The Wrong Half of Healthcare Got Automated

She had been my doctor for a few years at a health center in Ohio, the kind built specifically to provide care to the LGBTQIA+ community, patients who have often already been through a system that required them to justify their circumstances, or their identity, before receiving help. She remembered not just my chart but the things I’d mentioned in passing: the job stress, the bad year, the symptom I’d brought up once and then decided wasn’t worth following through on. That kind of attention is hard to describe and impossible to fake.

Then the requirements tightened. Federal policy mandated electronic health records and tied reimbursement to their use. Internal quality programs required tracking dozens of performance metrics, each one needing structured data entry to count. Liability doctrine — “if it wasn’t charted, it didn’t happen” — meant every clinical judgment had to be documented before the encounter could close. Payers added prior authorization workflows on top of that: permission-seeking, in writing, before tests could be ordered or treatments could proceed. Performance evaluations started measuring documentation completeness. Each layer arrived with its own logic, its own forms, its own fields that had to be filled before the system would let you move on. She lasted about two years before she retired.

Before that, my doctor in New York, in a long-established Upper East Side practice, pivoted to concierge medicine: a subscription practice that filtered the patient panel down to people who could afford to pay out of pocket. A private workaround. The problem was still there.

The doctor before that, in Atlanta, complained about the computer from his first appointment with me. Not about medicine. About the new screen and keyboard positioned between him and me, and about what it meant that he was looking at it more than he was looking at the patient. He said it once, directly: “I did not go to medical school to be a typist.” He kept practicing. I don’t know how.

Three doctors. Three responses to the same pressure. None of them were wrong. The pressure was wrong.

The system worked as designed

Physicians in ambulatory practice spend about 27% of their time with patients. They spend nearly 49% on EHR and desk work. For every hour of face time, there are almost two hours of documentation. After the clinic day ends, there are another 86 minutes spent catching up on charts that didn’t get finished during office hours. Researchers have a name for it: pajama time.

This is not a technology failure. Electronic health records were built for billing, compliance, and data extraction. They were designed to serve payers and administrators. The patient encounter was an input. The documentation was the product. Human connection was not in the requirements.

When we talk about physician burnout as a crisis, we tend to frame it as a cultural or emotional problem: doctors overwhelmed, disconnected, heading for the exits. The actual mechanism is more specific. We took people trained to pay close attention to other people, sat them in front of a screen, and made attention to the screen the job.

The attention followed the incentives. It always does.

We’ve been asking the wrong question

Most of the conversation about AI in healthcare focuses on the diagnostic end. AI reading radiology scans. AI flagging drug interactions. AI as clinical decision support. The question underneath all of it is whether AI can do what *doctors* do.

That is not the interesting question.

The interesting question is whether AI can do what doctors have been forced to do instead of medicine: the documentation, the prior authorizations, the billing codes, the structured fields that exist to satisfy payers rather than serve patients. That question has a cleaner answer, and the early evidence is striking.

When ambient AI handles clinical documentation in real time, listening to the encounter and generating structured notes automatically, something measurable happens. At Mass General Brigham and Emory, clinician burnout dropped from 52.6% to 30.7%. Physician eye contact with patients increased. At the University of Chicago, the share of encounters where physicians gave patients their undivided attention went from 49% to 90%.

No wellness program has produced numbers like that. No workflow redesign, no resilience training, no administrative simplification initiative. The variable that changed was documentation burden. Everything downstream shifted with it.

Here is the reframe: the problem was never that physicians stopped caring. The problem was that we built a system where caring, being fully present with another person, was structurally incompatible with getting paid.

What human connection actually produces

There is a community health worker program called IMPaCT, developed at Penn Medicine and now operating in eighteen states. It puts trained community health workers alongside Medicaid patients with chronic disease: listening, following up, navigating systems, meeting people in their circumstances and in their language over months.

A randomized controlled trial found a 30% reduction in hospitalizations. The return on investment, from the Medicaid payer’s perspective, was $2.47 for every dollar invested, within a single fiscal year.

That is not a soft outcome. That is what happens when human connection and trust between provider and patient are treated as infrastructure: designed, funded, and measured rather than assumed or improvised.

A different example: Hippocratic AI deployed a multilingual outreach agent to contact patients due for colorectal cancer screening at WellSpan Health. Spanish-speaking patients are historically underscreened. The agent called them in Spanish, with patience, at scale. The opt-in rate for Spanish-speaking patients was 2.6 times higher than for English speakers. Not because AI replaced the clinical relationship, but because it handled the logistics of getting to the door in a language that didn’t make the patient feel like an afterthought from the first word.

The pattern in both cases is the same. AI manages the transactional layer. Human capacity (or something designed to extend a human quality like language and patience) does the relational work that actually changes behavior. Together they outperform either alone.

The harder math

The Medicaid access problem is not, at its root, a compassion problem.

Medicaid reimburses physicians at roughly 72% of Medicare rates, and Medicare rates are already below private market. Physicians lose an estimated 17.6% of visit value to billing overhead: rejected claims, prior authorization delays, documentation requirements that exist to satisfy payers rather than serve care. The math is punishing. Only 74% of physicians accept Medicaid patients, compared to 95% or more for private insurance.

A gastroenterologist I know in the Midwest describes billing high and receiving low as a kind of fiscal fiction you maintain to keep the practice solvent. Everyone in the system understands the fiction. Nobody designed it to work exactly this way, but nobody redesigned it either.

When AI automates prior authorization, billing submission, and claims processing, the unit economics of serving underserved communities change. The physicians who stopped accepting Medicaid patients, or retired at fifty-three, or went concierge: many of them had a billing problem, not a values problem. The system made caring for the patients who needed it most into the least financially rational choice.

There is another dimension to this that resists easy quantification. People who have experienced homelessness, food insecurity, or who have had their dignity worn down by a system that requires them to justify their need before receiving care, often describe the intake process itself as a barrier. Not the cost, though the cost matters. Not the wait time, though that matters too. The experience of having to prove your circumstances are bad enough, in the right format, to the right person, in language that isn’t yours. The weight of that.

Automating the administrative process doesn’t fix structural dignity failures. But administrative complexity that makes providers unavailable to the communities that need them most is a design choice. Design choices can be changed.

Presence before prescription

Appalachian Regional Healthcare operates in some of the highest-need, lowest-resource communities in the United States. Rural Appalachian Kentucky and West Virginia, where chronic disease rates are among the worst in the country and institutional trust cannot be assumed. It has to be earned. The model there is built around something that cannot be automated: sustained presence. Local hiring. Continuity. Community workers drawn from the communities they serve. The relationship has to come before the prescription because the history of those communities with outside institutions does not give trust for free.

Technology can extend a model like that. It cannot originate one.

The University of Cincinnati and Kroger ran a randomized controlled trial of supermarket-based dietitian counseling for patients with cardiovascular risk, published in *Nature Medicine* in 2022. The dietitian sessions, guided by purchase history data and conducted in the grocery aisles where patients already shopped, produced significant improvement in dietary quality. Adding a bundle of digital tools improved outcomes further. The active ingredient was the human expertise. The technology reduced the friction around it.

The lesson is not “add an app.” The lesson is that sustained human attention, deployed in the right place: where people live, in the context of their actual lives, changes things. Technology makes that model more scalable. The human presence is what makes it real.

What this actually requires

The design question is not complicated to state. AI should handle documentation, prior authorization, billing, outreach logistics, scheduling, and the structured data requirements that payers and administrators have layered onto clinical encounters over the last two decades. Humans should handle presence, relationship, earned trust, and the conversations that change behavior.

These two things are not naturally in competition. They have been forced into competition by a system that priced them incorrectly: treating documentation as mission-critical and connection as a byproduct that happens, if it happens, in whatever time is left.

The early evidence suggests that rebalancing this isn’t a long-term research project. It is an implementation problem. The tools exist. Ambient AI documentation is commercially available and already deployed at major health systems. Multilingual outreach agents exist. Automated prior authorization exists. The results, where these tools have been seriously deployed, are not marginal.

The question isn’t whether we can build AI that handles the paperwork. We already can. The question is whether the people who run health systems are willing to design for what happens when the paperwork is done.