Insights from our conversation with Sandeep Pulim, M.D. Watch the full interview series here.
Open any healthcare technology newsletter, and you’ll find the same story told a dozen ways: AI is transforming medicine. Look closer, though, and most of that transformation is happening at the edges. The tools getting built and funded today are very good at things like automating paperwork. Yet the everyday friction of healthcare, from finding answers to reaching care, remains largely unaddressed.
In a recent conversation, Jon Shankman, Chief Analytics Officer at AMC Health, and Dr. Pulim unpacked that gap: where clinical AI falls short, why healthcare AI startups struggle to gain traction, what it takes to earn clinician and patient trust, and why rural America faces the highest stakes. Here’s what we learned.
Ask what AI is doing for clinical teams today, and the honest answer is mostly the routine. Today’s tools process lab results, handle refill requests, and streamline workflows for care teams. That work is undoubtedly useful, but it’s table stakes.
The bigger prize sits upstream of the visit. The real opportunity is not just handling documentation while a clinician is in the room with a patient. It’s surfacing intelligence about what’s happening with that patient before the encounter ever occurs, so the visit starts informed instead of from scratch.
That upstream shift points to a larger goal: giving patients more ways to manage their own care. New resources are making it easier for individuals and families to get answers and act without seeing a doctor or nurse in person for every question or concern. Preventing avoidable visits is not simply about convenience. For overloaded practices and underserved communities, it creates desperately needed capacity.
Getting there requires more than clever models. It requires trust and a regulatory framework that can keep pace with tools operating with some independence. Patients are increasingly willing to make a simple trade. They have data, and they want someone to look at it and help. The responsibility is to ensure oversight matches autonomy.
The answer, in a word, is workflow.
When a clinician evaluates a new AI tool, the questions are immediate and practical. How does this help me? How does it help my practice, my clinical team, my patients' workflow? How does it work with the systems that exist today, and how does it interface with current reimbursement? Startups that cannot answer those questions clearly don’t get adopted, no matter how impressive the underlying technology is.
Reimbursement deserves special attention, because it’s where good ideas often stall. Things are shifting with new access-to-care models, and Medicare is actively looking at how automation and AI-enabled services can get reimbursed. But many of the solutions on the market today depend on consumers paying out of pocket, and that’s a hard way to scale. The threshold question for any healthcare AI company should be: does this solution fit a use case that’s either reimbursable today or that clearly helps a practice monetize?
The good news is that a credible playbook is emerging. Several companies are doing the unglamorous work behind the scenes of making sure models are trained on properly tagged information and content, with clinical experts involved throughout. One startup recently built a voice digital biomarker library designed to help detect early neurologic, respiratory, and cardiac changes. Companies like these are standing up scientific and clinical advisory boards so that when they launch, they arrive with published peer-reviewed articles and real evidence that the work was done.
Respect for clinical medicine shows up in the fundamentals. Technology must fit the workflow, have a path to reimbursement, and prove its value with evidence.
Even a well-built tool can struggle to gain adoption. In healthcare, every new technology must overcome perceptions about what it can do, what it cannot do, and whether it belongs in the care pathway.
A recent example makes the point. Midjourney, known for its AI generation models, announced low-cost, ultrasound-based imaging hardware that could make it possible to visualize internal anatomy at a fraction of the usual cost.
The reaction exposed the tension. Skeptics pointed out that a ten-dollar scan is not healthcare. Interpretation, follow-up, and clinical decision-making still require expertise and access to the broader care system. Without that pathway, an inexpensive scan is just a picture, not a diagnosis.
Healthcare AI sits at this juncture. There’s real value in giving people affordable access to information about their own bodies. There’s also good reason to be cautious about assuming that information alone constitutes care. New technologies will inevitably be misunderstood, both by people who overestimate their value and by those who underestimate what they can enable.
The boundaries will become clearer with time. What matters now is being honest about what each tool can do, what it cannot do, and where it fits within the larger care journey.
Consider a thought experiment from outside healthcare.
Imagine a company that gives away free handyman services. In return, its workers wear GoPros that capture everything they do while repairing a pipe, fixing a furnace, or solving a household problem. The service itself is not the end product. The recordings are the asset. Over time, the company builds a massive dataset of how work happens in the physical world, creating training data for robots.
Now imagine applying the same logic to healthcare.
Aging safely at home is a physical-world problem. It involves the home itself, daily routines, mobility, caregiving, and the unpredictable moments when something goes wrong. Those details rarely appear in claims data, clinical notes, or medical images. Yet they increasingly determine whether an older adult can remain safely at home.
That gap matters because most healthcare AI is trained on what happens inside the health system. The next generation may need to learn from what happens outside it. Data captured in homes and communities could help create models that understand the relationship between people, their environments, and the care they receive.
The broader thesis is compelling. Healthcare AI may not just become better at reading the chart. It may become better at understanding the world around the patient.
Rural healthcare is where all these themes collide. Pre-visit intelligence can help clinicians make the most of limited capacity. Patient self-service can reduce unnecessary visits. Real-world infrastructure can help close the physical distance between patients and care.
The need is enormous. The median rural or community hospital has just 29 days of cash on hand, according to the Chartis 2026 Rural Health State of the State report. Hospitals are closing at an astonishing rate, leaving care deserts behind. The Center for Healthcare Quality and Payment Reform counts more than 700 rural hospitals, roughly one in three nationwide, at risk of closing, with hundreds at immediate risk within the next few years. As hospitals disappear, the problem is not simply that communities lose a building. They lose access, often across enormous distances.
Those distances matter most when specialized care is involved. Services like neonatal intensive care are especially scarce, with rural residents living an average of 50.6 miles from a hospital with neonatal services. That’s nearly five times farther than urban residents, and in parts of the rural West many communities sit more than 100 miles from the nearest hospital with obstetric or neonatal services. The same story repeats across other specialties, cancer care chief among them: about 20% of rural Americans live more than 60 miles from a medical oncologist, and one study of rural gynecologic cancer patients found they traveled a mean of 87 miles for care, with some traveling more than 200.
This is where AI and physical infrastructure begin to intersect. Drones and autonomous vehicles could help move supplies and services. Mobile care units could bring clinical capabilities directly into rural communities. ARPA-H is already exploring that model.
The bigger lesson is that healthcare access cannot be solved inside the hospital alone. It requires both the people who deliver care and the infrastructure that gets care to the patient. Technology that addresses both sides of that equation can do more than fill the gaps left by hospital closures. It can strengthen the rural systems that are still standing.
Pull these threads together, and a picture emerges. The AI we’re building today is real, but narrow: strong on documentation and routine workflow, weak on the harder problems of pre-visit intelligence, patient self-management, evidence-backed clinical tools, real-world data, and rural access.
Closing that gap is not primarily a modeling challenge. It is a trust challenge. Patients are ready to share their data in exchange for real help. Clinicians are ready to adopt tools that fit their workflow and their economics. Both need a framework where regulatory approval and oversight scale with the level of independence an AI model is given. Models that operate with more autonomy will, and should, face more scrutiny, and depending on the level of autonomy involved, some form of oversight and regulation will be required.
The teams that embrace that reality, building for workflow, for reimbursement, for evidence, and for the communities with the least access, will be the ones that close the distance between the AI we are building and the healthcare we actually need.
See how AMC Health helps health plans and rural providers extend clinical capacity with full-service remote patient monitoring here.
Sources