, but adoption and implementation vary widely between organizations. Mergers also leave many systems running multiple electronic health records and legacy applications, so agents still need connected systems and trustworthy underlying data to work well.
How does security apply to AI agents that access patient information?
Because agents can touch protected health information across workflows, systems must enforce HIPAA requirements. They should also apply least-privilege access to software as well as people, so each agent can reach only the data it needs, ideally through controlled, standards-based interfaces rather than one-off connections.
Why is centralized data integration preferred over building custom connections for each AI tool?
Standards-based interfaces and centralized data platforms give AI tools controlled access to information across the enterprise. This avoids fragile one-off connections, supports consistent security and oversight, and may make it easier to deploy new AI capabilities later because the integration and governance groundwork already exists.