AI

Doctors disagree on whether AI’s time savings will reach patients

A discussion of consumer chatbots turned into a debate over payment incentives, clinical accountability and how tightly to control AI as medical risk increases.

Author
HDM Staff
Save
Share
Ami Bhatt, Carolyn Jasik, Haider Warraich and Mario Aguilar seated beneath the Bridging the AI Gap session title and participant names.
Ami Bhatt, Carolyn Jasik, Haider Warraich and moderator Mario Aguilar discuss the gap between AI adoption and patient care. Photo credit: HLTH USA 2026

    Physicians developing health AI split Wednesday over a central promise of the technology, whether relieving clinicians of routine work will give patients more time with their doctors. Haider Warraich, an ARPA-H program manager and practicing cardiologist, challenged the assumption that efficiency gains would translate into more personal care under the healthcare system’s financial incentives.

    “What is the incentive for that?” Warraich asked, questioning the prospect of clinicians spending longer with patients simply because technology makes other work faster. Carolyn Jasik, associate chief clinical officer at Verily Health, argued that AI could restore time for meaningful conversations, while acknowledging that consumers themselves might end up paying for more human attention.

    The exchange occurred during Bridging the AI Gap, the closing STAT at HLTH panel in Las Vegas. STAT health technology correspondent Mario Aguilar moderated Warraich, Jasik and WHOOP chief medical officer Ami Bhatt, following an introduction from STAT host Matthew Herper. Their disagreement connected the design of clinical AI to questions about who benefits when care becomes more efficient.

    The payment question behind the promise

    Jasik described the appeal of meeting patients eye to eye with less attention diverted to paperwork. “I actually think that this is going to bring us closer to our patients,” she said, explaining the relationship she hopes the technology will make possible. Warraich challenged the economic assumptions underlying that expectation and asked what would motivate a health system to devote newly available time to longer conversations.

    Jasik agreed that payment arrangements constrain what happens next. “Our reimbursement system is the problem,” she said, then described consumers’ willingness to pay out of pocket for more connection with clinicians. She anticipated a two-tier system in which some patients purchase additional time, raising a question about access alongside the prospect of improved relationships.

    Jasik also questioned whether a single AI service could fairly be held responsible for an entire clinical outcome when many organizations contribute to a patient’s care. She favored paying for the service delivered and suggested that a continuously available autonomous agent might fit a subscription model, combined with monitoring and accountability. That approach would require decisions about the activity being purchased and how its performance is measured.

    Bhatt offered a different route to better access. She described patients who receive advice at infrequent visits and struggle to make changes between them, and proposed using continuous information and support to help them in daily life. In her account, that work could also free appointment capacity for patients whose worsening conditions need a clinician’s attention. These were proposed benefits of a care model, with the discussion focused on how to organize it.

    WHOOP’s July appointment of Bhatt reflects the company’s own interest in connecting consumer health information with clinical care. The company described plans to expand that role as she joined. Her contribution therefore came from an executive working on the same transition the panel was debating.

    Patients have already changed the workflow

    Aguilar opened with the observation that patients already use general-purpose AI to make decisions about seeking care. A March KFF poll found that 32% of adults reported using AI for health information or advice in the past year. Among users, 65% identified getting quick, immediate advice as a major reason, putting speed among the practical pressures on a more carefully controlled clinical service.

    Warraich described two patients who told him that ChatGPT had prompted them to seek care for serious symptoms. One had experienced difficulty speaking and writing while shopping, and another sought care after Christmas and was found to have had a heart attack. His accounts illustrated how chatbot conversations are entering clinical histories, with the consequences of their advice becoming part of a physician’s work.

    Those individual experiences left Warraich asking what advice a clinician should give patients about the technology. The discussion supplied anecdotes, without comparative clinical outcome data from the patients described. His proposed response was to develop evidence about how people use these systems and what happens to them afterward.

    Jasik called the consumer chatbot environment a “shadow health care system.” She described a family experience in which a response about symptoms and treatment led to further interpretation by a parent, requiring her intervention as a physician. Her point concerned the whole chain from a generated answer to a person’s subsequent decision, including choices the model itself may never explicitly have recommended.

    Risk changes what a system should be allowed to do

    Bhatt identified a gap between the availability of immediate information and the clinical responsibilities surrounding care. “We cannot find our way in when we’re having a symptom,” she said of patients’ difficulty accessing the conventional system. She paired that access problem with the need for an accountable clinical pathway when an interaction could affect someone’s health.

    Bhatt said continuous information can “help you know your own normal,” describing several levels of activity from health education through understanding changes in physiology to escalation into care. Each level places different demands on the organization behind a tool. In her account, consumer information and continuous monitoring need a way to connect to a clinical response when a problem exceeds what a person can manage through everyday actions.

    Warraich explained how that distinction shapes ARPA-H’s ADVOCATE cardiovascular AI program. Its heart-failure work combines conversational capabilities with more constrained decision-making as the potential consequence grows. He described the approach as using “more deterministic systems as you go up the risk ladder,” alongside studies and simulation designed to inform the eventual scope of autonomous activity.

    ARPA-H’s September awards support a four-year, $62.7 million program, including up to $33.7 million in its first year. The initiative aims to develop patient-facing clinical AI, supervisory systems and deployment approaches, with FDA authorization among its goals. Its work remains developmental, making the planned evaluation and permitted clinical scope consequential parts of the program.

    Warraich said the program’s implementation work would compare different organizational settings, including Kaiser Permanente and a university-health-system approach involving Duke and Ohio State. He described planned use across 21 Kaiser centers in Northern California and asked whether an integrated payer-provider structure could make the economics easier to manage. ARPA-H’s award announcement confirms the 21-center Kaiser plan, with shadow-mode deployments and randomized trials among its planned evaluation methods.

    Convenience can pull patients toward a different system

    Jasik raised a commercial difficulty for developers that invest in clinical safeguards. A service may take longer, ask more questions and sometimes direct a patient to be seen, while a general chatbot can provide a more immediate answer. She argued that government leadership is needed as companies build services under those competing pressures.

    Bhatt also described a telemedicine encounter in which a patient preferred receiving difficult news at home with family, despite her own initial wish to speak in person. The experience shaped her view that technology can strengthen a relationship when the format meets a patient’s needs. It supplied a concrete example of the human benefit she wants healthcare organizations to preserve as they deploy new tools.

    For health-system leaders, the disagreement brought the intended use of saved time into the same discussion as technical performance. The panelists identified clinical safeguards, patient expectations and payment incentives as forces that will shape what an AI-enabled service actually delivers. Their exchange left an operational question for organizations making these investments, how to turn greater capacity into care that patients can reach and relationships they value.