By Dr. Ashrith Amarnath, Chief Health Officer,
California Health Care Safety Net Institute
For health care leaders considering AI adoption in primary care, Amelia Sattler, MD, a family physician and associate medical director with Stanford Medicine’s Healthcare AI Applied Research Team (HEA₃RT), brought both implementation experience and strong conviction to her presentation at the 2025 CAPH/SNI Annual Conference’s clinical forum, “AI in Primary Care and Population Health.” In her role at HEA₃RT, she works with a team that rapidly and rigorously evaluates AI tools and thoughtfully integrates them into clinical care settings.
In her discussion, Sattler described a personal reason for wanting clinicians to have a central voice in how AI is used: although her father, a primary care physician, loved technology, the transition to electronic medical records drove him from practice. She is determined to help ensure the transition to an AI-enabled world does not repeat that pattern for other clinicians.
With that perspective, she walked attendees through five moments in the care journey of Juan, a hypothetical patient, to show where AI may help keep patient and clinician needs at the center, along with practical advice on adoption.
BEFORE VISIT
1. Quickly understand a patient’s condition before the visit
AI tool: Chart review support
To understand Juan’s clinical picture before seeing him, Sattler described how long it can take to review a patient’s chart. With information – from hospitalization records to imaging tests – scattered across the electronic health record (EHR), clinicians spend significant time navigating between screens and tabs to pull together a patient’s story. A 2024 Journal of General Internal Medicine study found chart review accounted for the second-largest share of ambulatory physicians’ EHR time.
To help alleviate this chart hunting, Stanford Medicine has developed an internal AI-enabled chart review tool called ChatEHR, which allows clinicians to query patient information through a conversational interface. In her example, Sattler described how tools like this may help clinicians prepare to see patients “in a fraction of the time.” That is in line with a 2021 JAMA study suggesting that AI can speed chart review without sacrificing accuracy.
For health care organizations without the resources to build their own AI chart-review tools, similar capabilities are beginning to emerge in the vendor market. But these tools do not yet appear fully mature or standardized.
“AI is very good at summarizing and synthesizing information. And I’m most excited for the AI chart review tools to mature and advance.”
— Amelia Sattler, MD, Associate Medical Director, Stanford Medicine HEA₃RT
DURING VISIT
2. Access fast-changing evidence on demand
AI tool: AI-assisted clinical decision support
Although Juan’s congestive heart failure is under control, Sattler noted that his records showed his diabetes was not, as reflected in his high A1C. For her, identifying the best next step in his treatment illustrated a broader challenge: optimizing care is increasingly difficult when medical knowledge is expanding faster than any individual clinician can keep up with alone. One widely cited estimate projected that medical knowledge would double about every 73 days by 2020.
Sattler said that AI-based tools trained in current scientific literature can give clinicians on-demand support as treatment decisions arise. Using a conversational interface, clinicians can query these tools and quickly surface relevant evidence and guidelines.
She cited medical evidence-retrieval tools, such as OpenEvidence, as one example of AI-assisted clinical decision support. In Juan’s case, it could suggest an appropriate medication for his poorly controlled diabetes, with the clinician assessing the recommendation. Stand-alone or lightly integrated AI tools such as these can make it easier for clinicians to access and apply the most current evidence at the point of care.
“We can’t keep up with all of the new guidelines coming out. But this [AI-assisted clinical decision support] helps us access the most up-to-date evidence, so I am giving patients the best possible care.”
— Dr. Sattler
DURING VISIT
3. Provide patients with personalized instructions
AI tool: Tailored patient instructions
At the end of Juan’s visit, Sattler said she would ordinarily give him
separate handouts for congestive heart failure, diabetes, and each of his medications, along with an action plan. In the limited time of a typical visit, patients like Juan may be left to piece those instructions together.
“That’s a lot for patients to digest and then ultimately act on,” she said.
Sattler noted that many of these post-visit materials are generic rather than tailored to a patient’s full clinical picture. In addition, they may not reflect the patient’s language, literacy level, or culture, making them harder to understand and use.
As an alternative, she pointed to an AI tool that can generate one cohesive set of after-visit information and guidance matched to the patient’s conditions, medications, language, and needs. Drawing on information from the chart and the visit’s documented plan, such as new orders, the tool can create a draft for the clinician to review.
Research suggests that this approach may be promising. A 2022 NeurIPS study found that physicians rated AI-generated draft patient instructions as more accurate and more helpful than comparison approaches. Importantly, the study envisioned those materials as drafts for clinician review and editing, not as a substitute for clinical oversight.
Depending on the health system, this capability may sit within the EHR or come through a tool integrated with the EHR workflow.
AFTER VIST
4. Preserve clinician time and reduce burnout
AI tool: Ambient AI scribe
Once Juan’s visit ends, Sattler’s documentation work continues. She still must update the record and write the clinical note. That burden is substantial: the 2024 Journal of General Internal Medicine study found that documentation accounted for the largest share of ambulatory physicians’ EHR time.
Sattler pointed to ambient AI scribes as one way to reduce that burden, much of which falls after hours. She noted their ability to free up clinician time so they can spend more time engaging directly with patients. These tools capture the visit conversation and generate a draft clinical note for the clinician to review and edit.
A 2025 JAMIA pilot study found that physicians using an ambient AI scribe reported lower burden and burnout, along with improved usability and perceived time savings. Even so, this use case depends on having an ambient AI scribe or similar documentation tool in place. Despite spreading rapidly, these tools do not yet appear to be standard across all health care systems.
“There’s incredible value from a clinician’s perspective to decrease the documentation burden. AI scribes help with that.”
— Dr. Sattler
BETWEEN VISITS
5. Respond to patient messages more quickly
AI tool: Portal message triage and drafting
One month later, Juan messages Sattler through the patient portal to say that he is short of breath and noticing swelling in his feet. She noted that since the pandemic, clinicians have faced a sharp increase in between-visit portal messages, driven by new symptoms, follow-up questions after visits, or queries about testing or imaging. For many clinicians, that volume now outpaces their capacity to respond as quickly as they would like.
Sattler pointed to AI tools that can draft portal-message replies for clinicians to review and edit. She emphasized that the clinician retains the option to keep, revise, or rewrite the draft, preserving human oversight.
In Juan’s case, she described an AI-supported exchange that could quickly gather additional clinical details, including his difficulty lying flat, and help route him to an on-call nurse, who could arrange a next-day visit. That timely escalation could allow earlier intervention, for example, by treating fluid overload before it leads to another hospitalization.
There is also emerging evidence that this approach may reduce clinician burden. A 2024 JAMA Network Open study found that, within five weeks, clinicians adopted AI-generated draft replies for about 20% of patient inbox messages, with reported improvements in burden and burnout measures, though not in time spent.
In practice, portal-message drafting is usually a capability within the EHR and patient portal rather than a separate standalone tool. This capability is gaining traction, though it is not yet universal.
Advice for adopting AI tools
Throughout her presentation and during the panel discussion, Sattler offered practical advice for health care leaders thinking about AI adoption. Her central message was that AI should reduce burdens that keep teams from working at the top of their licenses and limit meaningful interactions with patients. Some of her advice included the following:
Start with a clearly defined problem
Sattler urged organizations to begin by asking a simple but essential question: What problem are we trying to solve? She argued that AI is most useful when it is closely tied to a need that has already been clearly defined, rather than adopted in search of a purpose. She said Stanford Medicine’s HEA₃RT is “adamant” about this problem-focused approach because evaluating and implementing AI takes time and resources, and change management can be tricky.
“You are approached by AI companies all day long. But what is the problem you’re aiming to solve?”
—Dr. Sattler
Involve stakeholders early
Before deciding whether AI might help, Sattler recommended engaging the people closest to the problem, including patients, physicians, nurses, medical assistants, call center staff, care managers, and others. That early input from key stakeholders helps ensure any AI tool reflects the realities of clinical practice and frontline workflow. Without such input, organizations might prioritize the wrong tools or find that even promising solutions fall flat in practice.
“We can develop tools all day long, and they can look wonderful in a testing environment, but how they’re integrated into the workflow is more important than the tool itself.”
—Dr. Sattler
Look to your EHR first
When exploring AI solutions, Sattler recommended first looking to the EHR. Many AI capabilities relevant to primary care are already being offered through EHR vendors or are on their product roadmaps, which makes them a natural starting point. She noted that if an AI solution is already on the EHR vendor’s roadmap, for example, her institution tends to be hesitant to seek outside partnerships or build its own version, given the cost and complexity of integrating external tools. Where organizations do look beyond the EHR, she advised focusing on vetted products rather than prototypes, particularly in higher-risk settings.
“I’d be very wary of prototypes, especially for safety-net systems.”
— Dr. Sattler
Add an intelligence layer when needed
If EHR AI tools do not fully meet a health system’s specific needs, Sattler noted that organizations may need to build an additional intelligence layer on top of them. For example, that might include refining prompts so draft responses better match the desired tone or provide more information about the sender.
This type of layering helps move the tool beyond a one-size-fits-all design, so it can more closely align with the health system’s own operational needs, workflows, and communication style.
By the end of Sattler’s presentation and the discussion that followed, her message was clear: the current AI opportunity is not just about adding new tools, but about shaping a transition that better supports patients, clinicians, and the work of care itself.
As she said, “I hope that we can make a transition to an AI-enabled world that is supportive and helps us do our work better so we can all feel justified and excited about it.”