Healthcare is being sold two stories at once. In the first, artificial intelligence and biopharma breakthroughs rescue a system squeezed by costs, staff shortages and collapsing margins. In the second, the same technologies are already leaking patient records, inventing anatomy and opening attack surfaces faster than anyone can govern them. Both stories are backed by data. This article puts them side by side: the dream, the reality and what it takes to get the first without paying for the second.
By Jasmine Angelique, Practitioner in Chinese Medicine (Swiss-certified) and naturopath
Published: 17 November 2026 · Last updated: 17 November 2026
I treat lawyers, founders and clinicians in burnout through my Chinese medicine practice and by telemedicine. Several times a month, one of them arrives with a chatbot's opinion already in hand. That is why I wanted to look past the headlines at the actual numbers.
The numbers at a glance
- 7%: projected annual growth of total U.S. healthcare EBITDA from 2025 to 2030, slowing to about 4% a year through 2027 before accelerating to about 9% a year from 2027 to 2030.
- $62B to $9B: payer EBITDA from 2023 to 2025, a fall of roughly 85%. Recovery to about $84.5B is projected by 2030.
- $241B to $350–369B: provider EBITDA from its 2022 low to 2025.
- 11.2% to 8.9%: industry EBITDA as a share of national health expenditure, 2019 to 2024.
- 0.7%: average 2025 operating margin for nonprofit health systems with more than $20B in revenue. For-profit systems of the same size posted 16%.
- 6.7% vs 3.7%: projected annual EBITDA growth for ambulatory surgery centers versus hospitals, 2025 to 2030.
- $990B: projected U.S. gross drug expenditure by 2029, growing about 8% a year.
- $200–360B: estimated annual net savings (5–10% of U.S. healthcare spending) from AI adoption with today's technology.
- 192.7 million: people affected by the 2024 Change Healthcare ransomware attack, the largest healthcare breach in U.S. history.
- 3,182: patients notified after a Johns Hopkins Bayview employee uploaded their data to an unapproved public AI tool in 2026.
- #1: misuse of AI chatbots, ECRI's top health technology hazard for 2026.
- 40%: share of U.S. healthcare professionals who have encountered unauthorized ("shadow") AI tools at work (Wolters Kluwer Health survey, 2025/2026).
- 72%: share of U.S. healthcare organizations where AI tools or agents are deployed without formal IT approval at least occasionally (Imprivata / Vanson Bourne survey, 2026).

the dream – economic ai opportunity
Part 1: The dream
An industry under sustained economic strain
Costs remain elevated, reimbursement is constrained and productivity growth in care delivery has lagged the broader economy for years. Industry-wide profitability has not returned to pre-pandemic levels. McKinsey estimates aggregate EBITDA grew only about 2% a year from 2019 to 2024, well below the roughly 6% growth in national health expenditure.
The payer side tells the starkest story. Total health plan EBITDA fell from about $62B in 2023 to about $33B in 2024 and reached a trough of about $9B in 2025, as medical-cost trends outpaced premiums and government rates. In 2025 the major risk-bearing segments collectively turned unprofitable. McKinsey had earlier flagged that Medicaid managed care EBITDA alone could fall to about negative $7B by 2028, as healthier members exit and rate adjustments lag enrollment losses by 18 to 24 months.
Providers recovered from a 2022 low of about $241B in EBITDA to roughly $350–369B in 2025, yet still sit below historical benchmarks. General acute care hospital EBITDA margins rose from 6.9% in 2023 to 8.9% in 2025 and are projected to hold near 8.8% through 2030. Scale helps only up to a point: nonprofit systems above $20B in revenue averaged a 0.7% operating margin in 2025, against 4.2% for systems with $10–20B and 3% for those below $5B.
The path to 2030 is uneven. Growth slows through 2027, driven by the expiration of Affordable Care Act subsidies and Medicaid eligibility changes, then accelerates. Workforce pressure compounds the problem: McKinsey expects a shortage of roughly 400,000 nurses by the end of 2030, and the U.S. could be short about 86,000 physicians by 2036. Clinician exhaustion is part of that equation, which I explore in why burnout is so common in healthcare.
Where the value is moving
- Ambulatory surgery centers are expected to grow EBITDA about 6.7% a year to 2030, versus 3.7% for hospitals.
- Specialty pharmacy and infusion services capture a rising share of pharmacy economics, while the segment is reshaped by transparency rules, Medicare price negotiations, most-favored-nation pricing initiatives, a revised 340B rebate approach and site-of-care shifts.
- Healthcare services and technology remains the fastest-growing segment, with about 8% annual revenue growth and 9% annual EBITDA growth projected and total EBITDA expected to pass $110B by 2029. Growth concentrates in software, platforms, data and analytics rather than labor-heavy administrative services.
- Drugs keep expanding: gross U.S. drug expenditure is projected to reach about $990B by 2029. Cell and gene therapy spending alone could reach an estimated $28B by 2030, about three times the 2021 level.
Without deliberate action (productivity gains, portfolio optimization, selective exits and operating redesign), margins are expected to stay depressed. McKinsey estimates that technology could save $150–300M in administrative costs and $380–970M in medical costs for every $10B in revenue, and that AI available today could generate $24–48B in annual run-rate administrative savings for hospitals and $10–30B for physician groups within five years.
Five avenues for reinvention
- Payers become orchestrators, not claims processors. A large share of payer work is claims-related and largely automatable. With longitudinal electronic records and AI able to evaluate care against policies and benefits, the claim becomes less essential as a billing artifact. The durable role is orchestration: integrating data and intelligence to enable risk-aligned funding, shared-risk arrangements across the care continuum (for example in oncology or chronic disease) and consumer engagement platforms that reach into daily health behaviors.
- Primary care built for continuous access. Direct primary care models augmented by clinical AI could let physicians manage panels of several thousand patients instead of a few hundred, because protocol-driven chronic and preventive care can largely be executed by AI agents under physician oversight. Recent policy changes let health-savings-account holders use pre-tax dollars for membership fees. Early pilots have shown high concordance between AI and physician decisions on chronic medication renewals. Layered with personalized agents, direct-to-consumer diagnostics and pharmacy access, the model targets both the primary-care shortage and the gap between health span and lifespan.
- Acute care that sweats its assets. Hospitals face Medicare volume growth, site-neutral payment and the need to be profitable at lower rates. The levers: migrate to ambulatory settings where returns are stronger, shorten length of stay to free capacity and raise utilization of existing fixed assets such as operating rooms rather than defaulting to new construction.
- New ways to pay for high-cost therapies. For behavior-driven conditions, consumer-directed bundles combine direct-pay medications (such as GLP-1s), wearables, coaching and lifestyle support. For complex, high-cost treatments, the options include outcomes-based or annuity-style payments, centers of excellence that share risk among clinicians, systems and manufacturers, and carve-outs with reinsurance or securities hedges.
- Services and technology firms that are AI-native. Measurable reductions in total cost of care or operating expense are now table stakes. Advantage lies in tools embedded in clinician or operational workflows, or that act as enabling platforms. Firms that fail to adopt AI at scale face displacement even as the sector grows.
My honest read on the dream models
The large-panel, AI-run primary-care model and the GLP-1-plus-wearables package both look efficient on paper and both worry me in practice. A physician overseeing several thousand patients with AI agents handling protocol-driven chronic care may work for stable, well-resourced individuals who already know how to advocate for themselves; it is less plausible for people whose main problems are complexity, trauma history, or fluctuating energy. GLP-1s can produce clear metabolic benefits, yet pairing them with continuous data streams still leaves the questions of sustainability, muscle preservation, emotional drivers of eating, and what happens when the medication stops. Before I would trust either approach for my patients I would need transparent long-term outcome data that include quality of life, retention, and adverse-event rates in real-world populations that resemble mine—not only selected trial or early-adopter cohorts—plus clear lines of accountability when the system errs and a demonstrated ability for a human clinician to override or exit the protocol without penalty.
Operating at three speeds
AI capabilities improve exponentially, core technology modernization proceeds linearly and organizational change moves slower still. Leaders need visibility into where AI actually delivers value, partnerships beyond the industry's traditional boundaries and architectures (such as a semantic layer) that add capability without replacing every legacy system. Healthcare employment is still expected to grow because demand keeps rising, so the transition is about reskilling rather than mass displacement.

the reality – shadow ai data risk
Part 2: The reality
Adoption has outrun governance. The same data flows that power the dream (longitudinal records, AI agents, vendor platforms and consumer apps) multiply the places where patient data can leak, be stolen or be misused. These are not hypothetical risks. Here is the documented record.
Case file 1: Johns Hopkins Bayview Medical Center (2026)
On 21 September 2026, an employee at Johns Hopkins Bayview in Baltimore uploaded a spreadsheet to a public AI tool the hospital had not approved. It held names, medical record numbers, admission dates, medications and treatment details for 3,182 patients seen between May and September. The incident was found during an internal review the same day. The employee deleted the prompt, thread and document from the tool's history and asked the developer to delete the data, yet public reporting has not confirmed whether the AI company removed it from its servers. Notification letters went out on 22 October. This was a "shadow AI" incident: human error, not an external attack.
Case file 2: Weill Cornell Medicine (2026)
Between 10 and 20 June 2026, during a radiology research study, patient information from Weill Cornell Imaging at NewYork-Presbyterian and NewYork-Presbyterian Hospital was entered into AI tools and processed with machine learning without authorization. The data included medical record and radiology accession numbers, dates and times of service, facility and provider names and clinical details such as diagnoses, treatment plans and procedural histories. It did not include names, addresses, Social Security numbers or financial information. Weill Cornell learned of the incident on 6 September and published its patient notice on 5 November.
Case file 3: Change Healthcare ransomware (2024)
Not an AI attack, but the clearest proof of what happens when one critical vendor fails. In February 2024 the ALPHV/BlackCat group entered Change Healthcare, a UnitedHealth Group subsidiary, through a Citrix remote-access portal that lacked multi-factor authentication. Attackers moved laterally for nine days, exfiltrated data (they claimed about 6TB) and deployed ransomware. Claims and payment processing were disrupted for around two months, putting heavy financial strain on providers nationwide. A ransom was reportedly paid and secondary extortion attempts followed. On 31 July 2025, Change Healthcare told the HHS Office for Civil Rights that about 192.7 million people were affected, close to 57% of the U.S. population.
The wider context: in its 2024 report to Congress, OCR received 74,299 reports of smaller breaches on top of the large ones and named incomplete risk analyses, excessive user privileges enabling lateral movement and weak authentication as recurring failures. Every AI tool, ambient scribe and agent adds more systems, credentials and data flows to protect.
Case file 4: Tempus AI and the Ambry Genetics data (2025–2026)
Tempus AI bought genetic testing firm Ambry Genetics for $600 million in February 2025, including a database holding the genetic information of hundreds of thousands of customers. Multiple class actions filed in federal court in Chicago allege that Tempus took that data without the written consent required by the Illinois Genetic Information Privacy Act, used it to train AI models and shared it with pharmaceutical partners. One report says plaintiffs allege sales to more than 70 pharma and life-science clients in deals worth about $1.1 billion. Tempus says it shares only de-identified data. The allegations are unproven, but the case shows how data collected for one purpose can travel into AI training and commercial licensing after an acquisition.
What I see in my own practice
In my own practice I see patients arrive with ChatGPT or similar advice several times a month, especially younger professionals and those who have already spent time and money on conventional routes. The tools are often reasonably good at listing common differential diagnoses or summarizing mainstream lifestyle advice for straightforward problems such as mild seasonal allergies or basic sleep hygiene. Where they fail is in nuance and safety. I have had patients stop or reduce prescribed medication on chatbot suggestions that ignored interactions, others who delayed care for progressive symptoms because the model ranked a benign explanation highest, and a few who arrived convinced they had rare conditions that did not match their presentation. In each case the next step was the same: I took a full history, examined them (including pulse and tongue), explained the mismatch, and we adjusted the plan. The chatbot had not seen the person; it had seen text.

research evidence – membership inference imaging
The research evidence: what has been proven in the lab
Membership inference: anonymization has a tail problem. In a Nature paper published on 24 June 2026 (Knolle et al., Technical University of Munich, Imperial College London and Hasso Plattner Institute), researchers ran one of the first patient-level privacy audits of medical AI. Across six data types, including chest X-rays, mammograms, retinal images, dermatology images, ECGs and EHR data, membership inference attacks achieved near-perfect success for some individual patients even when aggregate performance looked like random guessing. The number of highly exposed patients grew with model capacity. Underrepresented groups, by disease status, self-reported race, insurance, sex or imaging protocol, faced disproportionately high risk. The lesson: an average privacy score can hide the patients who are easiest to unmask.
Adversarial attacks: images that lie. Researchers have shown that small, often invisible changes to medical images can add or remove findings. The best-known example, CT-GAN (Ben-Gurion University, 2019), injected and removed signs of lung cancer in CT scans convincingly enough to fool radiologists. Later work has added or removed brain tumors on MRI using readily available generative tools, degraded thyroid ultrasound segmentation and used universal adversarial perturbations to trigger high misclassification rates in chest X-ray models. Diagnostic outputs can be manipulated without visual cues.
Chatbot reliability: ECRI's number-one hazard. For its 18th annual list, ECRI ranked misuse of AI chatbots as the top health technology hazard for 2026. Its experts found chatbots suggesting incorrect diagnoses, recommending unnecessary tests, promoting substandard medical supplies and inventing nonexistent anatomy, and reported advice that could cause physical harm, such as electrode placement that risked burns. Scale matters: more than 40 million people a day turn to ChatGPT for health information, according to OpenAI. ECRI warns that rising costs and facility closures may push even more patients toward chatbots as a substitute for care.
Prompt injection, model inversion and agent errors. Large language models and agents can be tricked into leaking data or overriding their instructions through prompt injection, data extraction and model-inversion techniques. Agentic systems raise the stakes: an autonomous agent could expose protected health information, write incorrect data into a record or change a medication or dose under a clinician's apparent authority before anyone notices, and one error can spread across records quickly.
What AI cannot see
What AI cannot capture is the relational and pattern-based layer that Chinese medicine treats as primary. Pulse quality changes with the conversation; tongue coating shifts with stress, diet, and sleep in ways that are hard to reduce to a static image; constitutional tendencies and the precise way a person holds emotion in the body do not appear in most training data. A patient who is a high-performing lawyer in burnout may show a wiry, taut pulse and a red-tipped tongue while reporting "just stress" and normal labs; the model sees the labs and the self-report and often recommends more productivity tools. That gap matters most for the burnout and high-demand professional groups I treat—lawyers, founders, clinicians—because their presentations are shaped by sustained emotional and cognitive load that algorithms trained on average cases routinely under-weight.
The same pattern-based reading applies well beyond burnout, for example in how I approach hip arthritis through Chinese medicine patterns, where two patients with the same scan can need very different treatment.
Shadow AI and the governance gap
- In a Wolters Kluwer Health survey of 518 U.S. healthcare professionals (conducted by CITE Research in December 2025, split evenly between providers and administrators), 40% had encountered unauthorized AI tools in their workplace. Nearly 20% admitted using one themselves and about 1 in 10 had used an unauthorized tool for a direct patient-care use case.
- In Imprivata's report The Agentic AI Trust Gap: Why Healthcare Needs Identity-Led Governance (September 2026), based on a Vanson Bourne survey of 250 U.S. healthcare leaders responsible for identity security or AI strategy across health systems, hospitals and integrated delivery networks, 72% said AI tools or agents are deployed without formal IT approval at least occasionally.
- Clinicians under documentation pressure have pasted notes and protected health information into public large language models.
- Only a small fraction of organizations express high confidence in their vendor-risk assessments.
- Vendors may claim HIPAA compliance and say they do not train on protected health information, yet policies, business-associate agreements and actual data flows frequently diverge. Advertising-supported clinical AI adds further commercial incentives around data and attention.
Further systemic risks include clinician deskilling through over-reliance, unclear liability when AI contributes to an error, bias amplified from training data and the difficulty of fully de-identifying records once a model has learned from them. Criminal ransomware groups remain the main source of large-scale theft, while commercial collection and secondary use of patient data form a parallel privacy risk driven by the value of the information.
Part 3: Where the dream meets reality
How strong is the evidence?
Established: large-scale breaches through vendors and weak authentication (Change Healthcare), shadow-AI uploads of patient data (Johns Hopkins Bayview, Weill Cornell) and chatbots producing confident but wrong medical answers (ECRI).
Emerging: per-patient membership inference risk in medical models (Nature, 2026), adversarial manipulation of imaging AI (demonstrated in research rather than documented in live clinical attacks) and alleged commercial reuse of genetic data for AI training (Tempus, allegations not yet proven in court).
Speculative: that AI-run panels of several thousand patients or GLP-1-plus-wearable packages will deliver durable outcomes across diverse, complex populations. The long-term real-world data does not exist yet.

integrative medicine discussion in modern european clinic
Why this matters outside the U.S.
The incidents above are American, but the risk is global. Patients and companies in the UK and the Netherlands operate under GDPR, Brazil under the LGPD and Singapore under the PDPA, while my Florida and Miami clients sit under HIPAA. Each regime treats health data differently, and AI tools rarely respect borders.
With patient data I am deliberately restrictive. In the clinic and in telemedicine I use AI only for non-identifying administrative tasks—drafting appointment reminders, organizing my own lecture notes, or generating generic handouts that I then edit. I never enter names, dates of birth, full case histories, pulse or tongue findings, laboratory results, or anything that could re-identify someone. Voice notes stay on devices I control; I do not feed ambient recordings into cloud models. The revised Swiss FADP and GDPR make secondary use and cross-border transfers more constrained than the average HIPAA business-associate arrangement I have seen described for U.S. vendors. Both European regimes treat health data as sensitive by default and require clear purpose limitation and, in practice, explicit consent or a strong legal basis. That legal environment reinforces a habit I already prefer: keep the raw clinical picture offline and local. The risk of a model retaining or later leaking details is simply not worth the convenience for the kinds of cases I see.
What is working
Targeted applications with strong controls have shown measurable time savings in published deployments: ambient documentation under approved agreements, protocol-driven chronic-care support with physician oversight and selected administrative automations. These gains are real. They do not cancel out the vulnerabilities above.
The minimum standard for safe healthcare AI
- Technical controls that detect and block unauthorized AI tools, not just policies that forbid them.
- Verified business-associate agreements and data minimization before any protected health information leaves the organization.
- Regular adversarial and privacy testing of models, including per-patient membership inference audits rather than averages.
- Multi-factor authentication and least-privilege access on every vendor and remote-access entry point.
- Named human accountability for every AI output that touches care.
- Continuous monitoring of agentic systems, with limits on what they can change without sign-off.
- A preference for architectures that avoid training on identifiable patient data.
- Clear consent and provenance records for any data used in training or licensing.
Regulation lags capability. HIPAA and related rules were not designed for generative or agentic AI.
What I built instead
Having a master's in communication sciences, I have used my knowledge to come up with a crypto coin: MyHealer, a telemedicine platform for natural and preventive medicine that I first built in 2022 and have just updated. I like technology. But vigilance is key when we handle patient data and I want to make sure that is respected. So MyHealer starts from the opposite assumption to most health software: an attacker will eventually get in, so whatever they copy must already be useless. Patients hold the keys to their own encrypted records and use separate unlinkable identities with doctors and insurers. Decoy records trigger automatic key-shredding the moment someone tries a bulk theft. Private by default, provable on demand.
One rule for lawyers, founders and clinicians
If a lawyer, founder, or clinician asked me for one practical rule, it would be this: never put identifying or clinically sensitive information into any AI system you do not fully control, and treat every chatbot answer as an unvalidated draft that still requires a human who has examined you. That single habit prevents the majority of the privacy and reliability problems I have already watched occur.
Accelerated evolution or compounded exposure
The conditions for reinvention are present: persistent financial pressure, abundant capital and talent and genuine breakthroughs in AI and biopharma. So is the evidence of harm: 192.7 million people exposed by one vendor failure, thousands of records pasted into unapproved chatbots and peer-reviewed proof that "anonymized" models can still identify the most vulnerable patients. Organizations that pair productivity gains with rigorous data controls, clinical oversight and transparent accountability will define the next era. Those that treat AI as an unconstrained accelerator of old workflows risk eroding trust and inviting tougher external intervention.
If you are running on empty and want a clinician who examines you rather than your text, book a discovery call.
Sources
- McKinsey & Company, What to expect in US healthcare and The future of healthcare
- HLTH, McKinsey Projects U.S. Healthcare EBITDA Growth as AI Reshapes Industry Economics (12 Nov 2026)
- HIT Consultant, McKinsey Projects U.S. Healthcare EBITDA to Reach $500B+ (12 Nov 2026) and HST report (16 Jan 2026)
- Becker's Hospital Review, Largest nonprofit health systems post thinnest margins: McKinsey
- McKinsey, How the healthcare industry can weather ongoing challenges and McKinsey on healthcare: Foreword
- Digital Health Insights, U.S. healthcare will face years of financial pain before a potential rebound
- WMAR-2 News, Johns Hopkins alerts 3,000+ patients after data sent to AI tool
- Healthcare IT News, Johns Hopkins shadow AI use linked to data breach and Improper AI use leads to another data incident
- The Baltimore Banner, Johns Hopkins data exposure caused by unauthorized AI tool
- Radiology Business, Patient info improperly used in radiology AI research
- Medical Daily, Weill Cornell says radiology patients' medical details were entered into AI tools
- HIPAA Journal, Change Healthcare responding to cyberattack
- CyberInsider, Change Healthcare revises breach impact to 192.7 million individuals
- Paubox, OCR reports 242 million records exposed in 2024
- Knolle, M.A., Menten, M.J., Jungmann, F. et al. Disparate privacy risks from medical AI. Nature (2026)
- Technical University of Munich, Study reveals privacy risks in medical AI
- Imperial College London, Hidden risks in medical AI leave some patients more exposed
- Mirsky, Y. et al. CT-GAN: Malicious Tampering of 3D Medical Imagery using Deep Learning. USENIX Security (2019)
- ECRI, Misuse of AI chatbots tops annual list of health technology hazards
- Wolters Kluwer Health, Wolters Kluwer survey finds broad presence of unsanctioned AI tools in hospitals and health systems (22 January 2026)
- Imprivata, New Imprivata research finds 72% of healthcare organizations have AI tools or agents deployed without IT approval (September 2026)
- GovInfoSecurity, Health AI firm faces lawsuits over DNA data use
- Robinson+Cole, Tempus AI faces class action cases
- Proskauer, Recent class actions expose data privacy risk in healthcare AI acquisitions