Start with one workflow, not a grand AI strategy
Medical group leaders looking for a practical entry point into AI in health care do not need to begin with a sweeping transformation plan. The strongest use cases start in the places where clinicians and staff lose the most time: documentation, scheduling support, internal drafting tasks, and other repetitive work that drains attention from patient care.
Gena A. Graves, D.H.Sc., M.A., B.S., vice president of home and community services and value-based strategies at HealthPartners in Minnesota, framed AI as a problem-solving tool rather than a technology project. That approach matters because many practices fail when they adopt software before defining the clinical or administrative bottleneck it is supposed to fix. In a setting where every minute has downstream consequences, even modest gains can change the pace of a clinic day.
Her message was direct: identify the problem first, then choose the tool.
Ambient listening moved from pilot to scale ahead of schedule
HealthPartners traced its AI experience back to predictive risk scores used to identify patients who might benefit from specific programs. Those early systems did not carry the flash of newer generative tools, but they already relied on algorithmic pattern recognition to help guide population health and care management. In practice, that kind of work has been part of health care for years, even when no one called it AI.
A more visible example came with the organization’s rollout of an ambient listening tool. The system reached its targeted scale six months ahead of schedule and reduced the time clinicians spend on notes and after-hours work. In a profession where documentation burden is tightly linked to fatigue, the value is not abstract. Less evening charting can mean more cognitive reserve during patient visits and less friction at the end of the day.
Clinicians still review every AI-generated note. That safeguard is not a gesture; it is the core of safe deployment. Ambient tools can ease the administrative load, but they do not replace the clinician’s responsibility for accuracy, context, and clinical nuance. A note that looks polished can still miss the reason a symptom matters, the family history behind it, or the subtle detail that changes management.
Trust grows when the tool solves a visible problem
Skepticism around AI is understandable. Many clinicians have seen digital tools added on top of already broken workflows, creating more clicks, more alerts, and more cleanup. Graves described a different sequence: start with a real operational problem, show how the tool reduces friction, and allow the results to earn trust over time.
That order is clinically sensible. Workflow redesign is often as important as the technology itself. If intake, documentation, or task routing remains inefficient, AI simply speeds up the same flawed process. In that setting, automation amplifies disorder rather than relieving it.
The cleanest use cases are often the least glamorous. A tool that drafts a meeting agenda, updates a job description, or builds a new-hire orientation checklist may sound mundane, but these are exactly the tasks that absorb staff time without directly improving care. When a practice removes administrative drag, it returns attention to patients, trainees, and team coordination.
Human review remains the safety boundary
AI may alter the texture of work, but it does not erase the human side of medicine. Graves emphasized that health care is deeply human work and will remain so. The statement reflects a practical truth rather than a philosophy: patients do not present as data fields, and clinicians do not make decisions from one input alone.
In documentation, especially, human oversight is essential because the clinical note is not just a transcript. It is a legal record, a communication tool, and often the basis for follow-up care. An AI system can help construct the frame, but the clinician has to confirm the content, decide what belongs in the chart, and correct anything that distorts the encounter.
For that reason, AI should be treated as a drafting assistant, not an autonomous author. The technology can compress low-value work, but responsibility stays where it has always been: with the clinician and the organization that deploys the system.
Governance comes before experimentation
Before a practice starts testing AI tools, it needs to know which tools are approved, what policies govern their use, and who oversees those decisions. Without that structure, experimentation can drift into inconsistent use, privacy concerns, and uneven quality control.
Governance does not slow innovation; it gives it a boundary. When teams understand what is allowed, they can experiment more confidently within the rules. That clarity becomes especially important as practices move from simple writing assistance to systems that touch scheduling, triage, prior authorization, or population health workflows.
For leaders who are unsure where to begin, the advice was practical and unadorned: start small, practice, and explore. In the early stages, the highest-value questions are not about replacing staff or reorganizing the entire enterprise. They are about whether a tool saves time, reduces friction, and still allows the team to maintain quality and accountability.
AI will change work, but not the clinical relationship
The workforce impact of AI is likely to be uneven. Some tasks will shrink, others will change shape, and new oversight responsibilities will appear. What will not disappear is the patient’s expectation that a human clinician understands the story behind the data. A generated note cannot reassure a worried parent, interpret a hesitant answer, or notice when a patient’s tone changes in the room.
AI is already embedded in health care through risk scoring, workflow support, and documentation tools. The more important question now is not whether practices will use it, but whether they will use it deliberately. A narrow, well-chosen starting point is often more effective than a large vision that never leaves the slide deck.
For practices still at the beginning, the first step is modest: pick one repetitive task, confirm the policy framework, and see whether the tool actually helps. The broader lesson is less about technology than execution. AI works best when it is tied to a defined workflow, supervised by clinicians, and expanded only after it proves useful in real day-to-day practice.