The front door of American health care is being rebuilt, and health systems are not holding the keys. The moment a person decides that something is wrong and begins to look for help now happens on a screen, and increasingly inside a conversation with an artificial intelligence the health system neither operates nor sees. Three forces have converged to move that moment upstream. Traditional primary care, the historical entry point, is in measurable structural decline: among commercially insured adults, primary care visits fell 24.2 percent from 2008 to 2016, and the share of adults with no primary care visit in a given year rose from 38.1 percent to 46.4 percent.⁴⁰ Patients have filled the gap with self-service: of ChatGPT's roughly 800 million weekly users, about one in four, on the order of 200 million people, ask it health questions, and a nationally representative KFF poll fielded in early 2026 found that one in three adults had turned to an AI chatbot for health information.⁹⁶ Technology and retail companies have spent tens of billions of dollars trying to own the new entry point, with mixed results that are themselves instructive.

The strategic consequence is direct. Whoever owns the intelligence layer at first contact owns discovery, triage, routing, and, over time, the longitudinal relationship. A health system that does not operate its own AI front door across discovery, triage, scheduling, navigation, after-hours contact, and follow-up will find that first contact defaulting to consumer chatbots, to platform players such as Amazon, or to third-party marketplaces. The clinical evidence makes the stakes concrete. Contemporary large language models are increasingly accurate on curated benchmarks, some now exceeding 90 percent diagnostic accuracy on clinical vignettes,²¹ yet they remain unsafe for unsupervised patient use: a structured stress test found that ChatGPT Health undertriaged 52 percent of gold-standard emergencies,² and when real patients used otherwise high-scoring models, they identified the correct condition in fewer than 34.5 percent of cases, no better than controls.¹

This paper argues that first contact is being disintermediated, quantifies the forces driving it, and gives operators a framework to respond. The central recommendation is a sequencing discipline: stabilize clinician capacity first with tools that already have strong evidence, then build or license an integrated AI triage layer that routes patients within the network rather than away from it, govern it for safety and equity before scaling, and treat the front door as owned, measured intelligence rather than a marketing channel. The systems that do this will convert the access crisis into captured demand. The systems that cede it will become the back office of care delivery.

The phrase "front door" has always been a metaphor for the set of decisions and surfaces that surround a patient's first contact with the health system. Historically that door was physical and narrow: a patient called a primary care office, waited on hold, booked an appointment weeks out, and was seen in a brick-and-mortar exam room. The interaction was a directory lookup followed by a phone tree. Its intelligence, such as it was, lived in the head of a scheduler or a triage nurse.

That door is being replaced by an interactive reasoning surface. The functions that define first contact have not changed: symptom interpretation, discovery of a provider or site, triage and routing to the right level of care, scheduling, navigation through a fragmented system, after-hours contact, and post-visit follow-up. What has changed is that each of these functions can now be performed, or at least attempted, by a large language model that a patient can reach at midnight from a phone. The front door has become a stack of software functions rather than a single physical portal, and that stack is contestable. It can be owned by the health system, partnered out to vendors, or ceded to consumer technology.

The question this paper addresses is therefore not whether AI will reach the front door — by every available measure it already has — but who will own first contact. That question is being resolved now, over the span of two or three product cycles, largely without the participation of the health systems whose patients are involved. The remainder of this paper examines the demand-side pressure pushing patients toward AI, the evidence on how they use it and where it fails, the consumerized decision journey that now precedes any clinical encounter, the supply-side reconfiguration that has reshaped who competes for the entry point, the economics of the opportunity, and the strategic choices available to operators.

The decline of the traditional front door is not a pandemic artifact or a temporary backlog. It is a standing structural condition, visible across every major dataset and more than a decade of measurement.

Among commercially insured adults, primary care physician visits declined 24.2 percent between 2008 and 2016, from 169.5 to 134.3 visits per 100 member-years, while the proportion of adults with no primary care visit in a given year rose from 38.1 percent to 46.4 percent.⁴⁰ Visits for low-acuity conditions, the traditional bread and butter of primary care, fell 47.7 percent over the same period.^40,49 The pattern holds in public programs. Among traditional Medicare beneficiaries, primary care visit rates declined from 2.54 to 2.27 per person-year between 2017 and 2023, and the share of beneficiaries with at least one visit per year fell from 61.9 percent to 59.8 percent.²⁰ Nationally representative MEPS data confirm a decrease in the rate of any primary care contact from 2002 to 2017.⁴⁵

The visits that remain have grown more complex, running longer and addressing more diagnoses per encounter, but the volume decline is unambiguous.⁴⁸ Meanwhile the relationship at the center of primary care is fragmenting. Person-based usual sources of care, meaning an ongoing relationship with an individual clinician, declined from roughly 27 percent to 15 percent between 1996 and 2014, while facility-based sources rose.¹⁶ The average Medicare patient now sees two primary care physicians and five specialists across four different practices in a year, and the share seeing five or more physicians rose from 17.5 percent to 30.1 percent between 2000 and 2019.⁴² This fragmentation is not benign. High care fragmentation is independently associated with more emergency department visits, more hospitalizations, departures from clinical best practice, and nearly double the health care spending in the highest versus lowest fragmentation quartiles.^74-76

The supply side explains much of the pressure. Mean primary care physician supply fell from 46.6 to 41.4 per 100,000 population between 2005 and 2015, and hundreds of counties have no primary care physicians at all.⁷⁷ HRSA projects a total physician shortage on the order of 187,000 by 2037, with primary care shortages most severe, and estimates that 76 million residents live in primary care health professional shortage areas.^78,92 The AAMC projects a shortfall approaching 86,000 physicians by 2036.⁹³ More than 100 million Americans report they lack a regular source of care.⁹⁴ Primary care practices were less likely to offer extended or weekend hours in 2022 and 2023 than in the years just before the pandemic, and out-of-pocket costs for problem-based visits rose 31.5 percent from 2008 to 2016.^40,48

The demand side compounds it. When a patient cannot get a timely appointment, cannot reach the office by phone, or cannot afford the visit, the rational response is to look elsewhere. Urgent care, retail clinics, and telemedicine absorbed much of this displaced demand: among commercially insured adults, visits to alternative venues rose 46.9 percent from 2008 to 2016,⁴⁹ and by 2024, 27.6 percent of Americans had at least one urgent care visit and 19.0 percent at least one retail clinic visit in the prior year.⁵⁰ The newest venue is a chatbot. An access deficit this large and this durable manufactures standing demand for any lower-friction front door, and an AI that answers instantly, at no marginal cost, at any hour, is the lowest-friction door yet built.

Patient use of AI for health questions is mainstream, clinically consequential, and largely unsupervised. The scale of adoption is documented across multiple independent sources. Of ChatGPT's approximately 800 million weekly users, roughly one in four, on the order of 200 million people, seek health-related information, a signal strong enough that OpenAI launched a dedicated consumer health product, ChatGPT Health, in January 2026. A KFF Tracking Poll on Health Information and Trust, fielded in early 2026, found that one in three US adults had turned to an AI chatbot for health information in the past year, a share equal to those who use social media for the same purpose and up sharply from a 17 percent monthly-use baseline KFF measured in June 2024.^96-98 A US cross-sectional survey found 21.5 percent of respondents had used ChatGPT specifically for health information, with 39.3 percent of those users doing so two to three times a week or more. Users skew younger, with a mean age of 32.8 versus 39.1 for non-users, have lower educational attainment on average, and make greater use of transient care settings such as emergency departments and urgent care.³

What patients do with these tools is clinically consequential. Among ChatGPT health users in one survey, 90.2 percent sought information about a health condition, 47.4 percent used it to decide whether a consultation was necessary, and 46.2 percent explored alternative treatments. Patients are not merely reading; they are acting. In the same study, 35.6 percent requested a referral based on ChatGPT information, 31 percent changed a medication, and 81 percent believed ChatGPT was as useful as or more useful than their doctor.¹² The KFF work adds the motivation: cost and access barriers drive a meaningful share of this behavior, and patients frequently do not close the loop with a clinician afterward.^96,99

The accuracy evidence is mixed and improving — a combination that supports neither dismissal nor uncritical adoption. A systematic review and meta-analysis of 60 studies found ChatGPT displayed an overall integrated accuracy of 56 percent in addressing medical queries.⁷ Physician-graded evaluations are more favorable: across 284 responses graded by 33 physicians spanning 17 specialties, ChatGPT achieved a median accuracy between "almost completely" and "completely correct," with GPT-4 outperforming GPT-3.5.⁹ In head-to-head comparisons of written responses to patient questions, chatbots frequently outscore physicians on quality and empathy,¹⁰ though these comparisons typically pit brief physician forum replies against much longer chatbot responses, and a systematic review of 20 such studies found results split, with 12 favoring the models, six mixed, and two favoring clinicians.¹⁷ Google's AMIE system, tested against 21 primary care physicians in a randomized blinded OSCE study, was non-inferior in management reasoning and scored better on treatment precision and guideline alignment. Contemporary models have reached diagnostic accuracy around 93.75 percent on clinical vignettes.²¹

The safety gap is where the operator's responsibility and opportunity both sit. A physician-led red-teaming study found problematic response rates ranging from 21.6 percent for one leading model to 43.2 percent for another, with unsafe response rates of 5 to 13 percent.²⁵ A structured stress test of ChatGPT Health triage recommendations found that 52 percent of gold-standard emergencies were undertriaged, with diabetic ketoacidosis, for example, directed to a 24-to-48-hour evaluation window rather than the emergency department, and the system proved susceptible to anchoring bias, shifting toward less urgent recommendations when a hypothetical family member minimized symptoms.² Patients cannot reliably catch these errors. In one study, specialists rated harmful ChatGPT responses significantly lower on usefulness and correctness, but patients showed no such discrimination.¹⁹ When tested with real human users, models that scored 94.9 percent on conditions in isolation led participants to identify the correct condition in fewer than 34.5 percent of cases, no better than a control group.¹ This knowledge-practice gap, in which models achieve 84 to 90 percent on knowledge exams but only 45 to 69 percent on practice-based assessments,³⁷ is the central unsolved problem of patient-facing deployment.

Trust dynamics complicate the picture further. A nationally representative US experiment found that disclosing AI involvement in a diagnosis consistently decreased patient trust and intention to seek help, with large effect sizes, a pattern consistent across age, gender, education, and political affiliation.³⁵ Patients still prefer human expertise, and the aversion narrows only among frequent AI users. The synthesis for operators is precise: patients are using clinical AI at mainstream scale, often driven by cost and access, frequently without clinician follow-up, and they cannot themselves distinguish safe advice from dangerous advice. This constitutes both a patient-safety exposure and a market opportunity. An AI front door operated by the health system, governed for safety and connected to real clinical capacity, closes the loop that consumer tools leave open.

Even patients who never open a chatbot now begin their journey to care on a screen, and the structure of that journey has changed how patients select a point of entry. The discovery layer, the set of steps between deciding to seek care and booking an appointment, has become a consumer process that resembles how people choose a restaurant or a contractor more than how they once chose a doctor.

The evidence from the booking marketplaces is consistent. Zocdoc's 2025 What Patients Want report found that a positive personal connection ranks as the single most important factor in choosing a doctor, above ratings and above location, which tells operators that the discovery experience is emotional and relational even when it is digital. Patients now view an average of 21 provider profiles before selecting one, and 92 percent read a provider's bio before booking, up from 76 percent in 2018.¹⁰⁰ Reviews have overtaken referrals: 61 percent of patients weight online reviews over personal recommendations,¹⁰² 35 percent have chosen a provider based on social media, and 25 percent began using voice assistants to research providers in 2025.¹⁰³ The funnel is digital from the first step, and the first step is increasingly a conversation with an AI rather than a search box.

The most operationally damaging finding concerns the phone. A quarter of patients say they hate calling a doctor's office. Roughly half report they could not reach their doctor by phone and delayed care as a result, and about a third abandoned a booking after failing to reach the office.¹⁰⁴ Every one of those is displaced demand, a patient who wanted care from the system and did not get it because the front door was a phone line. The access center, staffed and expensive, is actively shedding the demand it exists to capture.

Two facts anchor the interpretation. First, delivery remains overwhelmingly in person: 93 percent of non-mental-health appointments were booked for in-person visits in 2025, including 92 percent among Gen Z, and 92 percent of patients stayed in network amid cost concern.^100,101 Telehealth, after surging to 35.3 percent of primary care visits in the second quarter of 2020, stabilized at a modest baseline, around 6.6 percent of Medicare primary care visits in 2023 and 4 to 6 percent at large academic systems.^20,57,58 Second, the discovery and access layer around that in-person delivery has been thoroughly digitized and is now being intermediated by AI. The contest is not for the exam room, which remains theirs, but for the discovery and routing layer that decides which exam room the patient reaches, and whether it is inside the system at all.

The supply side of the front door has been reshaped by an expensive natural experiment. Over the past three years, two theories of how to own primary care collided, and the results are now clear enough to guide strategy.

The first theory was that retail real estate and foot traffic could be converted into primary care relationships. It failed on unit economics. Walmart closed all 51 of its Walmart Health centers and its virtual care business in 2024.^105,106 Walgreens closed roughly 160 VillageMD clinics, about half the footprint, after a write-down measured in billions on an investment of comparable size.¹⁰⁵ CVS closed dozens of MinuteClinics even as it pursued value-based care through its acquisitions of Oak Street Health and Signify Health.^105,107 The common cause was structural: retail labor costs, thin and uncertain reimbursement, and real-estate overhead ran against panels that filled slowly, and the retail-adjacency thesis, that a shopper buying groceries would convert into a primary care patient, did not hold at the margins required. Babylon Health, a digital-first entrant that showed lower acute hospital costs in some settings, nonetheless failed on unsustainable financing, a cautionary parallel for anyone assuming that a better care model guarantees a viable business.

The second theory scaled. Amazon acquired One Medical for 3.9 billion dollars, closing in 2023, and bundled membership into Prime at nine dollars a month or ninety-nine dollars a year, positioning a primary care membership as a consumer subscription rather than a clinical relationship.^108,111 The membership-plus-platform model then repositioned from competitor to partner: One Medical established arrangements with health systems including Cleveland Clinic, Hartford HealthCare, Virginia Mason Franciscan under CommonSpirit, Hackensack Meridian, Montefiore, and Rush University System, with systems citing access and shifting patient expectations as the durable rationale.¹⁰⁹ In January 2026, Amazon layered an AI health-care tool onto One Medical for members and added pay-per-visit and pediatric telehealth options.¹¹⁰ The move is best read as the platform annexing the AI front door: a company that already sits inside tens of millions of households through Prime is now placing an AI intelligence layer in front of, and alongside, the health systems it partners with.

The vendors now occupying the front door fall into recognizable categories, each with a distinct buyer, a distinct job to be done, and a distinct threat or opportunity for a health system. An operator who cannot name the categories cannot make deliberate own, partner, or cede decisions across them.

Ambient consumer chatbots are the default front door and the one systems control least. ChatGPT, Google Gemini, and Anthropic Claude reach patients directly, and patients bring their own AI to every clinical question. This is the surface behind the 200 million weekly health queries and the one-in-three adoption figure. No health system can own this layer; the realistic posture is to influence it through content, integration, and patient education rather than to compete with it head-on.

Platform and retail-adjacent care is dominated now by Amazon One Medical, with the wind-down of Walmart Health and the retrenchment of Walgreens and VillageMD as the counter-examples. The buyer here is the consumer, often through an employer or a Prime subscription, and the job to be done is convenient membership-based access. The strategic threat is disintermediation of the longitudinal relationship; the opportunity is partnership that extends a system's access without ceding its specialty and acute franchises.

Patient-facing conversational agent vendors build the infrastructure a health system can license to operate its own front door. Hippocratic AI is the category reference, having raised a 126 million dollar Series C at a 3.5 billion dollar valuation in November 2025 to expand patient-facing agents, with named health-system deployments and a deliberately non-diagnostic safety posture.^112,113 One health system rated a post-discharge engagement program built on these agents at 9.0 out of 10 by patients.^114,115 The buyer is the health system itself, the job is to staff the front door with governed AI labor, and the reimbursement logic runs through labor substitution and capacity expansion rather than fee-for-service billing.

Clinician-facing and EHR-native front-door features are where adoption is deepest and evidence strongest. A national AHA survey found 31.5 percent of US hospitals were using generative AI integrated with their EHR by 2024, with another 24.7 percent planning to within a year,⁴⁴ and among Epic-using hospitals, 62.6 percent had adopted ambient AI documentation by mid-2025.⁸⁴ Epic has emerged as the most influential single force in health-system AI adoption: hospitals on Epic were far likelier to be early adopters than those on other EHRs,^44,46 and Epic has integrated large language models for patient message drafting at Stanford, UC San Diego, University of Colorado, and NYU Langone. Microsoft and Nuance, through DAX Copilot, lead the ambient documentation market, with clinical evidence showing reductions in physician task load and burnout. Google's health AI, the most extensively documented in the medical literature, spans Med-PaLM 2, Gemini, and AMIE. For most operators, the EHR vendor is the platform on which a front-door strategy should be built before standalone tools are considered.

Discovery and booking marketplaces such as Zocdoc and Healthgrades own the consumer discovery layer described in Section 4, and their reviews-plus-booking partnerships increasingly bundle reputation and scheduling into a single surface. The buyer is often the individual practice or system paying for visibility; the job is demand generation; the strategic question is how much of the discovery relationship a system is willing to rent versus own.

Independent evaluation infrastructure is the operator's lens on all of the above. The Peterson Health Technology Institute and its use of ICER methodology provide an assessment framework for digital health technologies, and PHTI's AI Taskforce offers a what-works-and-what-is-worth-it discipline. This is a governance category rather than a vendor one, and it belongs in every procurement conversation as the counterweight to vendor-reported outcomes.^116-118

The central observation from this map: patients arrive through surfaces the system does not control, and several of the best-capitalized entrants are consumer technology companies for whom health care is one vertical among many. That is the strategic problem the rest of this paper works to solve.

The front door is a profit-and-loss decision and a risk-economics decision, and it should be modeled as one rather than adopted as a marketing channel. The market context is large and fast-moving. The global AI-in-health-care market is projected to exceed 187 billion dollars by 2030, growing at a compound annual rate near 37 percent from about 15 billion in 2022. The US telehealth market rose to 17.9 billion dollars in 2020 and is projected to reach 140.7 billion by 2030. The FDA had authorized roughly 1,000 AI-enabled medical devices by mid-2024, up from near zero a decade earlier. These figures describe momentum, not proof of return, and the operator's task is to separate the two.

The most cited savings estimate, from a Harvard and McKinsey analysis referenced by Wachter and Brynjolfsson in JAMA, projects that AI could save 5 to 10 percent of US health-care spending, on the order of 200 to 360 billion dollars a year in 2019 dollars, primarily by reducing administrative waste rather than by making diagnoses. Administrative complexity is the single largest source of waste, and administrative simplification is where AI is most clearly feasible. The same authors caution that AI could accelerate waste if payers and providers enter an arms race, with models streamlining denials on one side and appeals on the other. On the payer side, AI has increased the share of complex claims processed without denial from below 80 percent to above 90 percent and cut associated administrative spending by roughly 30 percent, though only about 21 percent of prior authorizations are currently automated, and the same tools have been implicated in rising Medicare Advantage denial rates, which is why prior authorization is now the most common target of state AI legislation.

For the front door specifically, the return should be built as a driver tree rather than reduced to a single number. The inputs are the visible costs of the status quo: access-center labor, phone abandonment, no-shows, referral leakage, and unmet demand that never converts to a visit. The levers are the mechanisms by which an AI front door acts on those inputs: labor substitution and augmentation in the access center, conversion of previously unmet demand into scheduled visits, in-network capture that reduces leakage, reduction of no-shows and avoidable emergency department use, and after-hours coverage without proportional staffing. The outcomes are cost-to-serve per contact, net new revenue from captured demand, and alignment with risk-bearing economics. Several nodes in that tree require benchmarks a system must source from its own operations, and the honest ROI model marks those nodes rather than papering over them with vendor averages.

The evidence on adjacent economics is encouraging but immature, and operators should read it with discipline. A large real-world study of 366,195 members found virtual-first care episodes cost 9 to 33 percent less than in-person-first care for common acute conditions.⁶⁵ Yet a rigorous 2026 review of 117 health-economic evaluations of AI found that 63 percent assessed tools at early development stages, only 28 percent included implementation costs, and only 57 percent reported operational costs, and an earlier review found that 98 percent of studies concluded AI was cost-effective or cost-saving, a result so uniform that it more plausibly reflects publication bias and incomplete cost accounting than settled truth. This is exactly why the ICER and PHTI evaluation logic belongs at the center of procurement: treat clinical benefit and economic impact as the primary domains, treat user experience, equity, privacy, and security as modifiers, stratify by patient-risk tier, and buy on independent evidence rather than on vendor-reported outcomes. The front door is a P&L decision, modelable node by node, and the discipline of the model is what separates a defensible investment from an expensive experiment.

The convergence of declining primary care access, mainstream patient AI use, and aggressive entry by technology companies poses a single question to every health-system operator: will the AI front door route patients into your system, or away from it? The framework below is built to answer it in practice rather than in principle.

Start from the disintermediation thesis as a risk, not a slogan. First contact is moving to surfaces the system does not control. That is a demand-funnel risk, a routing risk, and, over time, a relationship risk. Naming it as a strategic risk, with an owner and a budget, is the precondition for acting on it.

Assess where you stand with a front-door maturity model. Most systems can locate themselves on a five-level scale. The point of the model is to make the current state legible and the next step concrete, not to reach Level 4 everywhere at once.

Decide own, partner, or cede function by function. Not every function should be built in house, and not every function can be safely ceded. Triage safety, routing logic, and navigation into owned assets are functions a system should own, because they determine clinical safety and in-network capture. Agent infrastructure and discovery marketplaces are natural partnership candidates, where licensing a governed platform beats building one. Ambient consumer chatbots are the function a system will most often have to cede in the sense that it cannot control them, and the correct response is to influence them through content and integration rather than to pretend they can be owned.

Sequence the investment against the evidence. The strongest evidence in all of health-care AI concerns clinician-facing tools rather than patient-facing ones. A five-site JAMA study found ambient AI documentation was associated with 13.4 fewer minutes of total EHR time and 16.0 fewer minutes of documentation time per eight hours of scheduled care, plus 0.49 additional weekly visits and an estimated additional 167 dollars per month in E/M revenue per clinician. The first randomized trial of ambient scribes found significant reductions in time-in-note and burnout, and a six-system study found burnout fell from 51.9 percent to 38.8 percent after 30 days of use. The sequencing discipline follows directly: stabilize clinician capacity first, reaching high adoption of ambient documentation and AI-assisted inbox management, because capacity is the prerequisite for any front-door strategy, then build the patient-facing triage layer on top of a workforce that is no longer at the edge of burnout.

Own the integrated triage layer, because the outcome data favor integration. Integrated telemedicine, delivered within a patient's own system by their own clinicians, consistently produces better outcomes than direct-to-consumer alternatives: lower antibiotic prescribing, comparable safety, and higher follow-up rates.^66,70,71 Virtual visits with an outside physician are associated with three times the risk of a seven-day emergency department visit compared with visits to a patient's own physician.⁶⁸ A system that licenses or builds AI-powered symptom assessment and triage, and routes patients to the right level of care inside its own network, captures both the clinical benefit and the downstream volume. A system that cedes triage to a consumer app exports both.

An owned AI front door is a safety instrument or a liability, depending entirely on how it is governed. The risks are concrete and the governance tools now exist, which means the operator's obligation is execution rather than invention.

The safety gap is the first mandate. Because roughly half of patients who use consumer AI for health do not follow up with a clinician,⁹⁹ and because patients cannot themselves distinguish safe advice from dangerous advice, the value of an owned front door lies precisely in closing that loop: escalating appropriately, monitoring for the undertriage failure modes documented in ChatGPT Health,² and routing to real capacity. The non-diagnostic boundary that leading vendors maintain is not timidity; it is the current standard of care for patient-facing agents, and it should be a procurement requirement rather than a nice-to-have.

Privacy and data governance are the second. Patients are uploading test results and clinical notes to tools that are not HIPAA compliant, and there are insufficient legal guardrails around that data. A system that offers a governed alternative, with clear consent and data handling, competes on trust as much as on convenience. Neither ChatGPT Health nor other consumer models are HIPAA compliant, and operators should communicate this clearly to patients.³²

Liability runs in both directions. The liability framework now discussed in the literature assesses the source of an error, the opportunity to catch it, the potential for harm, and the potential for legal redress, and it should guide which functions a system automates and how much human oversight it retains.⁸³ As AI tools become established, failure to use them may itself come to be seen as a departure from the standard of care, which means the liability calculus covers the risk of not adopting alongside the risk of adopting.⁸³

Equity is the mandate most easily lost and most consequential. AI adoption correlates with hospital resources, urban location, and system membership,^44,46 which creates a digital divide that can widen disparities. Consumer AI health tools can legitimize dangerous self-rationing among people facing structural barriers to care,³⁴ and users of digital health tools skew younger, more educated, and more technologically literate.⁵⁹ Yet the same tools, well designed, can narrow gaps: an NHS mental health study found an AI chatbot increased referrals from minority populations. The determining variable is execution. A front door deployed only at well-resourced sites widens disparities; one deployed across the full network, including rural and safety-net sites, with outcomes monitored by race, ethnicity, and socioeconomic status, and with interfaces written at an appropriate reading level rather than the college level typical of current model output,^27,30 can narrow them.

Governance frameworks are available and should be adopted before scaling. The Joint Commission and the Coalition for Health AI have developed certification processes covering AI governance, privacy and transparency, data security, ongoing quality monitoring, safety-event reporting, and bias assessment. A concerning empirical finding is that hospitals with weaker evaluation of predictive AI were likelier to be early adopters of generative AI — an inversion of the appropriate sequence.⁴⁴ With state legislative activity intensifying, the operator's move is to establish transparency, consent, and monitoring frameworks proactively rather than in response to a mandate or an adverse event.

The front door of American health care is being rebuilt around AI, and the window to influence who owns it is measured in a few product cycles rather than a few strategic-planning cycles. The forces are structural and will not reverse: primary care access will remain constrained, patients will keep using AI at mainstream scale, and well-capitalized technology companies will keep building toward the entry point. The clinical evidence is unambiguous on two counts that operators must hold together. AI is improving fast enough that patients will keep using it, and it remains unsafe enough for unsupervised use that someone accountable must govern the first contact.

That someone should be the health system. The three moves that matter this quarter are within reach of any operator willing to sequence them correctly. Stabilize clinician capacity first, with ambient documentation and AI-assisted inbox management, because the evidence there is strong and the workforce crisis is the binding constraint. Build or license an integrated AI triage layer that routes patients within the network rather than away from it, because the outcome data favor integration and the alternative is exporting both the clinical benefit and the volume. Govern the whole of it for safety and equity before scaling, because an ungoverned front door is a liability and a governed one is a trust advantage.

A final consideration is trust. The front door is where confidence in a health system is established or lost, and for a growing share of patients it now takes the form of a conversation with a machine at a moment of clinical uncertainty. Systems that operate AI-mediated first contact as a governed clinical service — instrumented for safety and designed for equity — are positioned to retain the patient relationship and the demand that follows from it. Systems that cede first contact will retain their delivery assets while losing the beginning of the care episode, which is where both the relationship and the downstream volume originate.