Here is a number the radiology AI industry would rather not talk about: 4.89.

That is the pooled odds ratio from a 2026 systematic review on automation bias in radiology - published by researchers at Rowan University School of Osteopathic Medicine and covering five controlled studies across mammography, chest radiography, and MRI. An odds ratio of 4.89 means that when an AI system flagged a finding incorrectly, radiologists were nearly five times more likely to miss the correct diagnosis than they would have been without AI assistance at all.

Meanwhile, the industry is celebrating a different number: 1,524. That is the count of FDA-cleared AI algorithms in medical imaging as of March 30, 2026. Seventy-six percent of those cleared algorithms - 1,163 of them - are in radiology. The FDA is clearing roughly 30 new algorithms per month. Sixty-eight were cleared in Q1 2026 alone. The pace is extraordinary. The outcome data tracking whether this wave of tools is improving diagnosis is not.

This issue examines the real tradeoff: what radiology AI delivers, what it costs, and why most programs are measuring the wrong thing entirely.

1. The Speed Revolution Is Real

Start with what is genuinely working. AI in radiology has produced measurable, meaningful improvements in workflow efficiency. The numbers are not small.

Average imaging turnaround time dropped from 11.2 days to 2.7 days in programs that deployed triage AI - an 76% reduction. Radiologist reading workload fell by 53% in high-volume settings when AI pre-screened studies for urgency. Scan reading speeds improved 30 to 75% depending on modality and AI system. These are operational wins that matter: patients wait less, critical findings surface faster, and radiology departments can handle volume growth without proportional staffing increases.

Aidoc's January 2026 clearance for a body CT AI covering 14 conditions achieved 97% sensitivity and 98% specificity across the cleared indication set. That is genuinely impressive performance. And 54% of US hospitals with 100 or more beds now report using AI in radiology. The infrastructure is in place.

The growth curve for FDA-cleared radiology AI tells a consistent story: from roughly 160 cleared algorithms in 2020 to 1,163 by early 2026. The field has scaled faster than almost any other area of medical AI.

The speed story is real. The problem is that it has become the only story being told.

2. The Accuracy Gap Nobody Is Measuring

Only 30% of radiologists currently use AI in clinical practice on a regular basis. That gap - between hospital systems that have deployed AI and the radiologists actually integrating it into reads - is a signal worth paying attention to.

The bigger gap is in measurement. Ask a radiology department how AI has affected their turnaround time and they can tell you precisely. Ask how AI has affected their false negative rate and most cannot answer. Ask whether their miss rate changes when AI confidence is low and the data does not exist.

The chart above captures the core problem. Programs tracking turnaround time and radiologist speed: over 80%. Programs tracking miss rate by AI confidence level: roughly 14%. Programs auditing AI versus unaided error rates over time: approximately 8%. The industry sold speed. The industry is measuring speed. Nobody is measuring what happens when speed and accuracy diverge.

3. Automation Bias: The Clinical Risk Hiding in Your Worklist

Automation bias is the tendency for humans to defer to automated systems even when those systems are wrong. In radiology, it means a radiologist accepting an AI finding without adequate independent review - or, more dangerously, missing a finding that AI did not flag because the absence of a flag substitutes for independent judgment.

The 2026 Rowan systematic review identified automation bias across every modality studied. The pooled odds ratio of 4.89 holds across different AI systems, different experience levels, and different imaging contexts. This is not a one-vendor problem or a mammography-specific problem. It is a structural problem with how humans interact with AI-assisted triage systems.

The most alarming finding involves experience level. For inexperienced radiologists - residents and early-career physicians - accuracy dropped from 79.7% to 19.8% when AI provided an incorrect recommendation. That is not a modest decline in performance. That is a fourfold collapse. The odds ratio for inexperienced radiologists accepting incorrect AI recommendations was 15.57, compared to 4.89 for the pooled sample.

4. The Expertise Paradox

Here is the uncomfortable irony: automation bias is lowest among the radiologists least likely to be using AI, and highest among the radiologists most dependent on it.

Experienced radiologists show meaningfully lower rates of automation bias in the literature. They are more likely to override incorrect AI flags, more likely to notice when an AI recommendation conflicts with clinical context, and more likely to perform the independent review that the AI workflow was supposed to supplement. Their baseline pattern recognition is strong enough to catch AI errors.

Trainees and early-career radiologists are in a different position. They entered training during the deployment wave. AI is woven into their workflow from day one. Their pattern recognition is still developing - which is precisely when AI assistance looks most helpful and is most dangerous. When the AI is wrong, they do not have the independent pattern recognition to catch it.

This creates a training pipeline problem that will compound over time. If the next generation of radiologists develops their diagnostic skills in an AI-first environment without structured oversight programs, the automation bias risk embedded in a 79.7% to 19.8% accuracy drop becomes a profession-wide liability rather than an individual one.

There is also a complementarity point worth taking seriously. Research published in Nature npj in 2025 found that AI and radiologists tend to make distinct categories of errors. AI systems miss findings that humans catch easily, and humans miss findings that AI catches reliably. This means thoughtfully designed human-AI collaboration can outperform either alone - but only when the collaboration is genuinely bidirectional, not when the human is simply ratifying the AI output.

5. Who Wins and Who Loses in the Current Model

The winner-loser split in 2026 radiology AI is not about which AI system you bought. It is about what you do with the data after deployment.

High-volume systems that have built audit programs - tracking miss rates by AI confidence score, comparing AI-assisted false negative rates to unaided historical baselines - are capturing the operational benefits of speed without creating the liability exposure that comes with unmonitored automation bias. These programs know when their AI is uncertain and have workflows to flag those cases for additional review.

Programs measuring only turnaround time and cost savings are flying blind. They know their throughput improved. They do not know whether their diagnostic accuracy changed. When a missed finding surfaces months later in a malpractice claim, the absence of outcome data is not a defense.

Trainees are in the highest-risk position in the current model. Their programs are deploying AI to help them handle volume, which makes sense operationally. Their supervisory oversight for AI-assisted reads may not be calibrated to the automation bias risk specific to inexperienced readers.

6. Deep Dive: What the 2026 Systematic Review Actually Found

The Rowan University study is worth examining in detail because it is the most rigorous synthesis of automation bias research in radiology to date.

The review identified five controlled studies that met inclusion criteria: randomized or crossover designs where radiologists read the same cases with and without AI assistance, with accuracy measured in both conditions. The studies spanned mammography screening, chest radiography, and MRI - three modalities with different visual complexity, different AI tool maturity levels, and different clinical stakes.

The pooled analysis found a consistent automation bias effect across all five studies. When AI provided an incorrect recommendation, radiologists were 4.89 times more likely to produce an incorrect read than when reading without AI. This held across modalities, suggesting the bias is not a mammography-specific phenomenon driven by the complexity of that particular task.

The experience stratification finding - OR of 15.57 for inexperienced radiologists versus a pooled 4.89 - was the sharpest finding in the paper. The authors note that this may partly reflect the fact that experienced radiologists have independent pattern recognition strong enough to catch AI errors, while inexperienced radiologists lack that baseline. The implication is that AI deployment without experience-stratified oversight protocols may transfer liability from high-experience to low-experience readers rather than reducing overall diagnostic risk.

One finding the authors highlight as a gap: none of the five studies tracked AI confidence scores against miss rates. This is the metric that would matter most operationally - knowing that when your AI outputs a confidence score below a certain threshold, human miss rates spike. That data does not exist in the published literature because programs are not collecting it.

What This Means For You

If you run a hospital system with AI in radiology: build the audit infrastructure before you need it defensively. Track miss rates by AI confidence score. Compare AI-assisted false negative rates to your historical unaided baseline. Stratify your oversight by radiologist experience. The operational case for AI is already won. The question now is whether you can capture the benefits without importing the risk.

If you are a radiologist integrating AI into your workflow: the Rowan data is a reminder that the AI recommendation is a second opinion, not a first one. The cases where AI is most confidently wrong are the cases where automation bias risk is highest. Interrogate the cases where AI flags something you would have dismissed, and the cases where AI is quiet about something that feels off.

If you are a radiology trainee or resident: you are in the highest-risk group in the current evidence base. Your program's AI deployment is not designed to be a substitute for developing independent diagnostic judgment. Use AI as a checker of your independent read, not as a triage system that tells you where to look.

If you are a radiology AI vendor: the speed pitch won. The infrastructure sale is done. The next competitive differentiator is outcome data. The vendors who can show diagnostic accuracy improvement - not just workflow efficiency - will own the next phase of this market. The vendors who cannot will face growing scrutiny as the liability exposure of unmeasured automation bias becomes clearer.

If you are a payer or health system administrator: the 30% clinical adoption rate among radiologists despite 54% hospital deployment tells you something. The gap between procurement and actual use is often a signal that the tool is not integrated into the workflow in a way clinicians trust. That trust problem is worth solving before it becomes an outcome problem.

The 1,524 cleared algorithms tell you how fast this field is moving. The 4.89 odds ratio tells you what the industry has not figured out yet. Both numbers are real. The question for radiology over the next three years is which one drives policy.

Speed is the feature. Accuracy is the product. The programs that understand the difference will be positioned very differently from the programs that do not when the accountability reckoning comes.

About the Author

Jonathan Govette is the CEO of Oatmeal Health, a company focused on early lung cancer detection in underserved communities. This newsletter covers healthcare policy, imaging AI, and the structural forces shaping community health. Subscribe for weekly analysis at the intersection of technology and access.

Key References

• Rowan University School of Osteopathic Medicine (2026). Automation Bias in AI-Assisted Radiology: Systematic Review and Meta-Analysis. 5 studies, pooled OR=4.89.

• FDA Medical Device Database. AI/ML-Enabled Medical Devices: 1,524 cleared algorithms as of March 30, 2026. 76.31% in radiology.

• Aidoc Medical (January 2026). FDA Clearance for Body CT AI Covering 14 Conditions: 97% sensitivity, 98% specificity.

• Nature npj Digital Medicine (2025). Complementary Error Patterns in AI and Human Radiology Reads: Implications for Collaborative Workflow Design.

• EJRAI (2026). Workflow Integration of Radiology AI: 30-75% Scan Time Reduction Across 12 Institutions.

• American College of Radiology (2026). AI Adoption Survey: 54% of US Hospitals with 100+ Beds Report Radiology AI Use; 30% of Radiologists Report Daily Clinical Use.

This is not primarily a story about insurance markets. It is a story about what happens to federally qualified health centers - the safety-net institutions built specifically to absorb coverage shocks - when the shock arrives faster, harder, and with fewer federal resources than at any point in the last two decades.

What Just Happened in New York

New York's Essential Plan was, until June 30, 2026, a model for how states could use ACA Section 1332 waiver authority to extend subsidized coverage to low-income residents. The Essential Plan provided $0-premium, $0-deductible coverage to residents earning between 138% and 250% of the federal poverty level - a population too income-high for Medicaid but historically priced out of the commercial market.

H.R. 1 eliminated federal matching payments for Essential Plan enrollees above 200% FPL effective July 1, 2026. New York faced a binary choice: absorb the cost entirely from state revenue, or end coverage for that income band. The state could not fill the gap in time. As a result, approximately 450,000 people - those earning between 200% and 250% FPL - lost their coverage on the first day of July.

The 1.3 million lower-income Essential Plan enrollees (below 200% FPL) retain their coverage under federal Medicaid matching. But the 450,000 who lost it face a replacement market with average premiums of $1,904 per month - up 114% from the $888 average in 2025, driven by the elimination of enhanced premium tax credits that H.R. 1 allowed to expire.

A family of three at 220% FPL - earning about $54,000/year - now faces monthly insurance premiums that exceed their rent. The math does not work. They will join the uninsured.

The National Picture Is Worse

New York's Essential Plan loss is dramatic because it happened all at once. But it is part of a broader, slower-moving collapse in coverage that has been building since 2025.

ACA exchange enrollment nationally fell from 22.1 million in 2025 to 19.2 million in 2026 - a decline of nearly 3 million people. The Congressional Budget Office projects that the reconciliation law's Medicaid cuts and premium credit rollback will produce 11.8 million more uninsured Americans by 2034. The Urban Institute estimates up to 7 million lose coverage by 2028 from work requirements alone.

The next major wave arrives January 1, 2027, when the Medicaid work requirement takes effect for expansion adults aged 19-64. CMS's own Interim Final Rule projects 2.3 million lose Medicaid in the first year, rising to 3.1 to 3.3 million in subsequent years. The CMS estimates are based on enrollment declines from administrative churn - people who are technically eligible but fail to document 80 hours of monthly qualifying activity. Of those projected to lose Medicaid, NACHC estimates 5.6 million are current health center patients.

FQHCs: The Institution Built for This Moment

Federally Qualified Health Centers exist precisely to serve patients regardless of ability to pay. They receive Section 330 federal grant funding, Medicare and Medicaid cost-based reimbursement, and are required by law to provide a sliding-scale fee schedule for uninsured patients. When coverage collapses, FQHCs are, by design, the place people go.

That design has a cost. And right now, that cost is accelerating past what many centers can sustain.

The numbers tell the story clearly. Health center net program margins averaged 1.6% in 2023 - thin but positive. By 2024, the sector had moved to a negative 2.1% margin, a 3.7 percentage point deterioration in a single year. The program posted a 2% loss in 2025. These are not individual centers struggling - this is a structural financial shift across the sector.

The proximate driver is the Medicaid unwinding that began in 2023 and has continued through 2025. The National Association of Community Health Centers reports that the average health center lost $595,000 in annual revenue from Medicaid unwinding alone - with large urban centers losing up to $20 million. The revenue loss is structural: patients who lose Medicaid don't disappear from the appointment schedule. They show up, receive care, and pay what they can on the sliding-scale fee - typically far less than the cost of service.

The $7 Billion Uncompensated Care Question

NACHC's projection is stark: $7 billion per year in additional uncompensated care costs for health centers if the coverage losses in H.R. 1 materialize as projected. To put that in context, the entire Section 330 federal grant program appropriation is approximately $7.4 billion annually.

This means the coverage loss projected by CBO and the Urban Institute would create an uncompensated care burden approximately equal to FQHCs' entire federal operating grant. The math is unfeasible without either new federal investment, significant state subsidy, or service contraction.

Service contraction at health centers is not a hypothetical. It means reduced hours at rural sites. It means longer wait times at urban centers. It means behavioral health and dental services - historically the first to be cut - disappearing from the care model. It means the 31 million Americans who currently receive primary care at health centers finding that care harder to access, in the same moment that their insurance coverage has been stripped away.

The Work Requirement Timeline: What Comes Next

The immediate crisis - NY Essential Plan - is severe. But the structural pressure builds further in 2027. Here is the timeline that health center CFOs and policy teams are working against:

• July 1, 2026: NY Essential Plan cut. 450,000 lose coverage immediately. Health centers in New York see uninsured volume rise.

• October 2026: State Medicaid redetermination cycles accelerate under H.R. 1 requirements. Additional churn expected.

• January 1, 2027: Federal Medicaid work requirement takes effect. 80 hours/month qualifying activity required for expansion adults 19-64. CMS projects 2.3M lose coverage in Year 1.

• 2027-2028: Urban Institute projects up to 7 million total coverage losses from work requirements. NACHC: 5.6 million of these are current health center patients.

• By 2034: CBO projects 11.8 million more uninsured Americans from the combined effects of the reconciliation law.

Each milestone represents not just an insurance statistic but an operational event for health centers: more patients, more uncompensated care, more financial pressure, and - absent policy intervention - more service reduction.

Winners and Losers

Not everyone loses in this scenario. State budget offices that had been absorbing Medicaid expansion costs see reduced outlays, at least in the near term. Commercial insurers face a smaller Medicaid-enrolled population competing for exchange market share. The federal government reduces direct Medicaid outlays - which is the stated fiscal rationale for H.R. 1's coverage provisions.

The losers are more diffuse but more numerous: FQHCs and community health centers facing $7 billion in projected uncompensated care. Uninsured patients earning 200-250% FPL who now face $1,904/month premium options or go without. Rural health systems with no safety-net alternative when their local FQHCs contract. Emergency departments that will absorb the uninsured surge that community health centers can no longer buffer.

What Health Centers Are Doing Right Now

The health centers that are navigating this best are doing several things simultaneously. First, they are accelerating grant diversification - pursuing state innovation grants, HRSA pilot programs, and foundation funding to replace eroding Medicaid revenue. Second, they are implementing sliding-scale fee optimization, ensuring that uninsured patients are being enrolled at the highest fee tier the data supports rather than defaulting to the lowest. Third, they are investing in eligibility assistance staff - workers who can help patients navigate remaining coverage options, including Medicaid for those below 200% FPL and potential future state coverage programs.

Some centers are also engaging aggressively with state policy processes. New York, for example, has signaled interest in creating a state-funded replacement for the lost Essential Plan coverage, though no funded program has been announced. Health center advocacy organizations are pressing for emergency supplemental Section 330 grant funding to offset uncompensated care increases.

The political reality is that federal supplemental funding is unlikely in the current legislative environment. Centers that are waiting for a federal rescue are planning poorly. The centers that survive and continue to serve their communities will be those that are restructuring their financial models now - before the 2027 work requirement wave makes the current moment look manageable by comparison.

The Data Health Executives Should Be Tracking

For health system leaders and policymakers monitoring this situation, here are the metrics that matter most in the next 12 months:

• Uninsured patient visit volume at FQHCs, tracked monthly. A sustained increase above 10% over 2024 baseline signals the financial inflection point is arriving.

• Section 330 grant adequacy ratio: uncompensated care cost as a percentage of 330 grant revenue. Any ratio above 80% indicates a center operating at structural risk.

• State Medicaid enrollment trends following H.R. 1 redetermination requirements. Month-over-month enrollment decline will indicate the pace of the coverage loss wave.

• Emergency department utilization rates at hospitals in FQHC catchment areas. Rising ED visits for primary care-treatable conditions signal that the FQHC buffer is failing.

• HRSA emergency grant applications and approval rates. A surge in emergency funding requests from centers will precede a wave of site closures or service reductions.

The Bottom Line

The 450,000 New Yorkers who lost coverage on July 1 are the first clear signal that the coverage architecture of the ACA era is being substantially dismantled. They are not the last. The January 2027 work requirement wave will be larger. The 2034 CBO projection will be reached through a series of smaller shocks, each one manageable-seeming in isolation, devastating in aggregate.

FQHCs were built to be the last line of health care access in America. That role is now being stress-tested at a scale and pace the system was not designed to handle. The centers that understand this, and are restructuring accordingly, have a chance to fulfill that mission through the next several years of policy turbulence. The ones waiting for the federal safety net to stabilize first may not survive long enough to find out what stabilization looks like.

The wave is not coming. It is here.

Sources: NACHC Federal Policy Report, June 2026; CBO H.R. 1 Coverage Effects Analysis, May 2026; Urban Institute Medicaid Work Requirements Impact Model, April 2026; CMS Medicaid Work Requirement IFR, Q2 2026; KFF Health Insurance Marketplace Report, 2026; New York Department of Health Essential Plan enrollment data; HRSA Health Center Program data, FY2024.

AI is reading chest X-rays faster than ever before.

Aidoc's January 2026 FDA clearance covers 14 conditions at 97% mean sensitivity and 98% mean specificity. AI triage tools are prioritizing worklists in real time, flagging critical findings in seconds, and promising to close the widening gap between imaging volume and radiologist capacity. The vendor dashboards look impressive. The benchmark numbers look convincing.

But in March 2026, the patient safety organization ECRI published its annual watch list and named "Navigating the AI diagnostic dilemma" as healthcare's single biggest patient safety risk of the year. Not data breaches. Not medication errors. AI diagnosis.

That is not a coincidence. That is a warning the field has not fully absorbed.

1. The Automation Bias Problem Nobody Wants to Name

Automation bias is not a theoretical risk. It has been documented across aviation, nuclear operations, and now, repeatedly, in clinical radiology.

The concept is straightforward: when a human works alongside an automated system, they unconsciously defer to the machine's output, especially under cognitive load. When the AI clears a scan, the radiologist reviews it less carefully. When the AI flags something urgent, they scrutinize it more. The AI shapes attention - even when the AI is wrong.

A peer-reviewed study published in Radiology found that incorrect advice from a purported AI decision support system measurably impaired radiologist performance across experience levels. The more trusted the AI appeared, the larger the impairment.

A separate study on cerebral aneurysm detection via time-of-flight MRA confirmed the same pattern. Radiologists working with AI assistance showed automation bias in a high-sensitivity imaging context. The AI's confidence score shaped where radiologists looked - and what they missed.

This is not a technology failure. It is a human-machine interaction problem that no accuracy benchmark on a vendor slide captures.

The fundamental issue with how most radiology AI is deployed today: vendors measure sensitivity and specificity on curated benchmark datasets. They do not measure how radiologist behavior changes after the AI makes a call. That behavioral shift is where diagnostic risk accumulates in production.

When an AI system marks 80% of scans as low priority, what happens to radiologist attention on case 73 in a 90-case worklist? That question does not appear in any sensitivity curve. But the answer matters more than the benchmark.

2. What ECRI's Warning Actually Means

ECRI's 2026 report is not speculative. It is grounded in documented failure patterns across actual clinical deployments.

Some machine learning models tested in simulated diagnostic scenarios failed to recognize 66% of critical or deteriorating health conditions. Not rare edge cases. Common presentations that a competent clinician would flag immediately.

When AI misses a critical finding, it does not return an error message. It returns a low-priority designation. A confident-sounding negative. That is the mechanism through which automation bias causes patient harm: not a software crash, but a subtle signal that reassures when it should not.

The specific failure modes ECRI identified include rare diseases and uncommon presentations, subtle early-stage findings in complex anatomies, and cases where imaging alone cannot tell the full story without clinical context. These are precisely the cases where a careful radiologist would ordinarily spend more time. When AI assigns them a low-priority flag, that extra time evaporates.

Radiology is also facing a structural workforce crisis that makes this worse. Practice turnover among radiologists increased 61% between 2013 and 2022, and radiologists are leaving practice entirely at more than twice the rate of a decade ago. Burnout affects 46% of private practice radiologists and 37.4% in academic settings. A burned-out radiologist reviewing AI-flagged worklists at the end of a high-volume shift is the exact profile most susceptible to automation bias.

3. The Accountability Gap at the Heart of AI Radiology

One of the starkest blind spots in how health systems are deploying radiology AI: almost no organizations are auditing AI-assisted reads against gold-standard re-reads at scale.

Vendors report benchmark accuracy. Health systems report turnaround time reduction. What almost no one is reporting is the false negative rate stratified by AI confidence score - the one metric that would directly reveal automation bias in their own deployment.

A 2025 prospective multicenter study on AI-assisted intracerebral hemorrhage detection found that AI-assisted radiologists achieved 98.91% sensitivity compared to 95.91% for AI alone. The critical finding: all 45 cases missed by AI were caught by radiologists, and all 12 cases missed by radiologists were caught by AI. Human and AI failures are complementary but do not overlap. An AI clearance that reduces radiologist scrutiny on a case the AI missed is exactly where harm concentrates.

The accountability gap has three dimensions most health systems are not addressing. First, liability: most radiology AI vendor contracts place responsibility for diagnostic errors on the signing radiologist, not the AI tool. Second, governance: CMS and the Joint Commission have not issued formal guidance on AI-assisted diagnostic reads, leaving most health systems writing their own policies with wildly varying quality. Third, operations: worklist design that routes radiologists to AI-flagged cases first and AI-cleared cases last means fatigue accumulates exactly where caution is most needed.

4. What the Data Actually Shows

The picture from peer-reviewed research in 2025 and 2026 is more nuanced than the vendor pitch or the automation bias critique alone would suggest.

In well-defined, high-volume applications, AI-assisted accuracy data is genuinely impressive. Breast cancer AI flagged 49.8% of interval cancers missed by human readers. AI decreased false positives by 37.3% and reduced unnecessary biopsies by 27.8% in breast imaging programs. For rib fracture detection, AI achieved a median inference time of 10.6 seconds versus 3.3 hours for human reports, with no compromise in accuracy on that specific task.

The challenge is that radiology is not a single well-defined task. It is hundreds of different diagnostic problems distributed across imaging modalities, patient populations, and disease stages. An AI that achieves 98% sensitivity on PE detection in a controlled trial may perform meaningfully differently when deployed across the heterogeneous scan volume of a community hospital.

A 2026 systematic review of AI generalizability found that performance metrics from controlled trials frequently do not replicate when the same tools are deployed in different clinical settings, patient populations, or imaging acquisition protocols. The gap between benchmark and real-world performance is not a bug. It is a structural feature of how AI is trained and how clinical settings vary.

Speed improvements are real. Accuracy improvements are conditional. The conditions are site-specific and modality-specific. That complexity is not reflected in how most health systems make AI procurement decisions today.

5. The Missing Metric: AI Confidence vs. Miss Rate

Here is the specific audit that most radiology AI deployments are not running, and should be.

Track the false negative rate stratified by AI confidence score. For cases where the AI assigned a low-priority or negative classification, what percentage contained a significant finding that was later identified? And how does that rate change as radiologist case volume increases across a shift?

This metric directly reveals automation bias in a live deployment. It is not complicated to generate. It requires structured re-reads of a sample of AI-cleared cases combined with outcome tracking. The information exists in most health systems' PACS and RIS systems. It is simply not being extracted and analyzed.

The most forward-thinking radiology programs in 2026 are running exactly this kind of audit. They are tracking miss rates by AI confidence bucket, measuring whether low-priority cases generate proportionally more subsequent imaging orders or escalations than the AI flag suggested, and using that data to calibrate how aggressively the AI's low-priority designations influence worklist positioning.

The math is not close. A quarterly audit of 200 randomly sampled AI-cleared cases costs roughly $10,000 to $40,000 in re-read labor at current rates. The avoided liability exposure from a single missed lung cancer diagnosis in most U.S. jurisdictions exceeds $1 million. The audit program pays for itself on the first miss it prevents.

6. The Regulatory Landscape Has Not Caught Up

The FDA has cleared more than 1,524 AI-enabled radiology devices as of mid-2026. The pace of clearances is accelerating: 68 new algorithms in the first quarter of 2026 alone. Breakthrough Device Designations are now being granted to generative AI systems that draft entire radiology reports, not just flag findings.

The governance infrastructure has not kept pace.

Current FDA clearance processes evaluate algorithm performance on training and test datasets. They do not require pre-market demonstration of real-world performance in heterogeneous clinical settings, post-deployment monitoring of false negative rates stratified by AI confidence, or vendor notification when site-specific performance drifts below benchmark thresholds.

The EU AI Act, fully effective in 2026, requires that high-risk AI systems including AI in medical diagnosis document training data curation, bias checks, and human oversight policies. European health systems deploying radiology AI are building those governance frameworks now. Most U.S. health systems are not.

The liability landscape is also evolving. Malpractice attorneys are actively building cases around the radiologist-as-rubber-stamp scenario: a radiologist who signed an AI-generated read without meaningful independent review on a case where the AI missed a significant finding. This is not hypothetical. The case law is beginning to accumulate.

Deep Dive: The Real Cost of an Unaudited Deployment

Let me walk through the math of what an unaudited AI radiology deployment actually costs when automation bias generates a missed finding.

In the best-case scenario, an AI system performing at 97% sensitivity on its benchmark dataset misses 3% of positive cases. In a high-volume setting reading 10,000 scans per month, that is 300 cases per month where a positive finding was not flagged by the AI. If automation bias causes radiologists to review AI-cleared cases with 20% less scrutiny - a conservative estimate from the automation bias literature - the proportion of those 300 cases that also get missed by the radiologist increases meaningfully.

Scenario A: High-Volume Site, No Audit Program

Reading 10,000 chest CTs per month. AI benchmark sensitivity: 97%. Real-world sensitivity estimate given distribution shift: 94%. Cases with positive findings: assume 15% of volume, or 1,500 cases. AI misses 6% of those: 90 cases per month with positive findings cleared as low priority. Automation bias increases miss rate on those cases by an estimated 20% relative to normal scrutiny. Result: roughly 18 additional missed findings per month introduced by unmonitored AI workflows. Annualized: over 200 missed findings per year that a properly governed deployment would catch.

Scenario B: Same Site, With a Quarterly Audit Program

Same scan volume. Same AI. But quarterly audit of 200 randomly sampled AI-cleared cases against a radiologist re-read. Miss rates are measured, reported, and used to adjust worklist protocols. Radiologists are briefed on their AI-assisted miss rate trends each quarter. Result: automation bias is detected and quantified within 90 days. Protocol adjustment alone cuts miss rates by an estimated 40-60%. Additional cost: roughly 800 re-reads per year. Avoided liability exposure from a single missed lung cancer diagnosis: well above $1 million in most U.S. jurisdictions.

The math is not close. The audit program pays for itself on the first miss it prevents.

What This Means For You

For FQHC executives and community health center leaders: if you are evaluating or using AI-assisted diagnostic imaging tools, ask every vendor for their false negative rate on cases they assigned low priority. If they cannot produce that number, that is your answer.

For health system administrators and CMOs: build the audit layer before expanding AI deployment. Measuring turnaround time without measuring miss rates is measuring the wrong thing. Implement structured re-read protocols for a random sample of AI-cleared low-priority cases quarterly.

For radiologists and radiology department chiefs: the legal exposure is yours. Until contract terms shift accountability to AI vendors, your defense against a missed finding claim depends on documentation that you conducted a meaningful independent review. Know your institution's protocol for AI-assisted reads.

For healthcare investors and founders: governance and audit infrastructure is the competitive moat in radiology AI that nobody has built yet. The first platform that provides site-specific AI confidence versus miss rate dashboards will be the one health systems lock into long-term.

For policy advocates: push CMS and the FDA to require post-deployment performance monitoring as a condition of continued AI clearance in radiology. The EU model is available to adapt.

Closing

The radiology AI story of 2026 is not one of failure. It is one of an industry moving faster than its own accountability infrastructure can follow.

AI tools that triage worklists, flag critical findings, and assist with structured reporting are genuinely valuable. The data on their performance in well-matched applications is real. The case for deploying them is not in question.

What is in question is whether health systems are building the oversight layer that makes AI-assisted radiology safer than unassisted radiology, not just faster.

ECRI did not name AI diagnosis the No. 1 patient safety risk of 2026 because AI does not work. They named it because AI works well enough that clinicians stop checking.

Speed without oversight is not an upgrade. It is a liability dressed as a solution.

The question worth asking in every radiology department reading this: do you know your AI-assisted miss rate on cases the algorithm cleared? If not, that is the most important metric your team is not yet measuring.

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise.

Subscribe at oatmealhealthjonathangovette.substack.com

Key References

ECRI 2026 Patient Safety Watch: "Navigating the AI Diagnostic Dilemma" - ranked No. 1 patient safety concern for 2026.

Radiology (RSNA): Automation bias study on AI-assisted mammography - incorrect AI suggestions impaired radiologist performance across experience levels.

Neiman Health Policy Institute: Radiologist practice turnover increased 61% between 2013 and 2022; radiologists leaving practice at twice the rate of a decade ago.

Prospective multicenter ICH study 2025: AI-assisted radiologists at 98.91% sensitivity vs. 95.91% for AI alone; all 45 AI misses were caught by radiologists.

ECRI 2026: Some ML models failed to recognize 66% of critical conditions in simulated cases.

On July 2, 2026, CMS released its Outpatient Prospective Payment System proposed rule for calendar year 2027. Buried inside 1,400 pages of regulatory language was a single payment rate change that will hit America's most fragile hospitals harder than any single Medicare policy in a decade.

The agency proposed cutting 340B drug payments from ASP+6% to ASP-33.4%. That is a 39.4 percentage point drop in reimbursement. The total estimated impact: $4.85 billion in reduced revenue for the hospitals, federally qualified health centers, and rural health clinics that serve the country's most vulnerable patients.

This is not a rounding error. This is a structural defunding of the safety net.

What CMS Just Did

The 340B Drug Pricing Program, created by Congress in 1992, allows covered entities to purchase outpatient drugs from manufacturers at deeply discounted prices. The program exists because safety-net providers carry a disproportionate share of uncompensated care, treat higher rates of complex and chronically ill patients, and operate with margins that leave almost no room for financial shocks.

For decades, covered entities purchased drugs at 340B prices and billed Medicare at the standard ASP+6% rate. The spread between those two numbers funded preventive services, care coordination programs, charity care, and specialty services that could not otherwise exist at these institutions.

In 2018, CMS first cut the 340B rate to ASP-22.5%, citing overpayment concerns. Courts ultimately ruled that cut invalid. In the years since, CMS restored the rate and accumulated an estimated $9 billion in obligations to covered entities for the years the cut was in effect. The new proposed rule addresses that recoupment while simultaneously cutting the ongoing 340B rate to its lowest level in the program's history.

The Math Behind the Shock

Under the current rate of ASP+6%, a hospital that purchases a drug at 340B discount prices and bills Medicare gets reimbursed at roughly 6% above the Average Sales Price. Under the proposed ASP-33.4% rate, that same drug bills at 33.4% below ASP. On a drug with a $200 ASP, the difference is $78 per unit. Multiplied across millions of drug administrations per year at thousands of covered entities, that produces the $4.85 billion annual revenue loss estimate.

CMS argues this is justified by two things: first, that the existing 340B rate overcompensated covered entities relative to their actual acquisition costs; and second, that the $2.3 billion in annual budget-neutral redistribution to other outpatient services will benefit those same hospitals through a higher base payment rate.

That second argument does not survive contact with the data. The 2.4% net outpatient base rate increase flows to all hospitals proportionally. Safety-net hospitals, which serve lower-income patients with fewer commercially insured visits, gain less from a base rate increase than from the 340B spread they are losing. The math does not work for hospitals that run on the thinnest margins.

Why This Hits Safety-Net Hospitals So Hard

Not all hospitals participate in 340B. To qualify, a hospital must meet the definition of a disproportionate share hospital (DSH) or operate as a critical access hospital, children's hospital, or freestanding cancer center, among other categories. These are precisely the institutions that serve high percentages of Medicaid, uninsured, and complex-need patients.

The concentration of the cut matters. CMS's own estimate acknowledges that the $4.85 billion reduction falls almost entirely on covered entities. Non-340B hospitals are largely unaffected and actually benefit from the budget-neutral redistribution. This is, at its core, a transfer of Medicare payments from hospitals that serve the poor to hospitals that do not.

Ashley Thompson, Senior Vice President of Public Policy at the American Hospital Association, put it plainly: 'This enormous cut will make drugs less affordable for America's most vulnerable patients and threaten funding for programs these patients depend on.'

The Budget Neutrality Shell Game

CMS frames its 340B cuts as part of a budget-neutral adjustment: the $4.85 billion in 340B reductions is offset by a higher base outpatient rate that benefits all hospitals. This framing is technically accurate and substantively misleading.

Budget neutrality means total Medicare outpatient spending stays approximately flat. It does not mean the impact is distributed evenly. A hospital that relies heavily on 340B drug revenue loses far more than it gains from the base rate increase. A hospital with minimal 340B exposure gains from the base rate increase without taking any cut.

The recoupment acceleration compounds this. CMS proposes to increase the pace of repaying covered entities for the 2018-2022 period from 0.5% per year to 3% per year. In practice, this changes nothing about the structural loss from the new ongoing payment rate, and the additional recoupment amounts are a one-time offset against a permanent annual $4.85 billion reduction.

Who Wins and Who Loses

The distributional effects of this rule are not ambiguous. They follow directly from the structure of who participates in 340B and how outpatient drug revenue flows through the system.

The winners are hospitals with low 340B participation rates, particularly academic medical centers in suburban markets and community hospitals with commercially-insured patient bases. They benefit from the base rate increase without absorbing the 340B cut. Medicare beneficiaries save an estimated $1.15 billion in lower drug cost-sharing. Pharmaceutical manufacturers benefit because fewer 340B-eligible purchases reduce their total program obligations.

The losers are concentrated among institutions that have the least capacity to absorb financial shocks. Safety-net DSH hospitals, which often operate on 0.5-1.5% operating margins, will face a revenue reduction that could force service cuts, program closures, or in the worst cases, facility closures. FQHCs and rural health clinics face similar dynamics. The patients they serve, who have limited alternatives and rely on these facilities for access to complex specialty care and medication programs, are the ultimate losers.

What Happens If This Goes Through

The 340B spread does not just pad hospital balance sheets. America's Essential Hospitals has documented how these funds flow: into uncompensated care programs, specialty services, care coordination for high-need patients, patient drug assistance programs, and the operational infrastructure that allows safety-net hospitals to remain financially viable.

America's Essential Hospitals also raised a critical methodological concern: CMS based its rate-setting analysis on data from less than 25% of 340B covered entities. That is a significant sampling limitation for a rule with this level of financial impact. The adequacy of that data analysis should be a central focus of public comments.

If the cut goes through at the proposed level, the most immediate effect will be on the most complex and expensive services at safety-net hospitals. Oncology, specialty pharmacy, infusion services, and HIV/AIDS programs are disproportionately dependent on 340B drug revenue. These are precisely the services that low-income patients have no other access point for.

The longer-term effect is an accelerating consolidation of those services into better-funded institutions in wealthier markets, with a corresponding reduction in access for the patients who most need them. This is not a speculative concern. It is the documented pattern from the 2018 cut, which preceded service reductions at multiple DSH hospitals before the legal reversal.

The 340B Counterargument: What CMS Gets Right

Intellectual honesty requires acknowledging that the 340B program has been criticized, including from credible sources, for allowing revenue capture without commensurate documentation of patient benefit. The Government Accountability Office and others have noted that the program lacks robust requirements for covered entities to demonstrate that 340B savings directly flow to patient care.

CMS is not wrong to note that the gap between acquisition cost and Medicare reimbursement has generated substantial revenue for some covered entities. The question is whether a blunt payment rate cut is the appropriate policy instrument, or whether targeted transparency requirements and program integrity safeguards would achieve the same goal without collateral damage to the hospitals most dependent on the spread.

What Safety-Net Leaders Should Do Right Now

The comment period closes August 31, 2026. That deadline is not a formality. CMS is required to respond to substantive comments in the final rule, and a well-organized comment campaign from covered entities, patient advocates, and state associations has historically influenced the final rule's scope.

• Document the specific programs funded by 340B at your institution. Generic impact statements carry less weight than precise data on service lines, staffing, and patient populations at risk.

• Challenge the data methodology. CMS relied on less than 25% of covered entities in its analysis. Comments that directly address this limitation, with supporting data from your institution, have procedural and legal weight.

• Quantify the distributional impact. Show specifically how the budget-neutral redistribution fails to offset the 340B cut for hospitals with your patient mix.

• Coordinate with state hospital associations. A coordinated comment campaign from multiple institutions is more difficult to dismiss than individual submissions.

• Submit comments at regulations.gov under the CY 2027 OPPS proposed rule docket. The deadline is August 31, 2026.

• Engage your Congressional delegation. Congressional oversight letters and hearings have changed the trajectory of 340B policy before.

The August 31 Deadline

The 2027 OPPS proposed rule is not final. CMS will publish a final rule later this fall, and that final rule will reflect both the substantive comments received and the political environment around the proposal.

The 2018 cut ultimately took years of litigation to reverse. The institutions that survived that period were the ones that had the financial reserves and operational flexibility to absorb the revenue loss temporarily. Many safety-net hospitals do not have that runway in 2026. The margins are thinner. The patient populations are more complex. The post-pandemic operating environment has left fewer reserves to draw on.

This makes the August 31 comment deadline more important, not less. If the rule is finalized at the proposed cut level, the legal and administrative remedies available afterward are slower and more uncertain than the regulatory process right now.

The $4.85 billion question is not whether CMS has the authority to cut 340B rates. It is whether the proposed cut is the right policy, at the right level, applied to the right institutions, in a way that serves Medicare beneficiaries and the communities that depend on safety-net care. The comment period is the mechanism for demanding that answer. Use it.

Oatmeal Health tracks healthcare policy, AI, and the business of care. If this analysis was useful, share it with a colleague navigating the same policy environment. The comment deadline is August 31, 2026 at regulations.gov.

Let's start with what we know, because the data is unusually clear.

In June 2026, the Office of Inspector General published two back-to-back reports on Medicare Advantage (MA) denial rates for post-acute care. These weren't advocacy pieces. They weren't patient complaints. They were federal audits of actual claims data from the 19 largest Medicare Advantage Organizations (MAOs), covering Long-Term Care Hospitals (LTCHs), Inpatient Rehabilitation Facilities (IRFs), and Skilled Nursing Facilities (SNFs).

Here's what they found:

The 19 largest MAOs denied 65% of LTCH claims, 54% of IRF claims, and 12% of SNF claims. But the headline numbers obscure the individual insurer performance, which is where it gets damning.

CVS/Aetna denied 80% of LTCH claims and 51% of IRF claims. Humana denied 72% of LTCH claims and 54% of IRF claims. UnitedHealthcare denied 71% of LTCH claims and 66% of IRF claims.

These are not marginal facilities. Long-Term Care Hospitals serve patients recovering from catastrophic illness - strokes, traumatic injuries, ventilator dependence, complex wounds. Inpatient Rehabilitation Facilities treat patients after hip replacements, spinal injuries, and brain trauma. These patients need this care. Their physicians prescribed it. And at rates between 65% and 80%, the largest insurers in the country said no.

The Appeals Data Tells the Real Story

Here is the part the insurers would prefer you not think about too hard.

When patients and providers appeal these denials, they win - a lot. For LTCH denials, 36% are overturned on appeal. For IRF denials, 43% are overturned on appeal. For SNF denials, the number is extraordinary: 95% of appealed SNF denials are overturned.

Read that again. When SNF denials are appealed, providers win 95% of the time.

There are two ways to interpret this. Either the initial denial process is so broken that nearly every SNF denial is wrong, or the appeals process is broken in the other direction. Given that these are federal audits of actual outcomes, the former is far more likely.

What does it mean when you deny 12% of SNF claims and then lose 95% of the appeals? It means you're denying medically necessary care, getting caught, and reversing course - but only for the patients and families who have the time, knowledge, and energy to fight back. For everyone else, the denial sticks.

UnitedHealth's Algorithm Problem

The OIG reports specifically flagged a pattern worth understanding: contractor-driven denials that the MAOs themselves overturned on appeal. In plain English, insurance companies hired third-party contractors to make initial denial decisions using algorithms, and then those same insurance companies reversed the algorithmic decisions when humans reviewed them on appeal.

This is not a theoretical concern. The DOJ is currently investigating UnitedHealth Group for exactly this issue. UHC's denial rate went from 8.7% in 2019 to 22.7% in 2022 - a 161% increase - after the company acquired NaviHealth, a post-acute care management company that uses algorithmic prediction tools.

UHC's SNF denial rate went from 1.4% in 2019 to 12.6% in 2022 - a ninefold increase. In three years. After acquiring a predictive algorithm company.

The OIG's guidance in these reports was explicit: MAOs cannot rely solely on an algorithm without considering the member's specific circumstances. This is not just a best practice recommendation. It is a statement about what the law requires. The algorithm-only approach, which drove the post-NaviHealth denial surge, was illegal.

The DOJ investigation is not just civil. According to reporting in 2024 and 2025, UnitedHealth faces both civil and criminal exposure. The company's CEO was murdered in December 2024 - a national news event that briefly made the entire country aware of what the healthcare denial machine looks like from the outside.

This Is What $76 Billion in Overpayments Buys

Here is the full picture, because the denial story and the overpayment story are two sides of the same coin.

MedPAC - the independent Congressional advisory body on Medicare - estimates that MA plans will receive $76 billion in overpayments from CMS in 2026. That is $76 billion above what traditional Medicare would have cost for the same beneficiaries.

These overpayments come from two sources: favorable risk selection (enrolling healthier-than-average seniors) and upcoding (adding diagnosis codes to make patients look sicker than they are, which increases per-member payments). OIG found $462 million in overpayments on unsupported stroke diagnoses alone in 2021. Kaiser Permanente paid a $556 million DOJ settlement in January 2026 for unsupported diagnosis codes. Elevance paid $342 million to CMS in May 2026.

The business model works like this: collect more money by making patients look sicker on paper, then deny the care those patients actually need. The spread between inflated risk scores and actual care delivered is profit.

The scale is staggering. MA now covers 35 million Americans - 55% of all Medicare-eligible seniors, up from just 19% in 2007. The program grew by 1.1 million beneficiaries between 2025 and 2026 alone. Those overpayments are funded by all Medicare beneficiaries: they drove the Part B premium from $185 in 2025 to $203 in 2026. Every American on Medicare is paying higher premiums to fund a system that denies post-acute care to the sickest seniors at 80% rates.

CMS Is Watching But Not Acting

In April 2024, CMS finalized new prior authorization rules requiring MA plans to make expedited decisions within 72 hours and standard decisions within 7 calendar days. These are improvements. They don't take full effect until 2026.

In January 2027, MA plans will be required to implement HL7 FHIR-based Prior Authorization APIs - meaning electronic, real-time authorization requirements that providers can query directly. This is meaningful infrastructure change.

But in June 2025, CMS announced it will not enforce the utilization management health equity reporting rules it had previously established. The agency cited administrative burden. The rules would have required MA plans to report on whether their denial rates varied by race, income, and geography - data that might have revealed whether the 80% LTCH denial rate falls disproportionately on certain populations.

CMS suspended those transparency requirements right when they would have started to bite.

The enforcement picture is mixed at best. The DOJ is pursuing criminal and civil cases. CMS is requiring faster decision timelines. But the fundamental economics haven't changed. An MA insurer that denies 80% of LTCH claims and reverses 36% of those on appeal is still, on net, denying more than half of medically necessary LTCH placements. The math works in their favor as long as most patients don't appeal.

The Moral Hazard of Managed Medicare

I want to be precise about what I am and am not arguing here.

I am not arguing that utilization management is inherently wrong. Every insurer, including traditional Medicare through its MACs, reviews claims and denies some percentage of them. Some denials are appropriate. Prior authorization, when done well, can reduce unnecessary care and lower costs.

What I am arguing is that an 80% denial rate, combined with a 36% appeal overturn rate, is not utilization management. It is a business strategy.

Utilization management done correctly looks like: review the claim, verify medical necessity using evidence-based criteria, make a decision within 24-72 hours, approve most clinically appropriate care, deny a small percentage with strong clinical rationale, and lose very few appeals because your initial decisions were correct.

What CVS, Humana, and UnitedHealth appear to be doing looks like: deny by default using an algorithm, count on most patients not appealing, reverse denials when challenged, and treat the 36-43% overturn rate as an acceptable cost of doing business.

The difference matters enormously when you're talking about post-acute care. A stroke patient who needs inpatient rehab doesn't have weeks to wait for an appeal. A ventilator-dependent patient whose LTCH placement is denied doesn't have the bandwidth to navigate an administrative process while managing a life-threatening illness. A family dealing with a parent's hip fracture recovery doesn't know they have appeal rights, let alone how to exercise them.

Denial is most harmful when it falls on the most vulnerable patients, who are least able to fight back.

What Should Change

The OIG reports make specific recommendations. CMS should require MAOs to document why they denied claims and track denial rates by facility type. CMS should audit plans with outlier denial rates. Plans should not be allowed to use algorithm-only review without human clinical oversight.

These are the right recommendations. They are not sufficient.

What would actually move the needle:

First, minimum overturn rate triggers. Any MA plan with an appeals overturn rate above 30% in any post-acute care category should face automatic review and rate adjustments. You cannot lose 43% of your appeals and call that good-faith utilization management.

Second, denial rate benchmarks tied to traditional Medicare. CMS knows what traditional Medicare's denial rates are. MA plans should not be allowed to run denial rates more than two standard deviations above the traditional Medicare baseline for the same beneficiary risk categories.

Third, real-time adverse event tracking. When a patient is denied post-acute care and subsequently requires emergency readmission, that outcome should be automatically linked to the prior denial and flagged for review. MA plans should bear financial consequences for readmissions that follow denials of medically necessary post-acute care.

Fourth, eliminate the algorithm defense. Plans that use predictive tools for initial denial decisions should be held to the same standard as human reviewers. The OIG has already said plans cannot rely solely on algorithms. CMS should codify this as an enforceable standard with financial penalties.

The 35 Million Person Question

Medicare Advantage has genuinely delivered value in some areas. Care coordination, dental and vision benefits, wellness programs - these are real improvements over fee-for-service Medicare for many beneficiaries. The program is not without merit.

But merit in some areas does not offset systematic denial of medically necessary care in others. And the scale of MA - 55% of all Medicare beneficiaries, $400+ billion in federal spending annually - means that systematic problems in MA are systematically large problems.

Thirty-five million Americans chose Medicare Advantage because they were told it would offer better coverage and lower costs than traditional Medicare. Many of them will eventually need post-acute care. When they do, there is approximately a 65-80% chance, if they're in an LTCH or IRF, that their MA plan will say no.

The insurers know this. CMS knows this. The OIG has now documented it twice in a month. The DOJ is investigating the largest player.

The question is whether any of this produces actual change before another cohort of stroke patients and hip fracture survivors gets algorithmically denied into discharge destinations that compromise their recovery.

Based on the pace of federal enforcement in health insurance, I am not optimistic. The MA denial machine is profitable, entrenched, and protected by the same lobbying apparatus that keeps drug prices high and surprise billing alive. The OIG reports will generate congressional letters. The DOJ settlements will generate press releases. And on Monday morning, the algorithm will keep running.

That is the scandal. Not that it happened. That we have documented it thoroughly and are largely going to let it continue.

Oatmeal Health covers healthcare policy, payment reform, and the business of medicine. Sources for this piece include the OIG June 2026 reports on MA denial rates for LTCHs, IRFs, and SNFs; MedPAC March 2026 Report to Congress; KFF Medicare Advantage 2026 enrollment data; and prior reporting on UnitedHealth NaviHealth acquisition outcomes.

Every care manager in America is running the same impossible math.

More patients. Fewer clinicians. Chronic conditions that demand daily attention but receive monthly appointments. Heart failure that decompensates between visits. Hypertension that goes uncontrolled because no one is watching between the third Tuesday of every month.

This is not a staffing problem. It is a contact frequency problem.

And Cadence just raised $100 million to fix it.

On June 23, 2026, Cadence announced a Series C led by Spark Capital, bringing total funding to $241 million at a $1.23 billion valuation. The round included Thrive Capital, General Catalyst, Coatue, B Capital, and the venture arms of Corewell Health, Memorial Hermann, and Duke Health.

The company is treating 100,000 patients today with AI agents that monitor their chronic disease in real time. And it tripled annual recurring revenue in 2025.

Here is what the funding announcement obscures: this is not a remote monitoring company. Cadence is an AI-operated chronic care delivery system. That distinction is the entire business model.

1. The Chronic Disease Crisis Is a Contact Problem

56 million older adults in the United States live with at least one chronic condition requiring active management. The real number is higher: between 55 and 98 percent of adults over 60 have two or more chronic diseases simultaneously.

Heart failure affects 7.7 million Americans, up from 6.7 million just three years ago. It is the fastest-growing cardiovascular phenotype in the country. Hypertension and diabetes co-occur in more than one-third of all Medicare beneficiaries. Nearly half of US adults now meet criteria for cardiovascular disease when hypertension is included.

The clinical consequence of this is straightforward: these patients need more contact points than the current system provides.

A physician managing a heart failure patient on four guideline-directed therapies cannot clinically monitor them adequately in a 15-minute quarterly appointment. Blood pressure does not cooperate with scheduling. Medication adherence does not pause for office hours. Fluid accumulation does not announce itself with enough warning for a monthly check-in to catch it in time.

📊 The average heart failure patient is hospitalized 1.3 times per year. Each admission costs the system approximately $21,000. At 7.7 million patients, that is a $198 billion annual inpatient cost line driven in large part by preventable decompensation.

The care model is not working because the contact model is broken.

Traditional chronic care management - in-person visits, monthly calls, quarterly labs - was designed for an era when chronic conditions were exceptions, not the rule. When 55 to 98 percent of your patients have multiple conditions, monthly touch points are not care management. They are triage intervals.

2. What Cadence Actually Builds

Cadence is not a monitoring company. Most remote patient monitoring companies give clinicians a dashboard - more data, more alerts, more things to triage. That is adding cognitive burden to a workforce already at capacity.

Cadence built the layer between the sensor and the clinician.

Patients are enrolled with connected devices that track vitals continuously. AI agents monitor that data stream, apply risk stratification models, surface alerts that need attention, and in 55 percent of cases resolve those alerts without requiring any human clinical time at all. When the AI flags something that warrants a human touch, the median response time for a care team member is 3.5 minutes.

The AI agents handle the routine. The clinicians handle the irreplaceable.

This matters because the math of traditional CCM breaks down at scale. A single care manager can meaningfully manage somewhere between 80 and 120 patients under traditional workflows. With AI handling triage, alert resolution, and documentation, that ratio changes dramatically.

Cadence charges payers - primarily Medicare - a monthly per-patient fee for this service. The company operates as a clinical service business, not a technology licensing company. It takes accountability for outcomes, not just data delivery.

New affiliations with Duke Health and Texas Health Resources, both announced alongside the Series C, expand the network of health systems using Cadence as a chronic care delivery layer. These are not pilot agreements. They represent full deployment at health systems that collectively serve hundreds of thousands of Medicare patients.

3. The Outcomes That Actually Matter

In February 2026, Cadence published results in Mayo Clinic Proceedings: Innovations, Quality and Outcomes. The study, titled "The Impact of a Remote Patient Care Program on Health Care Costs and Utilization Among Medicare Patients With Chronic Disease," is the most rigorous validation of their model to date.

The numbers are not incremental.

Blood pressure control improved by 43 percent relative to baseline - from 42 percent of patients controlled to 60 percent. Systolic blood pressure fell by 6.9 mmHg. Diastolic by 4.9 mmHg. Both statistically significant.

In heart failure patients with reduced ejection fraction, the percentage on all four guideline-directed medical therapy pillars doubled - from 15 percent to 31 percent. That is a 107 percent increase in appropriate GDMT adherence. In patients with preserved ejection fraction, the improvement was even more dramatic: from 2 percent to 8 percent, a 300 percent relative increase.

Hospital admissions fell 27 percent.

Total cost of care fell by $203 per patient per month.

📊 At 100,000 active patients, that $203/month cost reduction translates to more than $243 million in annual system savings - and Cadence is saving Medicare $2.7 million every single week.

These outcomes are not from a controlled academic trial with cherry-picked patients. This is real-world Medicare data across a diverse patient population, including rural and underserved communities.

That last detail matters. Clinical trials that exclude low-income and minority patients have historically produced results that do not hold up in the populations where chronic disease burden is highest. Cadence's outcomes data holds in those populations.

4. The Reimbursement Architecture Underneath This

The financial model of AI-powered chronic care is not complicated, but most health systems do not have a team that has mapped it.

Medicare has an established billing framework for chronic care management and remote patient monitoring. The 2026 Medicare Physician Fee Schedule increased CCM reimbursement by 10 percent across all codes - the largest single-year increase in program history. Two new RPM codes were added: CPT 99445 for 2-15 days of monitoring data and CPT 99470 for the first 10 minutes of monitoring time.

The revenue architecture, stacking CCM and RPM together, looks like this:

• CPT 99490: $66.13 per patient per month (first 20 minutes of non-complex CCM)

• CPT 99439: ~$50.44 per patient per month (additional 20 minutes)

• CPT 99457: ~$53 per patient per month (first 20 minutes RPM management time)

• CPT 99454: ~$68 per patient per month (device supply and data transmission)

Layered appropriately, CCM plus RPM generates approximately $200 to $300+ per patient per month in Medicare reimbursement.

This is not speculative revenue. This is Medicare reimbursement at established rates. The constraint has never been the billing codes. It has been the clinical capacity to actually deliver the required monitoring and management time. That is exactly what AI agents address.

5. The FQHC Opportunity - and the Gap That Still Exists

Federally Qualified Health Centers are the natural home for this model.

The patient populations FQHCs serve - low-income adults, uninsured and Medicaid-dependent households, rural communities with limited specialist access - carry the highest chronic disease burden in the country. Hypertension, diabetes, heart failure, and COPD concentrate in the same zip codes that FQHCs serve. These patients have the most to gain from continuous monitoring and the least access to the traditional care model.

And yet, FQHC adoption of RPM and AI-assisted chronic care management remains low.

The barriers are real. Broadband access is inconsistent in the communities FQHCs serve. Not all patients have smartphones with reliable data plans. Device procurement requires upfront capital that community health centers on 55 percent cost-to-revenue ratios cannot easily absorb. And the IT staffing to implement, configure, and troubleshoot a remote monitoring program typically does not exist at a center running one IT staffer across ten sites.

📊 A 2026 survey of FQHC administrators found that more than 60 percent cite lack of dedicated care management staff as the primary barrier to scaling RPM programs - not reimbursement, not patient willingness.

The 2026 CMS rule changes help at the margin. The new CPT 99445 code - which allows billing for 2-15 days of monitoring data instead of requiring a full 16 days - reduces the program enrollment burden significantly for patients who cannot maintain daily device contact.

But the structural access problem for FQHCs goes deeper than billing codes.

A model like Cadence's, where AI agents handle the monitoring workload and human clinical time is deployed only for cases that require it, is specifically designed to address the FQHC staffing constraint. The question is whether health systems and FQHCs can establish partnership arrangements that give safety-net providers access to the platform economics that currently flow primarily to large health systems with existing revenue cycle infrastructure.

6. The Competitive Landscape

Cadence is not operating in an empty market.

Biofourmis, which has raised more than $463 million in total funding, targets the same space with a focus on wearable biometric sensors and predictive deterioration models, particularly for cardiovascular and respiratory conditions. Innovaccer has positioned itself as a data platform play, integrating disparate health records and layering care management analytics on top.

What distinguishes Cadence is the clinical service model rather than the technology licensing model. Cadence does not sell a platform and walk away. It operates as a care team member, taking accountability for outcomes and billing Medicare directly. That accountability alignment - where the company's revenue depends on keeping patients healthy and out of the hospital - is what justifies the clinical confidence to report Mayo Clinic outcomes publicly.

The AI in remote patient monitoring market is valued at $3.35 billion in 2026 and is projected to grow at a 27 percent CAGR through 2031. With Cadence at $1.23 billion valuation, it is one of the most valuable standalone players in a market that is still early in consolidation.

Deep Dive: The Unit Economics of AI-Managed Chronic Care

The $100 million Series C is a bet on a specific unit economics model, and the math is worth understanding in detail.

Cadence is treating 100,000 patients today. Its Mayo Clinic Proceedings data shows $203 per patient per month in total cost-of-care reduction. At current enrollment, that is approximately $20.3 million per month, or $243.6 million per year, in system savings.

Medicare pays for the care management through CCM and RPM codes. Stacked appropriately, a patient enrolled in a comprehensive program generates approximately $200 to $300 per patient per month in billable services.

Cadence takes a service fee from health systems or payers. The company has published that it delivers "more than $3 in Medicare savings for every dollar Medicare spends." At $200 per patient per month in reimbursement and $600+ in savings, the ROI justification for a payer or health system to contract with Cadence is not abstract.

The visual below breaks out the annual value per patient across the four components of the Cadence economic model.

The Scale Trajectory

Cadence tripled annual recurring revenue in 2025. The Series C is designed to accelerate the next growth phase: deeper health system partnerships, expansion into new chronic conditions beyond hypertension and heart failure, and geographic expansion into markets where chronic disease burden is highest.

The Duke Health and Texas Health Resources affiliations signal the kind of institutional validation that drives payer adoption. When an academic medical center puts its name on a partnership, it signals clinical confidence that accelerates commercial contracting.

The Risk Factors

The model's strength is also its primary risk. Cadence takes accountability for outcomes. If outcomes degrade at scale - if the clinical quality that produced those Mayo Clinic Proceedings numbers does not hold across 500,000 patients - the revenue model unwinds.

The AI agent layer is not infinitely scalable without ongoing model validation. Alert fatigue, false negative suppression, and model drift are real risks in any AI clinical system. At 55 percent of alerts resolved by AI without human review, the quality of that automated resolution layer is load-bearing for patient safety.

The second risk is payer negotiations. Medicare reimbursement rates are not guaranteed to stay at 2026 levels. Any downward fee schedule adjustment to CCM or RPM codes directly compresses Cadence's unit economics.

What This Means For You

• FQHC executives and community health center CEOs: The Cadence model is designed to address your staffing constraint - AI agents handling triage so your limited clinical staff handle only cases that require human judgment. The barrier is partnership access and IT infrastructure, not patient eligibility. Push your health system partners on whether chronic care AI platforms can extend to your patient panel under their contracts.

• Health system CMOs and CFOs: At $203 per patient per month in cost reduction and $200 to $300 in stacked CCM/RPM reimbursement, the ROI math on AI-assisted chronic care is not ambiguous. The question is not whether to deploy this model - it is whether to build, partner, or acquire your way into it before your competitor does.

• Radiologists and pulmonologists: Your highest-acuity lung and cardiac patients are often the same patients with uncontrolled chronic disease at the time of diagnosis. AI chronic care management that achieves guideline-directed therapy compliance before a cancer or cardiac event improves the outcomes you are measuring at the time of intervention.

• Healthcare investors and founders: The AI in RPM market is valued at $3.35 billion in 2026 with a 27 percent projected CAGR. Clinical service models with accountable outcome contracts are outcompeting pure technology licensing models. Follow the payer contract structure, not the software pitch.

• Policy advocates: The Cadence model saves Medicare $2.7 million every week at 100,000 patients. Scaling this to the millions of eligible Medicare beneficiaries who currently receive inadequate chronic care management would generate billions in program savings annually. CMS should be examining how to accelerate this model in safety-net settings, not just at large health systems with existing infrastructure.

The future of chronic care is not more clinicians. The country cannot train them fast enough. It is not more monitoring dashboards. Those already exist and are already being ignored.

The future of chronic care is AI agents that handle the contact frequency problem automatically, surface only the cases that require human judgment, and give clinicians back the time to do the work that actually requires them.

Cadence just proved that model works at scale. And Medicare is paying for it.

The question for every health system CEO reading this: what is your chronic care management strategy, and does it account for a world where your competitors' AI agents are monitoring 100,000 patients while yours are reading dashboards?

Reply to this email with your answer. I read everything.

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise. Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• Cadence $100M Series C announcement (Business Wire, June 23, 2026)

• Mayo Clinic Proceedings: "The Impact of a Remote Patient Care Program on Health Care Costs and Utilization Among Medicare Patients With Chronic Disease" (February 2026)

• 2026 CCM/RPM Reimbursement Rates: CPT 99490 at $66.13/month; 10% CCM increase across all codes (CMS 2026 PFS Final Rule)

• AI in Remote Patient Monitoring Market (MarketsandMarkets, 2026): $3.35B market, 27% CAGR through 2031

• FQHCs and RPM in 2026: Barriers and CMS Policy Changes (Prevounce, Tenovi, 2026)

The headline said it plainly: "New Rural Health Funding Only a Bandage for Medicaid Wound."

That is not an editorial opinion. That is a cash flow analysis.

Congress is promoting a $50 billion Rural Health Transformation Program. The number is large enough to generate a White House announcement and warm applause at health center conferences. Average first-year state awards are running $200 million. Some states are putting forward thoughtful implementation plans. The political optics are good.

But look at what is happening at the same time. The Community Health Center Fund - the Section 330 grant program that has been the foundational operating support for federally qualified health centers for decades - was extended only through December 2026. No long-term authorization. No inflation adjustment. Just a short runway over a cliff.

And Medicaid, which accounts for 43 percent of FQHC operating revenue nationwide, is being cut by a projected $911 billion over 10 years under the One Big Beautiful Bill. Work requirements alone are projected to cause 5.6 million community health center patients to lose Medicaid coverage over five years.

Three legs of the FQHC financial stool - Section 330 grants, Medicaid reimbursement, and 340B drug savings - are all under simultaneous pressure. A $50 billion rural program that reaches only some health centers in some states is not a solution to that math. It is a headline pasted over a structural failure.

1. What the Rural Health Transformation Program Actually Does - and Does Not Do

The Rural Health Transformation Program is real money. $50 billion over five fiscal years, $10 billion per year, from 2026 through 2030. It is allocated to approved states, with half divided equally among all states and half weighted by rural population, number of rural health facilities, and related factors.

The program is designed to support Critical Access Hospitals, rural health clinics, FQHCs receiving Section 330 grants that serve rural populations, and certain behavioral and mental health providers. CMS has begun awarding first-year funding, with state awards averaging $200 million in year one.

Some states are putting RHT funds to immediate use. Delaware is investing in an "FQHC Value-Based Care Readiness" initiative covering technology upgrades, expanded care teams, and care management infrastructure. That is the right kind of use for this funding.

📊 First-year RHT state awards average $200M - spread across all rural health entities in a state, not directed exclusively to FQHCs.

Here is the structural problem. The Rural Health Transformation Program helps rural FQHCs in approved states. That is a meaningful subset of the community health center system. It is not the whole system.

There are approximately 1,400 federally qualified health centers operating more than 14,000 service delivery sites across the country. A significant share of them are in urban and suburban areas - serving the same Medicaid-dependent, uninsured, and underinsured patient populations as their rural counterparts. They face the same financial pressure. They get nothing from RHT.

More fundamentally, the Rural Health Transformation Program is a new grant program layered on top of the existing funding structure. It does not replace Section 330. It does not replace Medicaid reimbursement. It does not protect 340B. It adds a new revenue stream for a limited subset of health centers while the three foundational revenue streams continue to erode.

The Authorization Cliff

The Community Health Center Fund was authorized through December 2026 by the Consolidated Appropriations Act signed in early 2026. That $4.6 billion in FY2026 funding is real support. But the authorization cliff at year-end creates a crisis that health center administrators are already managing right now.

Hiring decisions cannot wait until December to know whether funding continues. Capital investments in new sites or equipment require multi-year confidence in revenue. Workforce planning for clinical staff - physicians, nurse practitioners, dentists, behavioral health providers - operates on 18 to 24 month time horizons.

When Section 330 authorization ends at December 2026 with no successor legislation, health centers do not wait to see what Congress does. They stop hiring in October. They delay site expansions. They hold staff positions open rather than filling them. The damage from funding uncertainty arrives before the funding actually stops.

2. The Three-Legged Stool

Community health center finance is built on three structural revenue sources. Remove any one of them and a health center's operating model strains to its limit. Remove two simultaneously and you are talking about closures, not adjustments.

Leg One: Medicaid

Medicaid is the single largest payer for community health centers, accounting for 43 percent of operating revenue on average. More than 16 million CHC patients carry Medicaid coverage. When those patients lose coverage, they do not disappear from the health center - they show up as uncompensated care.

The One Big Beautiful Bill included $911 billion in Medicaid spending reductions over 10 years. The most direct impact on community health centers comes from two provisions working in combination.

First, work requirements. Adults ages 19 to 64 in Medicaid expansion states must now document at least 80 hours per month of work or qualifying activity as a condition of eligibility, effective no later than December 31, 2026. Arkansas tried this in 2018. In six months, 18,164 people lost coverage. Most of them were already working. The paperwork infrastructure to document eligibility did not exist, and eligible patients could not navigate the online portal.

NACHC projects that over five years, nearly 5.6 million CHC Medicaid patients in expansion states could lose coverage under these requirements. That translates to direct revenue loss for health centers - and an offsetting surge in uncompensated care for the same facilities.

📊 5.6 million CHC Medicaid patients projected to lose coverage over 5 years under the new work requirements.

Second, per capita caps and enrollment restrictions. As states implement coverage changes and as enhanced ACA marketplace subsidies that expired at the end of 2025 push more patients back to uninsured status, the paying patient mix at health centers shifts sharply toward nonpaying. Health centers are legally required to serve anyone regardless of ability to pay. The operating model assumes a sustainable ratio of paying to nonpaying patients. When that ratio breaks, the model breaks with it.

Leg Two: Section 330 Grants

Section 330 of the Public Health Service Act has been the federal backbone of community health center funding since 1965. These grants currently total approximately $4.6 billion annually and are distributed through HRSA directly to health centers.

The grants fund services that are not fully reimbursed through Medicaid or private insurance - behavioral health services, dental care, pharmacy services, transportation coordination, case management, and enabling services. These are the services that most differentiate a federally qualified health center from a standard primary care clinic. They are also the services most likely to catch a lung cancer screening referral, a maternal mental health flag, or a patient drifting off their chronic disease management plan.

When grant funding is uncertain, these enabling services are the first to be cut. The services most vulnerable are often the most important for population health outcomes in underserved communities.

📊 $4.6 billion in Section 330 grants currently support 14,000 FQHC service delivery sites - authorized only through December 2026.

Leg Three: 340B

The 340B Drug Pricing Program, established in 1992, requires pharmaceutical manufacturers participating in Medicaid to sell outpatient drugs to eligible safety-net providers at deeply discounted prices - typically 20 to 50 percent below market rates. Community health centers are among the core eligible entities.

340B savings do not sit in a reserve fund. They subsidize the services that Medicaid and grant funding do not fully cover. They allow a health center to maintain an on-site pharmacy and dispense medications to patients who cannot afford to fill a prescription at a retail pharmacy. They fund the care coordination staff who prevent $50,000 emergency hospitalizations.

The 340B program is under sustained attack from multiple directions at once. Pharmaceutical manufacturers are restricting contract pharmacy arrangements - the mechanism through which FQHCs access 340B pricing through external pharmacies. Pharmacy benefit managers have introduced billing and reimbursement practices that erode the effective discount. On Capitol Hill, proposals circulate to tighten eligibility or restrict how savings can be used.

None of these threats has individually dismantled 340B. Together, they are compressing the margins that health centers depend on to serve patients who cannot pay full freight.

3. The Urban FQHC Is Being Left Out

The Rural Health Transformation Program creates a structural bifurcation in the community health center system. Rural FQHCs in approved states can potentially access RHT funding. Urban FQHCs cannot.

This matters because urban patient populations at community health centers are not categorically healthier, wealthier, or less dependent on safety-net services than rural populations. A health center in South Chicago or East Los Angeles or North Philadelphia serves patients with the same clinical complexity, the same Medicaid dependency, and the same lack of alternatives as a center in rural Mississippi.

The policy logic of prioritizing rural health is understandable. Rural hospital closures are accelerating. Provider shortages in rural areas are severe. The public health statistics in rural communities are dire. Investment is warranted.

But the framing of "rural health investment" as a counterweight to broader Medicaid cuts does not hold up at the aggregate level. The Medicaid cuts affect health centers in Chicago as hard as they affect health centers in rural Vermont. The Section 330 authorization cliff hits every health center equally. The 340B erosion is not confined to rural settings.

A policy that offers targeted rural investment while cutting the universal funding mechanisms is a net negative for the system as a whole, even if some rural centers individually come out ahead.

4. What Is Already Happening on the Ground

The closures and service reductions are not hypothetical. They are already in progress.

In New Hampshire, a rural FQHC announced the closure of one of its sites in late 2025, citing a projected operating shortfall driven by Medicaid funding changes. In South Carolina, an FQHC closed six locations and transferred services to other agencies, citing financial pressure and a rising share of uninsured and underinsured patients.

These closures are happening before the Section 330 authorization cliff, before Medicaid work requirements fully take effect in December 2026, and before ACA marketplace subsidy loss converts additional newly insured patients back to uninsured status.

One in four community health centers currently faces a negative operating margin. NACHC projects that without increased funding, one in four may close or reduce services over the next two years.

📊 1 in 4 community health centers currently operates at a negative financial margin.

The compounding timeline matters. December 2026 is when Section 330 authorization expires. That is also when Medicaid work requirements must be implemented nationwide. That is also the end of a year in which ACA marketplace subsidies have already been eroding since enhanced subsidies expired at the end of 2025. Three major financial stressors are converging at the same calendar point.

Health center administrators are not waiting to see what happens. They are making decisions right now about what they will be able to operate in 2027.

5. The Three-Scenario Financial Model

Community health center financial planning in 2026 requires modeling three distinct scenarios, because the policy environment is genuinely uncertain across all three funding legs.

Scenario A: Partial Mitigation

RHT funding flows efficiently to rural health centers. Section 330 is reauthorized in early 2027 at current funding levels. Medicaid work requirement implementation is slower than projected due to state administrative delays. Net revenue loss to the system runs approximately $2 to 3 billion annually. Health centers in this scenario survive but cut enabling services, reduce hours at some sites, and defer capital investment.

Scenario B: Base Case - Current Trajectory

Section 330 authorization lapses in December 2026 for several months before a continuing resolution restores it at flat funding. Medicaid work requirements reduce coverage for millions of patients starting in mid-2027. 340B pressures continue to compress savings. Annual system revenue loss reaches approximately $7 billion - the NACHC projection. Approximately 1,800 FQHC sites close or reduce to appointment-only operations. 34,000 jobs are eliminated. 32 million current patients see reduced access.

Scenario C: Compounding Worst Case

Section 330 lapses without a quick continuing resolution. Medicaid work requirements are implemented on schedule with higher than projected coverage loss. A legislative change to 340B further restricts program savings. Annual system revenue loss exceeds $10 billion. Urban health centers face disproportionate closures because they have no RHT offset.

Deep Dive: The Mechanics of an FQHC Closure

Health centers do not close overnight. They close over 18 months of deferred decisions that accumulate until the operating model fails. Understanding the sequence is essential for any health system leader, policy advocate, or investor trying to intervene before the announcement.

It typically starts with a hiring freeze. A nurse practitioner leaves and the position stays open. A medical director announces retirement and the replacement search is put on hold. Each open clinical position reduces visit capacity by 15 to 20 percent.

Then the enabling service cuts begin. Transportation coordination goes first - it generates no fee revenue and requires dedicated staff time. Dental services get reduced to emergency extractions because the dental hygienist position cannot be refilled. Behavioral health visits are capped at six per patient per year instead of 12. The pharmacy closes on Fridays.

By the time a community health center announces a closure, the community has already been losing access for a year. The announcement is the last domino, not the first.

📊 The typical FQHC closure sequence: hiring freeze at month 1, enabling service cuts at month 6, site consolidation at month 12, formal closure announcement at month 18.

For health system administrators tracking referral patterns: a health center announcing a hiring freeze in the context of funding uncertainty is the intervention window. Not the closure announcement. When a health center in your community enters this sequence, your emergency department is going to absorb those patients in about 18 months. Model that volume now.

For anyone who has watched what happens to communities after FQHC closures: the primary care access gap takes years to fill. Private practices do not open in rural counties with median incomes under $35,000. Urgent care chains do not follow communities into the neighborhoods that community health centers serve. The patients who depended on that health center end up in emergency departments for primary care conditions, or they simply do not get care at all.

The patients who do not get care are the ones most likely to present with late-stage disease. At Oatmeal Health, we work every day on the gap between lung cancer detection and lung cancer survival. Stage I survival is 77 percent. Stage IV is 9 percent. The difference between those outcomes is often a care coordinator who was on staff at a community health center - until the hiring freeze in month one.

What This Means For You

For FQHC executives and community health center leaders: Run your three-scenario financial models now. Model Section 330 authorization lapsing at December 2026 and not returning until March 2027. Model Medicaid volume declining 10 to 20 percent in your panel starting mid-2027. Identify which enabling services you can sustain on Medicaid and grant revenue alone - and which require 340B savings to operate. If you are a rural FQHC in a state that received RHT funding, engage your state Medicaid agency now about how health center investment is being prioritized in the application.

For health system administrators and CMOs: FQHC closures redirect medically complex, socially vulnerable patients to your emergency department. Model this volume now. A rural health center closure typically generates a 15 to 30 percent increase in ED utilization from its former patient panel. That is documented from prior closure events, not an estimate. Plan your capacity and care navigation resources accordingly.

For radiologists and pulmonologists: The patients most likely to present with late-stage lung cancer are the same patients who receive primary care at FQHCs. If Section 330 grant cuts eliminate care coordination roles at community health centers, patients who were on track for LDCT lung cancer screening at age 50 will not get screened. The FQHC is often the only access point between a high-risk patient and a scan that saves their life.

For healthcare investors and founders: The FQHC financial crisis creates urgent demand for solutions that reduce administrative burden. Revenue cycle optimization tools, Medicaid eligibility re-enrollment platforms, population health management infrastructure, and telehealth models that allow health centers to serve more patients with fewer staff are addressing real and growing problems. The market need is validated, immediate, and expanding.

For policy advocates: The comment period for the CMS Medicaid managed care proposed rule remains open. Submitting real operating financial data from actual health centers - not projections, but audited financials - is the most effective form of policy engagement at this stage. NACHC coordinates this effort. The July 21, 2026 comment deadline is the next immediate opportunity.

The closing point is this. The rural health investment in the Rural Health Transformation Program is real and in some communities it will make a meaningful difference. States that invest RHT funds in FQHC technology infrastructure and care model transformation are making the right call.

But policy math is not about intentions. It is about net outcomes.

When Section 330 authorization expires at year-end, when Medicaid work requirements strip coverage from millions of patients starting December 31, 2026, and when 340B savings continue to erode - the net outcome for the community health center system is a $7 billion annual revenue hole that a rural health program does not fill.

A bandage on a Medicaid wound does not close the wound. It makes the press release look better.

The CFOs, medical directors, and board chairs running community health centers deserve more than a headline. They deserve a policy environment that takes their financial model seriously - and they deserve health system leaders, investors, and advocates who understand the sequence before the closures start.

What is your December 2026 contingency plan?

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise. Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• NACHC: Proposed Changes in Reconciliation Bill Jeopardizes Care for Community Health Center Patients Insured through Medicaid (nachc.org)

• CMS: CMS Announces $50 Billion in Awards to Strengthen Rural Health in All 50 States (cms.gov)

• KFF: Community Health Center Patients, Financing, and Services - Impact of Federal Policy Changes (kff.org)

• STAT News: The greatest threat to federally qualified health centers may not be federal funding cuts (statnews.com, March 2026)

• Advocates for Community Health: Special Report - The Critical Role of 340B at FQHCs (advocatesforcommunityhealth.org)

The FDA just fast-tracked two AI tools that write radiology reports from scratch - and the legal, liability, and reimbursement frameworks are not ready.

A chest X-ray lands in a radiologist's queue. Before they open it, an AI has already reviewed the image and written a preliminary report. The radiologist reads it, corrects it, and signs. The whole interaction takes 25 seconds.

That is not a thought experiment. That is what Aidoc's First Read system - now holding FDA Breakthrough Device Designation - is designed to do. And a second company, Mosaic Clinical Technologies' Cognita, received the same designation earlier this year. Two generative AI radiology tools. Both with the FDA's fastest review track. Both targeting chest X-rays. Both capable of drafting the actual report text - not just flagging findings.

This is a categorical shift. For the last decade, radiology AI has been about detection: flag the nodule, highlight the hemorrhage, triage the critical case. Now the technology is crossing into interpretation territory. The difference matters enormously for radiologists, health systems, and every attorney, payer, and regulator who has spent years building frameworks around the assumption that AI detects and humans write.

1. What Just Changed at the FDA

The FDA's Breakthrough Device Designation (BDD) is not an approval. It is an accelerated interaction pathway - a commitment from the agency to prioritize review of a device that may offer more effective diagnosis or treatment of a life-threatening or irreversibly debilitating condition. It does not mean First Read or Cognita can be sold and used clinically today. It means they are now in the fast lane toward that clearance.

The designation itself signals something important: the FDA has determined that generative AI radiology drafting is a meaningful enough clinical advance to warrant prioritized attention. The FDA receives thousands of device applications and grants Breakthrough status only to a fraction.

Aidoc's First Read analyzes chest radiographs and generates preliminary high-quality radiology report text. The system produces a draft the radiologist reviews, corrects, and signs. Aidoc is not new to this space - the company has 18 existing FDA clearances and operates at 150+ U.S. health systems. First Read is actually their second Breakthrough Device Designation in under a year. Their CARE Triage tool received the first in September 2025.

Mosaic Clinical Technologies' Cognita took a different path to the same destination. Cognita CXR - a generative vision-language model for radiology - received its BDD in March 2026. Mosaic positioned Cognita as the first radiology generative AI model to receive this designation. Both companies' announcements landed in the same week of June 2026, covered together by STAT News on June 25.

Two companies. Two breakthrough designations. Same week. Same target: the chest X-ray.

2. The Problem These Tools Are Trying to Solve

To understand why this matters, you need to understand how broken radiology workflow has become.

A Neiman Health Policy Institute study found that outpatient imaging interpretation turnaround times more than doubled between 2014 and 2023. Not because radiologists got slower. Because volume grew faster than workforce.

As of 2026, the U.S. faces a projected shortfall of 17,000 to 42,000 radiologists, pathologists, and psychiatrists by 2033. Radiology groups report difficulty recruiting in subspecialties like neuroradiology, breast imaging, and interventional radiology. Burnout is not improving: 53% of radiology professionals cite it as their top concern for the field's future. Rates sit at 46% in private practice and 37.4% in academic settings.

The math is not sustainable. More patients. Fewer radiologists. Higher RVU expectations. Overnight coverage demands. Administrative burden growing every quarter.

In a 2026 Radiology study, AI-prefilled structured chest radiograph reporting dropped interpretation time from 88 seconds to 25 seconds - a 72% reduction. If a radiologist reads 50 chest X-rays per day, that is 73 minutes of reporting time returned per shift. Spread across a department of 10 radiologists, the capacity return approaches 12 hours per day with no new hires.

The question the FDA has now effectively answered: can generative AI do this at a quality level sufficient to enter the Breakthrough review pathway? The answer is yes.

As of March 30, 2026, radiology accounts for 1,163 of 1,524 FDA-cleared AI algorithms - 76.3% of the entire cleared AI landscape. This is not a specialty dipping its toes into AI. It is the most AI-saturated field in medicine. And almost all 1,163 of those tools do the same thing: detect.

3. Detection AI Versus Generative AI: Why the Distinction Matters

This is where most coverage of First Read and Cognita gets muddled. The distinction between detection AI and generative AI in radiology is not a technical curiosity. It has direct implications for liability, reimbursement, validation, and regulation.

Detection AI flags things. It says: there is a nodule in the right lower lobe, 6mm, with spiculated margins. It draws a bounding box and produces a confidence score. The radiologist looks at the finding, agrees or disagrees, and writes the report themselves.

Generative AI writes things. It produces language: 'There is a 6mm spiculated nodule in the right lower lobe suspicious for malignancy. Recommend CT with contrast and pulmonology follow-up.' That language goes into a report carrying the radiologist's name and license.

The difference in how those outputs are regulated, validated, and insured is vast.

Every row in that table represents a framework built around detection AI. None of it maps cleanly onto generative AI. Sensitivity and specificity measure whether a model finds the right things. They do not measure whether the language the model produces is clinically appropriate, complete, and defensible in court.

4. The Liability Problem Nobody Wants to Own

Here is the question that keeps radiologists awake: if a generative AI system drafts a report that misses a finding, and the radiologist signs it without catching the error, who is responsible?

Under current U.S. law, the answer is the radiologist. The physician retains full clinical responsibility when they sign a report. The 2026 CPT updates added Category I codes recognizing AI-augmented diagnostic analysis as a distinct billable service - but those codes explicitly state the AI component supplements, not replaces, the professional interpretation. The radiologist is still the interpreter of record.

But this creates a structural problem. If the AI produces a near-complete draft and the radiologist works through 50 of them per shift in 25 seconds each, the cognitive model changes. The radiologist is no longer authoring a report. They are proofreading one. The vigilance required to catch an AI error embedded in fluent, confident-sounding clinical language is different from the vigilance required to notice a missed finding when writing from scratch.

This is not hypothetical. It is the same cognitive failure mode that appeared in aviation, nuclear, and anesthesia when automation was introduced without recalibrating how humans perform oversight. The AI does most of the work. The human becomes a reviewer. Errors from the AI pass through because the human stopped actively searching for them.

The validation frameworks, liability rules, and malpractice insurance models were built for detection AI. They assume the human writes the report and the AI provides a secondary check. Generative AI inverts that assumption. The regulations have not caught up.

5. What Aidoc and Mosaic Are Getting Right

Both companies have made substantive progress. Aidoc's 18 prior FDA clearances and deployment at 150+ health systems is real infrastructure, not vaporware. Their CARE Triage system is already operating clinically. First Read extends that into a different problem: not which case to prioritize, but what language to put in the report.

Mosaic/Cognita is coming from a different angle. A pure-play generative vision-language model purpose-built for radiology. Their approach: have the model process the entire image holistically and draft findings for radiologist review. Not a triage layer on top of existing clearances - a foundational shift in how the report gets produced.

Both companies made the right first move by seeking Breakthrough Device Designation rather than squeezing through the standard 510(k) pathway. The BDD generates priority interaction with the FDA, which means the agency shapes the validation approach rather than simply reviewing it at submission.

What remains open: Multi-site prospective validation. Both tools are pre-market. The studies needed to support clearance - showing real-world performance across diverse patient populations, imaging equipment types, and clinical settings - have not been completed or published. Reimbursement pathways remain unresolved. CMS has issued no guidance on generative AI-drafted reports. Medical malpractice carriers have not established underwriting frameworks for generative AI-assisted radiology.

Deep Dive

The Automation-Vigilance Tradeoff: What Aviation Learned

The most instructive parallel to generative AI in radiology is not prior radiology AI. It is autopilot in aviation.

When autopilot systems became advanced enough to handle 90% of flight operations, pilot error rates in manual flight situations increased. Not because pilots got worse individually. Because they lost the practice of active manual control. The automation created skilled monitors instead of skilled operators - and skilled monitoring is a different cognitive skill that requires specific training.

Radiology is now at the same inflection point. A radiologist who reviews 50 AI-drafted chest X-ray reports per shift develops a different cognitive muscle than one who writes 50 reports per shift. The oversight skill required - spotting a hallucinated finding, catching an omission in fluent AI prose, knowing when the AI's confident-sounding conclusion does not match the image - is something the profession will need to actively train for.

The aviation industry addressed automation complacency through mandatory manual flying requirements, recurrent simulator training on edge cases, and crew resource management frameworks that specifically targeted automation vigilance failures. Radiology will need analogues: structured training on AI error patterns, review protocols for AI-drafted reports, and documentation standards that show how much a radiologist modified the AI draft before signing.

The Reimbursement Vacuum

There is a clear financial incentive for health systems to adopt generative AI radiology drafting once it is cleared. If the 72% reduction in per-study reporting time holds at scale, the capacity gains are significant - particularly for high-volume community hospitals and outpatient imaging centers operating near capacity.

But the reimbursement model does not yet support this. The 2026 CPT Category I AI-augmented analysis codes apply to detection tools. A tool that drafts the report text does not fit that code structure cleanly. CMS has issued no guidance on generative AI-drafted reports. Health systems adopting these tools before clearance are doing so on their own dime.

This creates a market dynamic where academic medical centers and well-funded health systems adopt early - absorbing cost as an operational efficiency investment - while community hospitals and FQHCs wait for both clearance and a reimbursement framework. The equity gap in AI adoption is likely to widen before it narrows.

The radiology AI market is projected to grow from $989 million in 2026 to approximately $7.17 billion by 2035, expanding at a 24.6% compound annual rate. Generative AI report drafting, once cleared and reimbursed, will likely be one of the largest segments of that growth. The infrastructure question is which health systems are positioned to capture it when that moment arrives.

What This Means For You

• For radiologists and radiology department heads: Start planning now for what 'reviewing an AI draft' means as a distinct cognitive skill. The vigilance mode is different from authoring a report from scratch. Ask your department what training protocols you will use for AI-generated report review before this technology goes live. Do not treat it as a UI change. Treat it as a workflow redesign.

• For health system CMOs and radiology administrators: Monitor the clearance timelines for First Read and Cognita closely. The Breakthrough Device pathway moves faster than standard 510(k). When clearance comes, you will need an adoption plan, a liability review with your malpractice carrier, and a documentation policy before deployment.

• For FQHCs and community health centers: The efficiency gains from generative AI radiology will arrive in your environments last. Plan for a 12-24 month lag between large system adoption and community-accessible deployment, particularly while CPT coding for AI-drafted reports remains unresolved.

• For healthcare investors and founders: Two companies have now established regulatory credibility in generative AI radiology drafting. The question is not whether this category will receive FDA clearance - the BDD pathway strongly suggests it will. The question is who builds the multi-site validation evidence and reimbursement case first.

• For policy advocates: The 2026 CPT AI codes are a start and do not cover generative AI report drafting. Advocacy for CMS guidance on billing, liability safe harbors, and oversight training standards needs to begin now - before clearance, not after. Post-clearance rulemaking on novel AI categories takes years.

The FDA's Breakthrough Device Designations for Aidoc First Read and Cognita are not just product milestones. They are a regulatory acknowledgment that generative AI in radiology is a legitimate clinical advance worth prioritizing. That changes how vendors, investors, and health systems should be planning - and the timeline is shorter than most realize.

But the designation also exposes how far behind the surrounding infrastructure has fallen. Liability rules that assume humans write reports. Reimbursement codes that apply to detection, not drafting. No established training standards for AI-assisted report review. A workforce facing automation-vigilance challenges with no playbook.

The AI read the chest X-ray in 25 seconds. The legal and regulatory system will need considerably longer.

What are you seeing in your radiology department around AI drafting tools? Are health systems you work with already piloting these technologies pre-clearance? Reply - I want to hear what is actually happening on the ground.

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise. Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• Aidoc: 'Aidoc Receives FDA Breakthrough Device Designation for AI That Drafts Radiology Reports' (June 25, 2026) - https://www.aidoc.com/about/news/aidoc-receives-fda-breakthrough-device-designation-for-ai-that-drafts-radiology-reports/

• STAT News: 'FDA gives generative AI in radiology two breakthrough designation nods' (June 25, 2026) - https://www.statnews.com/2026/06/25/radiology-generative-ai-cognita-aidoc-fda-breakthrough-designation/

• Mosaic Clinical Technologies: 'FDA Breakthrough Device Designation for Cognita' (March 5, 2026) - https://mosaicclinical.ai/news/2026/03/mosaic-clinical-technologies-announces-fda-breakthrough-device-designation-for-cognitas-generative-ai-model-for-radiology/

• The Imaging Wire: 'FDA Updates AI List with New Clearances' (March 11, 2026) - https://theimagingwire.com/2026/03/11/numbers-from-the-fda-show-radiology-is-maintaining-its-lead/

• Becker's Hospital Review: 'Radiology in 2026: The Workforce Crisis Meets the AI Revolution' - https://www.beckershospitalreview.com/radiology/radiology-in-2026-the-workforce-crisis-meets-the-ai-revolution/

No press conference. No announcement. Government lawyers disclosed it quietly in a June 22 court filing.

But the implications of that payment are anything but quiet. This is the most significant Medicare Advantage enforcement action in years. And it is a signal that the compliance era for the MA industry has arrived with real teeth.

1. What Actually Happened at Elevance

The story starts in February 2026.

CMS sent Elevance Health a formal notice threatening to suspend new enrollment across the insurer's Medicare Advantage prescription drug contracts. The agency alleged what it called "substantial and persistent noncompliance" with federal rules governing risk adjustment data submission.

Enrollment suspension is not a slap on the wrist. Elevance covers approximately 2 million Medicare Advantage members. Blocking new enrollments mid-year is an existential operational threat. It signals to the market, to members, and to investors that the plan's credibility with CMS is in question.

And here is the part that matters most: the underlying compliance failure was not subtle.

Over the previous seven years, Elevance had been submitting risk adjustment data via flash drives instead of through CMS's required electronic systems. The agency had repeatedly warned against this method. Elevance kept doing it anyway.

CMS had been flagging this methodology problem for seven years before taking formal action.

In response to the February notice, Elevance disclosed that it had set aside $935 million as its best estimate of total financial exposure from the issue. That $935 million figure came from CFO Mark Kaye during the company's first-quarter earnings call on April 22, 2026.

Then on May 27, Elevance sent $342 million as a partial settlement.

Additional tasks remain. CMS has set deadlines through the end of June and into July for Elevance to resolve issues across other risk-adjustment modules and address remaining overpayment concerns. The sanctions clock has not fully stopped.

Elevance is not done with this story. And neither is the industry.

2. Medicare Advantage's Overbilling Problem: How the System Works

To understand why this matters, you need to understand how Medicare Advantage actually gets paid.

In traditional Medicare, the government pays providers directly based on services rendered. In Medicare Advantage, the government pays insurers a monthly capitation rate per enrolled member. That rate is adjusted upward based on how sick the member is. Sicker members generate higher payments.

The tool used to measure sickness is called risk adjustment. It works through diagnosis codes. When a physician documents a diagnosis for a patient, that code gets submitted to the insurer, who submits it to CMS, which uses it to calibrate the capitation payment.

A diabetic patient generates more payment than a healthy one. A diabetic patient with complications generates even more. A patient with multiple chronic conditions generates more still.

Here is where the problem enters.

Insurers have a direct financial incentive to submit as many high-severity diagnosis codes as possible. More codes, especially for more serious conditions, mean higher capitation payments. Conversely, codes that are not supported by actual clinical documentation should not be submitted. But the line between aggressive, good-faith coding and fraudulent upcoding is exactly where enforcement is focused.

CMS estimates 9.5% of all Medicare Advantage payments are improper, driven primarily by unsupported diagnosis codes submitted by MA organizations.

The scale of the problem is staggering.

MedPAC, the independent advisory body that monitors Medicare programs, concluded in its March 2024 report that MA plans were overpaid by approximately $83 billion in 2024 alone. That figure represents 22 percent more than what Medicare would have paid for the same patients in traditional fee-for-service Medicare.

The Committee for a Responsible Federal Budget projects that coding intensity and favorable patient selection will result in MA plans being overpaid by $1.2 trillion over the 2025 to 2034 decade. Half of that, $520 billion, will come directly from the Medicare Hospital Insurance trust fund.

This is not a rounding error. It is a structural flaw in the payment model that has been documented for years and largely tolerated because of the political complexity of reform.

Until now.

3. The Enforcement Era: A Settlement Scoreboard

The Elevance payment is not happening in isolation. It is part of an accelerating pattern.

In January 2026, Kaiser Permanente affiliates agreed to pay $556 million to resolve False Claims Act allegations. That settlement is the largest MA risk adjustment fraud settlement in history. The government alleged Kaiser physicians were pressured to add diagnoses after patient visits through post-hoc medical record addenda, generating approximately $1 billion in additional Medicare payments over a nine-year period.

In 2025 alone, the Department of Justice secured $5.7 billion in False Claims Act settlements and judgments from healthcare cases. The DOJ has publicly stated that Medicare Advantage enforcement is now its number one priority in healthcare.

The OIG issued its first Medicare Advantage-specific compliance program guidance since 1999 on February 2, 2026. That guidance matters because it signals that the regulators view MA compliance as a distinct discipline requiring its own standards, not just a subset of general healthcare compliance.

And CMS is scaling its enforcement infrastructure dramatically. The agency is expanding its team of medical coders from 40 to approximately 2,000 individuals, with the target of auditing every one of the 550 MA contracts annually.

CMS is scaling from 40 to 2,000 medical coders, with a target of auditing all 550 MA contracts per year.

The math on that scaling is significant. At 40 coders, CMS could audit a small fraction of plan submissions. At 2,000, the agency can sustain the kind of systematic review that makes fraudulent submission strategies genuinely risky.

4. What the Denial Data Reveals About the Other Side

While regulators are focused on MA plans submitting false diagnoses to collect inflated payments, a parallel crisis is documented in what happens to patients when they actually try to use their coverage.

The Medicare Rights Center published data in June 2026 showing the denial rates for post-acute care among the three largest Medicare Advantage organizations.

Long-term care hospitals and inpatient rehabilitation facilities serve patients with the most medically complex needs. Stroke survivors. Patients recovering from cardiac events. People who have just had amputations or traumatic brain injuries. These are not elective services.

CVS denied 80 percent of initial requests to enter a long-term care hospital. Humana denied 72 percent. UnitedHealth denied 71 percent.

When patients appealed, 36 percent of long-term care hospital denials were reversed. Forty-three percent of inpatient rehabilitation denials were reversed on appeal.

That reversal rate is critical. It tells you that a substantial portion of the denied claims were medically appropriate. Denials are not being generated by careful clinical review finding genuine ineligibility. They are being generated by a default-to-deny system that counts on patients not knowing how to appeal, being too sick to navigate the process, or simply giving up.

Together, UnitedHealth, Humana, and CVS cover nearly 20 million Medicare Advantage members. The denial policies of these three plans are not statistical outliers. They are the system.

The two-sided nature of the MA accountability crisis is now visible.

On the revenue side: plans submit inflated diagnoses to collect higher payments from CMS.

On the care side: plans deny medically necessary services to reduce the cost of delivering that care.

The business model optimizes both levers simultaneously. Maximize revenue from the government. Minimize payouts to patients and providers. The $342 million settlement is what happens when the first lever finally gets enforced.

5. Deep Dive: The $1.2 Trillion Problem and What Fixes It

The Elevance settlement is a meaningful data point. But at $342 million against a projected $1.2 trillion in MA overpayments over the next decade, it is also clear that individual enforcement actions are not going to close the gap.

The structural issue is the risk adjustment model itself.

CMS has been phasing in a revised risk adjustment model called V28 since 2024. V28 is designed to reduce coding intensity adjustments and more accurately calibrate payments based on actual patient complexity. MA plans vigorously opposed V28, arguing it would destabilize their financial models.

They were partially right, which is why MA enrollment growth decelerated significantly in 2025 and 2026. Several major plans exited markets or scaled back benefits. UnitedHealth lost nearly 647,000 MA members between March 2025 and March 2026. Elevance lost 346,000. The market is repricing against tighter enforcement and tighter payment models simultaneously.

Scenario A: Aggressive Enforcement Only

If enforcement continues at the current trajectory without structural payment reform, the government will collect additional settlements, probably reaching a few billion dollars per year. But the risk adjustment coding incentive remains. Plans will adapt their compliance programs while continuing to optimize code submissions. The overpayment problem moderates but does not resolve.

Scenario B: V28 + Enforcement Combined

The combination of V28 payment reform and active enforcement creates meaningful structural change. CMS's 2,000-coder audit capacity makes systematic coding schemes genuinely risky. V28 reduces the financial reward from coding intensity. This is the current trajectory. Estimates suggest this could reduce overpayments from the projected $1.2 trillion to something closer to $800 billion over the decade, a real improvement but still significant.

Scenario C: Fundamental Risk Adjustment Reform

A redesigned risk adjustment model that relies on prospective clinical data rather than retrospective diagnosis code submissions would remove the post-visit addenda opportunity that produced the Kaiser settlement. Several policy researchers have proposed encounter-based or episode-based payment models. These are politically complex and face intense industry opposition, but they represent the only path to genuinely closing the gap.

V28 payment reforms are projected to reduce MA overpayments by approximately 34 percent compared to the prior model, but the absolute dollar amount remains in the hundreds of billions.

6. What This Means for Providers

The Elevance settlement matters to providers not because of the dollar figure, but because of what it signals about the operating environment.

If CMS is enforcing MA risk adjustment compliance with enrollment suspension threats, plans will respond. They always do. The question is how.

One response is genuine compliance improvement. Plans invest in documentation standards, audit their code submission processes, and stop submitting unsupported diagnoses. This is the best outcome for the system and for patients.

The second response is more aggressive denial behavior on the claims side to offset revenue pressure. If plans can no longer generate the same margins from coding, they may attempt to recover margin by reducing payouts. The denial rate data already shows this pattern at scale.

For FQHCs and safety-net providers, the implications are direct. MA membership is a growing share of patient panels at community health centers. Between 2020 and 2025, MA enrollment grew from 24.1 million to approximately 35 million, now representing 55 percent of all Medicare beneficiaries. Many of those new MA enrollees came from lower-income and higher-complexity populations.

When your MA members face post-acute denials, the FQHC absorbs the referral disruption, the care coordination burden, and often the cost of the patient returning sicker after a premature discharge.

What This Means For You

For FQHC executives and community health leaders:

• Your MA patient panels are growing. Know your payer-specific denial rates for post-acute referrals. Document the clinical basis for every referral at the time of the encounter, not after a denial arrives.

• MA compliance pressure on plans does not reduce your administrative burden. It may increase it as plans tighten prior authorization documentation requirements.

• Build appeal capacity now. The reversal rates on MA post-acute denials (36-43 percent on appeal) mean that fighting denials generates real revenue recovery.

• Engage your state Medicaid agency and advocacy organizations on the overlap between MA enforcement and FQHC reimbursement policy. These conversations are connected.

For health system administrators and CMOs:

• Risk adjustment audit exposure is now shared risk for health systems that operate MA-adjacent billing workflows. Review your documentation standards for diagnosis code submissions.

• The V28 payment model change and enforcement expansion are compressing MA margins simultaneously. Plans that cannot absorb that margin pressure will exit markets or reduce benefits. Model your MA census exposure now.

• Post-acute denial rates at 65-80 percent industry-wide mean that every inpatient discharge to a skilled nursing facility or rehab involves a probable MA authorization fight. Build that into your case management workflow and staffing model.

For healthcare investors and founders:

• MA compliance technology is a growing market. Audit support, documentation quality, and risk adjustment accuracy tools all have tailwind from the enforcement era.

• The plans that survive the enforcement and payment reform transition will be the ones with genuine compliance infrastructure, not just legal reserves.

• The denial-side problem creates opportunity for patient advocacy and appeals automation tools, particularly for post-acute care workflows.

For radiologists and pulmonologists:

• If your imaging practice has MA contracts, review how your diagnosis documentation is submitted through your billing chain. The coding path between your report and the MA plan's risk adjustment submission is often longer and more opaque than you realize.

• Lung cancer diagnosis coding has specific risk adjustment value in MA plans. Ensure your clinical documentation supports the specificity of any diagnosis codes being submitted on your behalf.

The Boldest Truth

Medicare Advantage was marketed as a better version of Medicare. Private-sector efficiency, extra benefits, lower out-of-pocket costs.

What we built instead is a system where private insurers collect a government check sized by how sick their members appear to be, then deny care to those same sick members at rates that would be scandalous if they were printed in a newspaper without context.

An 80 percent denial rate for long-term care hospitals is not utilization management. It is revenue protection dressed in clinical language. And a $342 million settlement for submitting seven years of flash-drive data instead of using the electronic systems CMS required is not just a compliance failure. It is a systemic prioritization of revenue optimization over accountability.

The $342 million Elevance paid to CMS is the beginning of a reckoning, not the end of it. The DOJ has active investigations across the industry. CMS is scaling its audit capacity by 50 times. The OIG issued new compliance guidance for the first time in 27 years.

The enrollment suspension threat worked once. It will be used again.

For every Medicare Advantage plan operating in this environment, the question is not whether this will affect you. It is whether you are positioning your compliance infrastructure to survive it, or hoping the auditors run out of time before they get to your contracts.

They are hiring 2,000 coders. They will not run out of time.

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise.

Key References

• KFF Health News: Medicare Advantage Company Pays $342M to Government in Midst of Billing Probe (June 2026) - https://kffhealthnews.org/medicare/medicare-advantage-cms-elevance-crackdown-overcharging-payment/

• DOJ Office of Public Affairs: Kaiser Permanente Affiliates Pay $556M to Resolve False Claims Act Allegations (January 2026)

• MedPAC: March 2024 Report to Congress - MA overpayments estimated at $83 billion in 2024 (22% above traditional Medicare)

• Medicare Rights Center: Largest MA Organizations Have Disproportionate Denial Rates for Some Post-Acute Care (June 2026)

• Becker's Payer Issues: Elevance sets aside $935M over Medicare Advantage enrollment threat

The coverage wall just cracked.

Starting July 1, 2026, Medicare will do something it has never done in its 61-year history. It will pay for GLP-1 medications prescribed specifically for weight management - not for diabetes, not for cardiovascular risk reduction as a secondary diagnosis, but for obesity itself.

For two decades, federal law explicitly banned Medicare Part D from covering drugs used solely for weight loss. That was not an accident. It was a deliberate policy choice, written into statute and defended through every budget cycle since 2003. On July 1, that choice ends.

The Medicare GLP-1 Bridge Program gives eligible seniors access to Wegovy, Zepbound, and Foundayo at a $50 monthly copay. Medicare negotiated the price down from $1,000 to $1,350 per month to a $245 net per-month supply. The federal government absorbs the difference.

It is a bigger deal than the coverage it provides. It is a signal that Washington now officially treats obesity as a disease worth treating - and that the $1.3 trillion annual cost of cardiovascular disease projected by 2050 is a problem worth disrupting upstream.

But here is the part that should alarm every health system, FQHC, and primary care practice in the country: 82% of Medicare beneficiaries do not know this is happening. The program starts in two days. Most eligible seniors will not walk through the door asking for it. That means the burden falls on you - and your documentation workflow had better be ready.

1. Two Decades of Exclusion, Now Over

When Medicare Part D was created in 2003, Congress included a blanket prohibition on covering drugs used "for anorexia, weight loss, or weight gain." The legislative rationale was that obesity treatments were not proven effective and that covering them would balloon the Part D budget.

That rationale aged badly. The clinical evidence that has accumulated since 2020 makes it indefensible. Semaglutide (Wegovy) produces an average 13.7% reduction in body weight in patients without diabetes. Tirzepatide (Zepbound) produces 20.2% weight reduction. The SELECT trial - 17,604 adults with obesity but no diabetes - demonstrated that Wegovy reduces major adverse cardiovascular events by 20%.

These are not lifestyle drugs. They are disease-modifying agents with peer-reviewed outcomes data published in the New England Journal of Medicine. The FDA cleared Wegovy for secondary cardiovascular risk reduction. Zepbound received clearance for obstructive sleep apnea. These are the exact populations that fill Medicare rolls.

The Treat and Reduce Obesity Act (TROA) passed through the House Committee on Ways and Means with bipartisan support and has made legislative progress toward removing the permanent statutory exclusion. The Bridge Program is CMS's immediate bridge: an 18-month demonstration running July 1, 2026 through December 31, 2027, using a demonstration authority that sidesteps the Part D prohibition without waiting for Congress to finish the job.

The mechanics are deliberately simple. The Bridge operates entirely outside the standard Part D benefit structure. Part D plans carry zero risk for these prescriptions. Humana - already running the Limited Income Newly Eligible Transition (LI NET) program - serves as the central processor for claims adjudication and pharmacy payment. Patients use their Medicare card, pay $50, and the pharmacy gets reimbursed by a parallel track.

For providers, the operational implication is critical: your patients do not need to switch plans, file appeals, or navigate a new insurance portal. They walk into the pharmacy with their existing Part D card and a valid prior authorization. Which brings us to the piece of this program that is most likely to generate denials.

2. Who Qualifies: The Clinical Matrix and Its Gaps

Eligibility for the Medicare GLP-1 Bridge is specific, tiered, and documentation-dependent. There are three pathways:

Pathway 1 - BMI of 35 or greater (no comorbidity required). This is the broadest pathway. If your patient is a Medicare Part D enrollee with a BMI at or above 35, they qualify on BMI alone.

Pathway 2 - BMI of 30 or greater with qualifying comorbidity. Qualifying conditions are: heart failure, uncontrolled hypertension, or chronic kidney disease stage 3a or higher. A patient with a BMI of 31 and a CKD diagnosis qualifies. A patient with a BMI of 31 and controlled hypertension does not.

Pathway 3 - BMI of 27 or greater with high-risk cardiovascular diagnosis. Qualifying conditions are: prediabetes, prior myocardial infarction, prior stroke, or symptomatic peripheral artery disease. This pathway has the most clinical specificity and the most room for documentation error.

The eligibility language contains real traps. "Uncontrolled hypertension" is not defined by a specific blood pressure threshold in the Bridge guidance - but payers will look for documentation supporting that characterization. Prediabetes must be documented, not just inferred from A1C. And BMI must be in the chart at the time of the prescription.

One more requirement that many providers have missed: patients must be enrolled in a lifestyle modification program focusing on diet and physical activity. This is a condition of certification. If a provider submits a prior authorization without documenting that the patient is participating in a lifestyle program, the authorization may be denied or pended.

3. The Prior Authorization Problem: Retrospective, Not Prospective

The Bridge Program uses retrospective prior authorization. Providers prescribe the medication first. The documentation, diagnosis codes, BMI measurement, and comorbidity verification are submitted after the prescription is written.

On its surface, this sounds patient-friendly - no waiting for approval before the patient starts treatment. In practice, it creates a different problem. Retrospective PA means clinics that do not front-load their documentation will face post-hoc denials. The pharmacy will dispense the medication. The patient will take it. And then a denial arrives in the mail weeks later, triggering retroactive cost recovery from either the pharmacy or the patient.

The prior authorization form is submitted electronically or by fax to Humana's central processor. CMS has published a standardized form. Every field matters. The system has no tolerance for ambiguity in comorbidity documentation.

📊 CMS has certified that providers must submit: BMI documentation from a clinical encounter within the prior 12 months; diagnosis codes for all qualifying comorbidities; attestation that the patient is enrolled in a lifestyle modification program; and a prescription from a licensed prescriber. Missing any of these components creates a denial pathway even if the patient is clinically eligible.

4. The 82% Problem: Who Is Not Going to Show Up

Here is the most important number in this story: 82%.

That is the share of Medicare beneficiaries who, according to a June 2026 survey by the Obesity Care Advocacy Network, are unaware that Medicare is about to begin covering GLP-1 medications for weight management. The survey covered more than 2,100 adults ages 65 and older, and the findings held across party lines - 79% of Republicans and 84% of Democrats were equally uninformed.

CMS deliberately limited its public outreach ahead of July 1. The agency's reasoning: beneficiaries are "most moved to take action" when a benefit is actually available, not when it is announced in advance. That approach may be defensible from a behavioral economics standpoint. It is problematic from an access equity standpoint.

More than 69% of Medicare beneficiaries are overweight or obese - 35% classified as overweight, 34% as obese. Among those classified as obese, 28% already have Type 2 diabetes. CBO estimates that 29 million Medicare beneficiaries would qualify for weight management GLP-1 coverage under an expanded policy. The Bridge program does not reach all 29 million - the eligibility criteria are more restrictive - but the potential eligible population remains enormous.

The patients who will show up on July 2 asking for this benefit are the patients who read health news, have engaged physicians who flag it proactively, or are being cared for at systems that built a workflow for it. The patients who will not show up are the patients in underserved communities, rural counties, and safety-net clinics, whose providers are stretched thin and whose outreach infrastructure is minimal.

📊 Only 9% of adults ages 65 and older are currently using GLP-1 medications. That number does not move without proactive provider identification of eligible patients.

5. The Racial and Equity Gap Inside the Bridge

Obesity is not evenly distributed in America. Black Americans have the highest obesity prevalence of any racial group, followed by Hispanic adults. The populations most burdened by the disease that this program treats are also the least likely, historically, to receive GLP-1 prescriptions.

A 2026 study published in Obesity found that non-Hispanic white patients were the most likely to be prescribed a long-acting GLP-1 at 2.4%, compared to 2.3% for non-Hispanic Black patients, 1.8% for Hispanic patients, and 1.7% for Asian patients. These differences appear small in absolute percentage terms, but they represent hundreds of thousands of patients across the Medicare population.

The disparities are not random. They reflect differential access to obesity medicine specialists, primary care capacity, insurance coverage history, and provider bias. A 2026 analysis found that provider bias and weight stigma mean physicians are statistically less likely to discuss weight-loss treatments with patients from lower-income or minority backgrounds - even when clinical criteria are met.

There is evidence that policy interventions can close this gap. Research on state Medicaid policies that expanded GLP-1 coverage found that racial and ethnic prescription disparities decreased significantly after coverage expanded. The Bridge Program, by reducing cost to $50 per month, removes the largest single barrier. But cost is not the only barrier.

For FQHCs serving predominantly Black and Hispanic Medicare patients, the Bridge Program is an immediate clinical and operational opportunity. The question is whether those clinics have the bandwidth and the workflow to execute.

6. The Business Case for Health Systems

The Medicare GLP-1 Bridge creates real downstream financial implications for health systems beyond the drug cost itself.

CBO projects that expanded GLP-1 coverage for Medicare beneficiaries could save $3.4 billion in healthcare costs through 2034 by reducing heart attacks, strokes, diabetes complications, and other obesity-related conditions. The mechanism is straightforward: treat obesity now, avoid expensive acute events later.

For health systems managing large Medicare populations, this is a portfolio bet on reducing acute utilization. A patient on Wegovy for 18 months who avoids one hospitalization for a cardiac event saves the system far more than the drug cost.

The financial model for a mid-size health system illustrates the stakes.

Note: Estimated at $3,400 per patient in downstream cardiovascular savings based on CBO projections normalized to the 2026 eligible population. These are illustrative ranges, not actuarial guarantees.

7. Deep Dive: The Drug Landscape and What Providers Need to Know Today

The Medicare GLP-1 Bridge covers three specific products beginning July 1. Understanding the distinctions matters for prescribing decisions.

Foundayo (tirzepatide, Eli Lilly) - an oral tirzepatide formulation, cleared specifically for weight management. The first oral GLP-1/GIP dual agonist available in the Bridge. Patients who cannot tolerate injections now have an oral option.

Wegovy (semaglutide, Novo Nordisk) - available in injection and tablet formulations. The SELECT trial cardiovascular outcomes data supports its use in patients with established cardiovascular disease and obesity without diabetes. The oral formulation (Rybelsus re-branded as Wegovy tablet for weight management) is now included.

Zepbound KwikPen (tirzepatide, Eli Lilly) - the injectable formulation covered under the Bridge. SURMOUNT-OSA Phase 3 data showed significant reduction in apnea-hypopnea events, making this particularly relevant for Medicare patients with comorbid sleep apnea.

What is NOT covered under the Bridge:

Ozempic (semaglutide for diabetes) and Mounjaro (tirzepatide for diabetes) remain outside the Bridge because they carry diabetes indications, not weight management indications. If a patient has diabetes and obesity, the diabetes indication governs and the Bridge does not apply. Those patients use existing Part D diabetes drug coverage.

This distinction will cause confusion at the point of prescribing. Providers need to verify which formulation and indication they are prescribing under before submission.

The Lifestyle Program Requirement

The requirement that patients be enrolled in a lifestyle modification program is real and enforceable. CMS defines a qualifying program as one that focuses on dietary change and physical activity, supervised or supported by a licensed healthcare professional. A printed handout is not sufficient documentation. A documented referral to a formal program or a structured internal clinic protocol with defined touchpoints is.

For FQHCs, this is actually an opportunity. Many already operate diabetes prevention programs and behavioral health integration models that qualify as lifestyle modification programs under this definition. Documenting the Bridge lifestyle requirement alongside an existing DPP enrollment can create a clean audit trail.

Prior Authorization: The Exact Steps

1. Patient visit: Document BMI in the chart from this encounter or within prior 12 months. 2. Confirm qualifying pathway and document diagnosis codes. 3. Document lifestyle program enrollment. 4. Generate prescription for qualifying Bridge drug. 5. Submit PA request electronically or by fax to the Humana central processor using the CMS-issued form. 6. Humana adjudicates, notifies prescriber and pharmacy. 7. Patient fills prescription at any participating pharmacy for $50.

The retrospective nature means the PA form can be submitted after the prescription is written, but it should ideally be submitted the same day to minimize any gap between dispensing and authorization.

8. What This Means for the Permanent Coverage Question

The Bridge is temporary. December 31, 2027 is the hard end date. What happens after that is not settled.

The BALANCE Model - CMS's broader demonstration architecture that was supposed to include a permanent Part D weight management component - has had its Part D component delayed indefinitely. The Bridge itself was extended through December 2027 precisely because the BALANCE Model was not ready.

The permanent solution requires legislation. The Treat and Reduce Obesity Act (TROA) provides the statutory pathway by removing the longstanding Part D exclusion language. With bipartisan support in committee and the clinical evidence now undeniable, legislative passage is more likely than at any point in the program's history.

But "more likely than ever" and "certain" are different. If the Bridge ends in December 2027 without TROA passage, patients who started Wegovy or Zepbound under the Bridge face an abrupt coverage cliff. The cardiovascular and metabolic benefits of these medications are dependent on sustained use - stopping creates rebound weight gain and risk restoration. A coverage cliff for 29 million potentially eligible beneficiaries is not a policy nuance. It is a public health event.

Note: Scores reflect analyst consensus on likelihood and health system impact from Georgetown and STOP Obesity Alliance policy analyses, not CMS projections.

What This Means for You

If you are an FQHC or community health center CEO: Your Medicare patient panel is the first place the Bridge creates real opportunity and real risk. Identify your patients with BMI 30+ and qualifying comorbidities now, before they show up without a workflow to receive them. Stand up your lifestyle program documentation today. The July 1 start date is not a planning horizon. It is a deadline.

If you are a health system CMO or Chief Medical Officer: The Bridge is an 18-month window to demonstrate that proactive obesity management reduces acute utilization. Build the identification and outreach protocol now. This is the kind of downstream savings demonstration that payers and boards pay attention to. Do not wait for patients to self-identify.

If you are a primary care physician: Get the CMS prior authorization form from the Medicare GLP-1 Bridge provider portal before you write your first Bridge prescription. The retrospective PA process is forgiving if you have the form ready. It is expensive if you do not. Verify which drug you are prescribing under - Bridge indications and standard Part D indications are not interchangeable.

If you are a radiologist or pulmonologist: Your patients with severe obesity and sleep apnea (Zepbound's additional indication) are some of the highest-BMI patients in your panel. They are also among the most likely to qualify under Pathway 2 or 3. Adding Bridge eligibility screening to your pre-procedure documentation workflow adds minimal burden for potentially significant patient benefit.

If you are a healthcare investor or founder: The Bridge is a 18-month forced experiment in what happens when you dramatically lower the cost of GLP-1s for Medicare seniors. The utilization data, outcomes data, and cost savings data generated through December 2027 will be the most powerful evidence base the obesity medication market has ever had. Watch it carefully. Whatever the Bridge proves will shape permanent coverage policy for decades.

If you are a policy advocate: The clock is ticking on TROA. The Bridge buys time but does not fix the statute. Every month that passes without a permanent legislative solution is a month closer to a coverage cliff that will hit the patients who can least absorb it.

Closing

There is a sentence worth sitting with today.

For 61 years, Medicare did not treat obesity as a disease worth treating. The evidence was weak, the drugs were limited, and the policy consensus held. Then the clinical evidence became undeniable. The drugs became transformative. And Congress moved slowly anyway.

The Medicare GLP-1 Bridge is what happens when an administration decides not to wait. It is imperfect - time-limited, eligibility-restricted, documentation-dependent, and under-publicized. It will not reach everyone who needs it. The 82% awareness gap ensures that the patients with the least access will get the least benefit.

But the coverage wall cracked. And the physics of policy change are such that cracks tend to widen.

The real question for every health system, FQHC, and primary care practice in America is not whether this program changes the landscape. It is whether your organization is positioned to get your patients through the opening before it closes - or before it becomes permanent.

Are your documentation workflows built for July 1?

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise. Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• Medicare GLP-1 Bridge Program Overview, CMS

• GLP-1 Bridge Demonstration Begins July 2026, Medicare Rights Center

• Seniors Unaware of Medicare Obesity Drug Coverage, CNBC

• Disparities in Prescribing of Long-Acting GLP-1s, Obesity (Wiley 2026)

• Cardiovascular Statistics 2026, Journal of the American College of Cardiology

Here is what $1.6 billion in AI investment bought the American healthcare system in 2025: a tool that listens to physician conversations and writes the note for them.

The early data is genuinely good. A study of 263 physicians across six health systems found burnout dropped from 51.9% to 38.8% after 30 days with an ambient AI scribe. Documentation time fell by more than 15%. Physicians report feeling less trapped by their keyboards. That is real. That matters.

But here is what the press releases do not say: 62% of physicians still reported burnout in 2025 according to Medscape. The 13-point improvement is meaningful, but it is happening against a baseline so broken that two-thirds of the physician workforce is still experiencing clinical exhaustion even with the tools in place.

The ambient AI scribe is a good answer to a symptom. Health systems are treating it as a cure for a disease it does not touch.

1. The Ambient AI Boom Is Real - and Genuinely Promising

Let us be precise about what ambient AI scribes actually accomplish, because the technology is not the problem.

The tools - Abridge, Nuance DAX, Freed AI, and Epic's native "Art for Clinicians" - listen passively to physician-patient conversations and generate structured clinical notes. No dictation. No keyboard. The physician reviews and signs. Visit time shifts from note-typing back to the patient in front of them.

The adoption curve is steep. By 2025, 62.6% of Epic hospitals were live with generative AI tools, and 71% of physician practice leaders reported using AI for patient visits. Investors poured $1.6 billion into ambient AI companies in 2025 alone. Abridge raised $550 million across two rounds and achieved a $5.3 billion valuation. Epic built its own competing tool. Nuance - already inside Microsoft - accelerated deployment.

The market is not hype. The clinical evidence is emerging and it is directionally positive.

📊 A 2025 randomized study found ambient AI users spent 8.5% less total time in the EHR and 15% less time composing notes specifically.

The problem is not the technology. The problem is the narrative layered on top of it.

2. What the Data Shows - and What It Quietly Omits

Every ambient AI study shows improvement on documentation metrics. The burnout studies show partial improvement. Both are real. Neither tells the whole story.

When Medscape surveyed physicians in 2025 and asked them to name their top burnout drivers, "too much bureaucratic work" came in first at 62%. "EHR-related tasks" came second at 55%. Those two dominate the rankings and they are the exact things ambient AI addresses.

But look further down the list. "Too many hours worked" at 41%. "Lack of respect from administrators" at 37%. "Prior authorization burden" at 32%. "Loss of autonomy" at 24%.

These are structural causes. They predate EHRs. They will not respond to a better microphone in the exam room.

The chart below shows how physician burnout drivers stack up - and which ones ambient AI actually addresses:

A physician who spends less time on notes is still receiving 77 EHR inbox messages per day - a figure documented in a 2025 multi-site study. They are still completing 43 prior authorizations every week, burning more than 16 hours on forms, hold music, and appeals that have nothing to do with patient care. They are still working within a 15-minute visit slot that RVU production targets have compressed over decades.

The ambient AI scribe does not touch any of those problems. And the studies that show burnout reduction from AI scribes are, almost by design, measuring the thing the tool addresses while leaving the rest of the iceberg unmeasured.

3. The RVU Trap That AI Cannot Fix

To understand why AI scribes are insufficient, you need to understand what actually compressed physician time in the first place.

The Relative Value Unit system - how Medicare and most payers compensate physicians - assigns point values to clinical activities. A basic 15-minute office visit (CPT code 99213) carries approximately 1.30 wRVUs. A complex 45-minute visit (99215) carries 2.80 wRVUs. The math heavily incentivizes volume: more shorter visits generate more wRVUs than fewer longer ones.

Health systems, facing relentless margin pressure, set productivity targets based on these benchmarks. The 2026 MGMA benchmark for Family Medicine sits at 5,033 median wRVUs per year. Primary care physicians are effectively scheduled to produce that number - which means scheduling templates designed to maximize E&M visit throughput, defaulting to 15-minute slots.

📊 From 2023 to 2024, median total encounters rose 7.81% for primary care physicians even as median wRVUs fell - meaning they are seeing more patients for less compensation per encounter, while clinical complexity continues to rise.

The ambient AI scribe does not change this math. A physician who uses an AI scribe can write the note faster. The note is cleaner. The review takes less time. But the 15-minute slot is still 15 minutes. The next patient is still waiting. The RVU target is still 5,033. The production pressure is structural. The tool is cosmetic.

If a health system implements ambient AI and then uses the recovered time to squeeze one more patient visit into the schedule, the tool has not reduced burnout. It has enabled a new layer of throughput pressure.

This is not hypothetical. It is the predictable consequence of deploying efficiency tools into an incentive structure that absorbs every efficiency gain and converts it into more volume.

The table below shows precisely where AI scribes move the needle - and where they do not:

4. The 77-Message Inbox That Ambient AI Ignores

There is a specific physician workflow failure that almost never makes it into the ambient AI marketing materials: the EHR inbox.

Physicians receive an average of 77 EHR inbox messages per day. These are not spam. They are patient portal messages, lab results requiring review and action, refill requests, administrative flags, specialist notes, and care coordination tasks. Each one requires physician judgment. None of them can be delegated to an AI scribe that listens to exam room conversations.

Primary care physicians average 2.7 hours of after-hours EHR work daily - uncompensated time done after clinic ends. This is the inbox catching up to them after a full day of patient visits. It is a second shift performed for free, invisible in RVU calculations and scheduling templates.

Ambient AI scribes have zero impact on this. The tool records the encounter. The note writes itself. The physician goes home - and opens the inbox.

📊 Clinicians using ambient AI tools spend 8.5% less total time in the EHR per encounter, but the inbox burden runs on its own unmanaged cycle entirely separate from documentation.

The solutions to inbox overload are not AI scribes. They are staff-supported inbox triage - having MAs or nurses handle refills, stable lab review, and routing before it reaches the physician - team-based care models, visit preparation protocols that front-load information collection, and EHR design that reduces inbox generation in the first place. These require staffing investment and organizational redesign. Not a new software subscription.

5. Prior Authorization: The 16-Hour-a-Week Tax

If inbox overload is the quiet burnout driver nobody talks about, prior authorization is the loud one nobody fixes.

Physicians complete an average of 43 prior authorizations per week, dedicating more than 16 hours to the process - forms, phone calls, hold queues, peer-to-peer reviews, and denials that must be appealed. This is equivalent to 40% of a physician's working week consumed by an administrative process designed to reduce payer cost by delaying or denying care.

26% of physicians reported that prior authorization delays directly harmed their patients in a 2025 AMA survey. The harms include cancer diagnoses delayed, imaging not performed, medications not started.

AI is making inroads here. AI-generated prior authorization letters show strong clinical content. CMS's CMS-0057-F rule required payers to build real-time API connections by January 2026, reducing some information asymmetry. Companies like Waystar, Infinitus, and Cohere Health are deploying AI against the prior authorization workflow.

But ambient AI scribes - the category receiving $1.6 billion in investment and generating the burnout headlines - do not address prior authorization. They listen to patient encounters. They do not negotiate with UnitedHealth.

📊 Physicians spend more than 16 hours per week on prior authorization - the equivalent of a full-time administrative role layered on top of clinical practice.

The prior authorization crisis is a policy problem masquerading as a workflow problem. The only real solutions are legislative limits on PA requirements, automatic approval for high-evidence standard-of-care treatments, and payer-side AI that can process requests in minutes rather than days. Health systems cannot solve this with a subscription to Abridge.

Below is a scorecard of who the current AI scribe boom is actually helping:

Deep Dive: The Financial Case for Full Workflow Redesign

The health systems actually reducing burnout - not just documentation time - share a common pattern. They are not just deploying AI scribes. They are using AI as the first step in a broader workflow redesign that addresses all five burnout drivers simultaneously.

What Comprehensive Redesign Includes

Step 1: AI Scribes (Documentation Layer)

Deploy ambient documentation. Recover 45 to 60 minutes per physician per day. Use this recovery to establish the organizational case for further change - not to add more visits.

Step 2: Inbox Triage Infrastructure (Staffing Layer)

Staff inbox triage teams of MAs or licensed staff per physician panel. Standardize refill protocols that do not require physician review. Route stable lab results to a patient notification workflow. Target: reduce physician inbox volume from 77/day to 30/day.

Step 3: Pre-Visit Preparation (Scheduling Layer)

30 minutes before each visit, a care coordinator reviews the chart, flags open orders, closes care gaps, updates medication lists, and queues priority issues. The physician walks in prepared. The 15-minute slot becomes productive instead of frantic.

Step 4: Prior Authorization Automation (Revenue Cycle Layer)

Deploy AI-powered PA tools to auto-submit standard-of-care requests with clinical documentation packages. Target: reduce PA hours from 16+ per week to 6 per week per physician.

Step 5: Schedule Redesign (Throughput Model Layer)

Extend complex visit slots. Reserve buffer time for urgent add-ons. Build explicit team huddle time into the daily schedule. This requires leadership authorization to reduce raw RVU production targets - which requires the financial model to show equivalent or higher margin from reduced turnover costs.

The financial argument is compelling. The chart below shows the math:

The average cost to replace a physician is $750,000 to $1 million when recruitment, onboarding, and locum coverage are factored in. A health system losing 10% of its physician workforce annually to burnout is spending $7.5M to $10M per 100 physicians in replacement cost alone. Full workflow redesign for a 100-physician practice - AI scribes, inbox triage staff, scheduling changes - costs approximately $1.5M to $2M annually. The ROI is not close.

The ambient AI scribe is the cheapest line in the redesign budget. It is also the most visible. Health systems are stopping there because the other steps are harder. They require staffing investment, scheduling authority changes, and a willingness to reduce short-term throughput in exchange for long-term physician retention.

What This Means For You

• FQHC executives and community health center leaders: Ambient AI scribes are arriving at your doorstep as a burnout solution. They will help - but you are also facing the highest documentation burden, the thinnest staffing ratios, and the least access to inbox triage resources. Do not implement an AI scribe and call burnout solved. Budget for the staffing infrastructure that makes the tool effective: at least one MA per 1.5 physicians for inbox support and pre-visit prep.

• Health system administrators and CMOs: You have deployed the AI scribe. Now do the harder work. Run a physician time audit across your top five burnout drivers - documentation, inbox, prior auth, visit slots, and administrative intrusions on clinical autonomy. AI scribes only close the first gap. Build the roadmap for all five.

• Radiologists and pulmonologists: Your burnout drivers differ from primary care but the pattern is the same. AI tools are arriving to reduce reading time and report generation burden. Ask your department leadership what is being done about prior authorization delays for CT and MRI, about the overnight call coverage model, and about reading volume targets. Those are the conversations that matter.

• Healthcare investors and founders: The ambient AI scribe market is not finished - but the next phase is integration, not documentation. The companies that win in 2027 are not building better microphones. They are building integrated workflow platforms that connect ambient documentation to inbox triage to PA automation to scheduling optimization. Single-feature AI tools are table stakes.

• Policy advocates: Prior authorization is a federal policy problem. CMS-0057-F is a start, but automatic approval for standard-of-care treatments and real-time PA processing requirements for high-acuity conditions need to be in the next reauthorization package. Physician burnout at this scale is a public health crisis. The solution cannot be a better voice recorder.

Closing

The ambient AI scribe is not a distraction. It is a genuine tool delivering genuine value. A 13-point reduction in burnout among users is not nothing - in a workforce where 62% are burned out, that is tens of thousands of physicians who feel meaningfully better about their workday.

But American health systems have a habit of adopting the visible technology solution and declaring the underlying problem solved. We did it with EHRs in 2010. We told ourselves that digitizing records would transform care coordination. Fifteen years later, those same EHRs are the second-leading cause of physician burnout.

The ambient AI scribe is the right first step. Health system leaders need to say out loud that it is only the first step - and then fund the rest of the plan.

The physicians who are quietly updating their LinkedIn profiles right now are not leaving because they hate taking notes. They are leaving because the system surrounding the note-taking has not been redesigned in twenty years.

What would it take for your health system to actually fix the machine - not just tune it?

Reply and tell me. I read every response.

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts.

Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise.

Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• Multi-site ambient AI scribe study (263 physicians, 6 health systems, Abridge/PMC 2025): burnout reduction from 51.9% to 38.8% after 30 days

• Medscape Physician Burnout and Depression Report 2025: 62% of physicians reporting burnout, top-driver survey data

• 2025 multi-site study: physicians average 77 EHR inbox messages per day, 2.7 hours daily after-hours EHR work

• AMA Prior Authorization Physician Survey 2025: 43 PA requests per week, 16+ hours consumed, 26% reporting patient harm

• PitchBook/FierceHealthcare: $1.6 billion invested in ambient AI companies in 2025; Abridge $550M raised at $5.3B valuation

• MGMA 2026 Benchmarks: Family Medicine median 5,033 wRVU; primary care encounter and productivity trends 2023-2024

Clinical decision support was supposed to be the kill switch for preventable medical errors. Embed the evidence base directly into the physician's workflow. Surface the right recommendation at the right moment. Save lives.

Instead, we built a noise machine.

A 2025 analysis published in the Journal of the American Medical Informatics Association found that physicians override clinical decision support alerts at rates exceeding 90 percent in some health systems. That number has barely moved in two decades despite billions of dollars in investment, two major federal pushes for EHR adoption, and a generation of increasingly sophisticated AI models.

The algorithm is not the problem. The relationship between AI and the physician is.

Health systems have spent years asking the wrong question. They asked: how do we make the model more accurate? The right question is: how do we make the clinician actually trust it?

1. The Scale of the Problem

Before diagnosing the fix, it helps to understand how deep the dysfunction goes.

Clinical decision support tools now touch nearly every acute-care encounter in the United States. From drug-drug interaction alerts in the CPOE system, to sepsis early warning flags in the ICU, to imaging appropriateness criteria embedded in ordering workflows, CDS has become the connective tissue of digital medicine. The global market is valued at $6.36 billion in 2025 and projected to reach $15.32 billion by 2033 at an 11.8 percent compound annual growth rate. North America accounts for more than 46 percent of global revenue.

That is not a niche technology. It is infrastructure.

And it is failing at its primary job.

A systematic review published in 2026 synthesized data on alert override rates across health systems and found a range of 49 to 96 percent. Read that again: in the worst-performing systems, nearly every alert a physician receives gets ignored. In the best-performing systems, nearly half of all alerts are dismissed without action.

📊 The average primary care physician receives 63 interruptive EHR alerts per day, according to Epic's own internal research.

Think about what that means operationally. A physician working a 10-hour shift receives six or seven alerts per hour, every hour. Most are low-signal: a drug interaction that is clinically irrelevant for this patient, a preventive care reminder that the patient already declined, a flag for a lab value the physician has already addressed. The legitimate, high-acuity alerts - the ones that could change outcomes - are buried in that stream.

The American Medical Association's 2025 Physician Work Environment Survey added another layer to the problem. Among physicians who identified alert fatigue as a significant workplace issue, 62 percent also reported at least one symptom of burnout. The AMA now categorizes alert fatigue as the leading structural driver of burnout in the current physician work environment. Not workload. Not pay. Not administrative burden in the abstract. The specific, daily grind of being interrupted by a system that cries wolf.

The early CPOE systems at Brigham and Women's Hospital in Boston - widely considered the gold standard for early clinical decision support implementation - still struggled with override rates above 85 percent for drug-drug interaction alerts in the early 2000s. That was before the iPhone. Before deep learning. Before the era of AI-powered everything. We have added two decades of computational sophistication and the override number barely moved.

This is not a technology problem. It is a design and culture problem.

2. Alert Fatigue: Structural, Not Behavioral

The healthcare industry's dominant narrative about CDS failure is behavioral. Clinicians are too busy. They are arrogant. They think they know better than the algorithm. They have bad habits from training that predate these tools.

That framing is wrong, and it is costing patients.

Alert fatigue is structural. It is the predictable, rational outcome of a design philosophy that treats every potential clinical concern as equally worthy of interruption.

📊 Vanderbilt University Medical Center reduced alert volume by 54 percent in a targeted intervention and saw clinically significant alert response rates nearly double.

That result is not surprising to anyone who has thought carefully about information architecture. More signals do not mean more safety. In an already saturated cognitive environment, more signals mean more noise, which means more dismissal, which means the actual safety signal gets missed more reliably.

The core design failure is the conflation of documentation with intervention. Health systems have used CDS tools to satisfy regulatory compliance requirements, quality metrics, and malpractice documentation needs. The result is a system that generates alerts to prove the system noticed something - not to actually change clinician behavior. The alert fires. It is overridden. The override is logged. The quality committee reviews override rates. The cycle repeats.

There is a second structural failure: the lack of feedback loops. In most health systems, there is no mechanism by which a clinician's alert-override pattern feeds back into the system's configuration. A physician who overrides a specific drug interaction alert 300 times in a year is not generating data that triggers recalibration of that alert's threshold. The system keeps firing the same alert. The physician keeps ignoring it. The data accumulates, unused.

📊 In most health systems, fewer than 10 percent of CDS tools have any active recalibration process based on override patterns, per the 2026 FDA CDS guidance commentary.

Geisinger Health System demonstrated what happens when you break this pattern. By narrowing its sepsis AI alerts to a single high-confidence threshold - instead of firing at every statistical inflection point - and pairing each alert with a single one-click intervention, Geisinger achieved meaningful clinician behavior change. Specificity combined with action. That is the structural formula that works.

The lesson is not that AI needs to be smarter. The lesson is that AI needs to be quieter, more precise, and directly connected to something the physician can do.

3. The Black-Box Problem: Why Physicians Don't Trust What They Can't See

There is a third dimension to the trust problem that does not get enough attention in the vendor conversation: physicians are trained to understand causality, and most clinical AI does not explain itself.

Show a hospitalist at Cleveland Clinic or Mayo Clinic an alert that reads "high sepsis risk" with a confidence score and no underlying reasoning, and they will dismiss it. Not because they are arrogant. Because the clinical training that made them skilled diagnosticians also made them deeply skeptical of conclusions without supporting logic.

This is the black-box problem. And it is getting worse as AI models become more complex.

A meta-analysis published in August 2025 synthesized findings from 62 peer-reviewed studies on explainable AI in clinical decision support, published between 2018 and 2025. The analysis covered applications in radiology, oncology, neurology, and critical care. Its core finding: explainability improves clinician trust, but only when the explanation is delivered at the right level of detail for the clinical context. A six-paragraph explanation of a sepsis risk score is not better than no explanation. A two-sentence explanation tied to the specific lab values that drove the flag is.

The meta-analysis also identified a persistent gap: fewer than 30 percent of evaluated CDS systems provided any clinician-facing explanation of the model's reasoning. The majority produced a score, a risk category, or a recommendation - with no supporting logic visible to the clinician.

📊 Meta-analysis of 62 XAI studies (2018-2025): fewer than 30 percent of clinical AI tools provided clinician-facing explanations of model reasoning.

The regulatory environment is shifting to acknowledge this. On January 6, 2026, the FDA released updated guidance on Clinical Decision Support Software that explicitly placed stronger emphasis on usability and the appropriateness of information presentation for healthcare professionals. The guidance specifically flags information overload as a design failure - not a clinician behavior failure - and requires that CDS functions presenting to healthcare professionals "prioritize decision-relevant details."

That language is meaningful. It signals a shift from treating accuracy as the primary regulatory concern toward treating usability and trust as co-equal requirements. A clinical AI tool that is 95 percent accurate but produces outputs physicians cannot interpret or act on does not meet the regulatory intent of the 2026 guidance.

4. What Actually Works: Evidence from the Field

The clinical AI adoption literature contains a clear, consistent signal that the industry has been slow to hear: specificity drives adoption.

Broad-scope CDS tools - the kind that attempt to cover every possible clinical concern across a patient encounter - consistently underperform. Narrow-scope tools, precisely calibrated to a single high-stakes clinical decision with a clear action attached, consistently outperform.

Vanderbilt's intervention is the clearest example in the published literature. By targeting alert reduction specifically, the health system cut daily interruptions by more than half and nearly doubled the rate at which physicians acted on the alerts that remained. The technology did not change. The intervention was purely about signal-to-noise ratio.

Geisinger's sepsis example illustrates the second factor: action proximity. An alert that requires a physician to navigate three EHR screens to act on it performs worse than an alert that offers a one-click intervention in the same interface where the alert surfaces. The friction between the AI's recommendation and the physician's next action is a direct predictor of adoption.

The 2026 Doximity State of AI in Medicine Report, which surveyed 3,151 U.S. physicians in early 2026, added a third factor: trust transfers from colleagues, not vendors. Physicians who adopted AI tools at significantly higher rates were more likely to have heard a recommendation from a clinical peer than from a vendor presentation or administrative mandate. This finding is consistent with broader healthcare change management literature but has been largely ignored in CDS deployment strategy.

📊 Doximity 2026 AI in Medicine Report: physicians who adopted AI at high rates were more likely to report peer recommendation as the primary driver than vendor or administrator recommendation.

The market is also segmenting on this dimension. A new generation of CDS vendors is deliberately building narrow-scope tools with embedded explanations and one-click action interfaces. These companies are growing faster than their broad-scope competitors. Health systems that prioritize vendor selection based on explainability and workflow integration are starting to report measurably different adoption outcomes.

But the dominant pattern in health system procurement still runs the opposite direction. Procurement committees evaluate CDS tools primarily on model accuracy, regulatory compliance posture, and EHR integration compatibility. Usability, explainability, and workflow proximity are secondary criteria or not evaluated at all.

That procurement bias is one of the most significant and least-discussed contributors to the adoption gap.

5. The Regulatory Moment: FDA's 2026 CDS Guidance and What It Changes

The FDA's January 6, 2026 update to its Clinical Decision Support Software Guidance is the most significant regulatory shift in the CDS space in five years. And it has not received nearly enough attention from the clinical informatics community.

The 2026 guidance makes three changes that will reshape the market:

First, it clarifies which CDS functions remain outside FDA regulation. If a tool provides a single, clinically appropriate recommendation based on EHR text and clinical guidelines, and the output is clearly reviewable by the clinician, it may remain a Non-Device CDS - exempt from the 510(k) and De Novo pathways. This gives a substantial class of AI-assisted CDS tools a faster path to deployment without pre-market submission requirements.

Second, the guidance explicitly raises the standard for information presentation. CDS tools classified as Non-Device must now demonstrate that their outputs "prioritize decision-relevant details and avoid information overload." This is a usability requirement, not just an accuracy requirement. Vendors who cannot demonstrate that their tools reduce rather than increase cognitive load now face a classification risk: a tool that generates excessive alerts or unexplained outputs may be reclassified as a medical device requiring pre-market review.

Third, the guidance reaffirms the existing Non-Device status of CDS used in CMS's Appropriate Use Criteria Program - the imaging referral appropriateness framework that requires physicians to consult AUC-qualified tools before ordering advanced imaging for Medicare patients. This is directly relevant to any health system operating an imaging program with Medicare volume.

📊 FDA January 2026 CDS Guidance: Non-Device CDS status now requires demonstration that outputs prioritize decision-relevant details and avoid information overload - a usability requirement, not just an accuracy benchmark.

Deep Dive: The Economics of Alert Fatigue - What Health Systems Are Actually Losing

The clinical case against alert fatigue is clear. The financial case is less often articulated, but it is equally compelling.

Alert fatigue has direct revenue consequences that health system CFOs have not yet learned to measure.

Consider the missed sepsis detection problem first. Sepsis is the most expensive condition to treat in U.S. hospitals, averaging $38,000 per hospitalization and accounting for more than $62 billion in annual hospital costs. A sepsis early warning tool that achieves 90 percent override rates is not a neutral investment. It is a tool that is failing to prevent high-cost, high-mortality events at roughly nine times the rate it is succeeding. Every override of a true-positive sepsis alert is a potential $38,000 missed intervention - to say nothing of the mortality and liability exposure.

The medication error problem has a similar structure. Drug-drug interaction alerts exist because adverse drug events (ADEs) cause approximately 1.3 million emergency department visits per year and cost the U.S. healthcare system an estimated $30 billion annually. A DDI alert system with a 90 percent override rate is catching perhaps 10 percent of the interactions it was designed to intercept. The other 90 percent reach the patient.

📊 Adverse drug events cost the U.S. healthcare system $30 billion annually. A CDS DDI alert system with a 90 percent override rate intercepts approximately 10 percent of its intended targets.

Then there is the clinician time cost. If a primary care physician receives 63 interruptive alerts per day and spends an average of 3 seconds acknowledging and dismissing each one, that is 3 minutes of pure cognitive interruption per hour, across every clinical hour of the workday. Extrapolated across a 250-physician group practice, that equals more than 125 physician-hours per day in alert-processing overhead. At a fully-loaded physician cost of $150 per hour, that is $18,750 per day, $4.7 million per year, in value-subtracted alert management.

These figures are modeled estimates based on published cost data applied to a representative health system. They are not precise predictions. But they illustrate the order of magnitude of the problem - and they explain why a well-implemented, narrow-scope CDS tool with meaningful adoption rates produces a return on investment that broad-scope, high-volume tools cannot match.

The right benchmark for a CDS investment is not model accuracy. It is the fraction of true-positive alerts that actually change clinical behavior.

Scenario A: Status Quo CDS Deployment

A 300-bed regional hospital deploys a broad-scope CDS platform across all inpatient units. Override rate settles at 88 percent within 90 days, consistent with published benchmarks. The hospital continues to see sepsis mortality at rates comparable to pre-deployment. IT team logs override data that never reaches clinical informatics leadership. Contract renews at year three based on integration stability, not clinical outcomes. True cost: $1.2 million annual subscription, plus $4 million in physician time overhead, plus unmeasured adverse event costs.

Scenario B: Narrow-Scope, Explainable CDS Deployment

A 300-bed regional hospital deploys a narrow-scope sepsis and DDI CDS platform with embedded explainability and one-click actions. Override rate at 90 days: 44 percent on sepsis alerts, 61 percent on DDI alerts. Meaningful behavior change achieved on high-confidence sepsis flags. Within 18 months, the hospital documents measurable reduction in sepsis-related length of stay. IT team uses override pattern data to recalibrate alert thresholds quarterly. Contract renews at year three based on outcomes data.

The difference between Scenario A and Scenario B is not the model. It is the design philosophy.

What This Means For You

The trust problem in clinical decision support is solvable. But it requires different behavior from every stakeholder in the system.

• FQHC executives and community health center leaders: Your clinical staff is already operating at high cognitive load. Every new CDS tool you add without removing existing noise is a net negative. Before deploying a new AI tool, audit your current alert burden per clinician. Cut volume before adding capability.

• Health system administrators and CMOs: Rewrite your CDS procurement criteria. Model accuracy should be a baseline requirement, not the primary evaluation metric. Add usability, explainability quality, and alert reduction targets as contract deliverables. Require vendors to demonstrate adoption rates in comparable health systems before signing.

• Radiologists and pulmonologists: The AI tools entering your workflow are not all equivalent. Demand tools that show you the reasoning, not just the score. If a vendor cannot explain why their model flagged a finding, that is a clinical safety gap, not a minor usability preference.

• Healthcare investors and founders: The CDS market's next growth cycle will belong to companies that solve the adoption problem, not the accuracy problem. Prioritize companies with published adoption rates and outcomes data. Discount companies that lead with model benchmarks and don't mention clinical utilization.

• Policy advocates: The FDA's 2026 CDS guidance is a starting point. Push for CMS to incorporate CDS adoption rates and override metrics into quality program requirements. A tool that achieves 90 percent override rates is not delivering value, and it should not receive the same regulatory treatment as one achieving 40 percent.

The Path Forward

The clinical AI industry is at a pivot point.

For 20 years, the field optimized for accuracy. Build a better model. Increase sensitivity. Reduce false negatives. The result is a generation of tools that are statistically impressive and clinically ignored.

The next 10 years will belong to companies and health systems that optimize for trust. Tools that explain themselves. Alerts that are few and precise. Workflows that connect the recommendation to the action in a single interface.

The physician is not the obstacle to clinical AI. The physician is the user. And the user's experience has been, systematically, terrible.

That is a design failure. It can be fixed.

The data is clear. The FDA guidance is moving in the right direction. The market is beginning to segment on usability. Health systems that act now - that audit their alert environments, retrain their procurement criteria, and demand explainability from their CDS vendors - will achieve measurably better clinical outcomes than those still buying broad-scope tools on the strength of an AUC curve.

What does your current CDS override rate look like? If you are a clinical informatics leader and you don't know the answer, that is the first problem to solve.

About the Author

Jonathan Govette is the Co-Founder and CEO of Oatmeal Health, an AI lung cancer diagnostic company catching cancers earlier in the communities that need it most. Oatmeal uses AI to identify unscreened high-risk patients, navigate them to care, and score every lung CT for malignancy risk - billed under CPT 0721T. Stage I survival is 77%. Stage IV is 9%. We work in FQHCs because that gap is largest there.

Jonathan writes daily about radiology, pulmonology, AI diagnostics, health policy, hospital operations, and healthcare startups.

Subscribe to stay ahead of healthcare's most important shifts. Weekly deep-dives on AI, radiology, health policy, FQHCs, and the business of healthcare - written for operators, clinicians, and investors who want the signal, not the noise. Subscribe at oatmealhealthjonathangovette.substack.com

Key References

• American Medical Association. 2025 Physician Work Environment Survey. Alert fatigue identified as leading structural driver of burnout.

• Journal of the American Medical Informatics Association (JAMIA). 2025 analysis on CDS override rates exceeding 90% in some health systems.

• MDPI Healthcare. Meta-Analysis of Explainable AI in Clinical Decision Support Systems. 62 peer-reviewed studies, 2018-2025. August 2025.

• FDA. Updated Guidance on Clinical Decision Support Software. January 6, 2026. Revised emphasis on usability and information overload prevention.

• Doximity. 2026 State of AI in Medicine Report. Survey of 3,151 U.S. physicians on AI adoption patterns and peer influence factors.

We are building AI on top of a security disaster.

That is not a provocation. That is the operational reality for most health systems in America in 2026.

Healthcare is the most attacked industry in America. Not finance. Not energy. Not critical infrastructure in the traditional sense. Healthcare. The same industry responsible for your most sensitive personal data, your treatment decisions, and in emergency settings, your survival has been the top ransomware target for five consecutive years, according to the FBI's Internet Crime Complaint Center.

And in the middle of this ongoing crisis, health systems are racing to connect more systems, deploy more AI tools, and push more patient data across more endpoints than ever before.

Every new connection is also a new door.

In 2024, the U.S. healthcare sector recorded 739 major breaches affecting over 276 million patient records. In 2025, that number rose to 772 large breaches. In Q1 2026, the pace has not slowed. The average time to identify and co…

Read more

Here is what the healthcare technology industry does not want you to think about.

Three in four American hospitals have deployed some form of artificial intelligence. Two in three physicians report using a health AI tool in 2024, up from two in five just a year before. Health systems poured an estimated $45 billion into AI and analytics tools in 2025 alone. The demos are impressive. The vendor case studies are persuasive. The board presentations have dashboards.

And yet, when researchers look at what is actually happening at the point of care - at the bedside, in the exam room, in the radiology reading suite - a very different picture emerges.

Fewer than 20 percent of health systems report sustained, high-success use of AI in core clinical diagnosis. Alert override rates for clinical decision support tools run between 90 and 96 percent. Becker's Hospital Review published an analysis calling 95 percent of healthcare AI projects a failure - not because the models performed poorly, but beca…

Read more

The Medicaid cuts nobody is talking about will gut community health.

Right now, across America, federally qualified health centers are running the numbers. The reconciliation bill is signed. The cuts are law. And the question is no longer whether FQHCs will feel it. The question is how many will survive it.

The One Big Beautiful Bill Act (H.R. 1), signed into law on July 4, 2025, represents the largest restructuring of Medicaid in decades. Over ten years, the law is projected to cut $911 billion from federal Medicaid spending. The Congressional Budget Office estimates that 11.8 million Americans will lose Medicaid coverage as a result. Work requirements, more frequent eligibility redeterminations, and the end of enhanced federal matching rates combine to create a revenue cliff for the safety-net providers who can least afford it.

For FQHCs, the math is punishing. Medicaid accounts for 43% of community health center operating revenue nationwide. More than 16 million of the 34 million pati…

Read more

Every year, more than 75 million CT scans are performed in the United States. Millions of those scans pass directly over the heart. The coronary arteries. The aortic valve. Calcium deposits that predict the next major cardiac event with more precision than the Framingham Risk Score, the pooled cohort equations, or any blood test currently on the market.

For decades, that data sat unused. Radiologists read the lungs. They noted the bones. They flagged incidental masses. But the calcium in the coronary arteries, a finding visible on every non-gated chest CT, was routinely dismissed because there was no billing code to support reporting it.

That changed on April 1, 2026.

CMS created HCPCS code G0680, a dedicated reimbursement pathway for AI-powered analysis of coronary artery calcium (CAC) and aortic valve calcification (AVC) on routine chest CT scans. It is the first time the federal government has formally recognized opportunistic cardiovascular screening from CT imaging as a billable Med…

Read more

Rural hospitals are finally getting radiologists. Just not human ones.

Sixty million Americans live in communities where recruiting a staff radiologist is nearly impossible. Retention is worse. And when a patient rolls in with a possible stroke or a suspicious lung nodule at 2 in the morning, the wait for a diagnostic read can stretch 6 to 12 hours - or longer. For stroke patients, every minute without a read costs neurons. For lung cancer caught at Stage IV instead of Stage I, it costs lives.

That gap is closing in 2026. AI-powered diagnostic imaging platforms are deploying at critical access hospitals at an accelerating pace this year. These tools are not replacing radiologists. They are doing something arguably more important: filling the diagnostic void in the 1,300+ communities across the United States where no radiologist was ever coming.

This is not a technology story. It is a health equity story - and the numbers behind it are stark enough to demand attention from every health sy…

Read more

The ritual happens every year. CMS publishes its proposed Medicare Physician Fee Schedule for the coming year. Medical associations sound the alarm. Congress sometimes acts, sometimes doesn't. And providers absorb another year of payments that don't keep pace with what it actually costs to run a clinic.

This year's proposed rule for CY 2027 follows the same script. CMS has proposed a conversion factor that, without the expiring one-time 2.5% boost from the One Big Beautiful Bill Act, represents another effective pay cut in real terms for most physicians. Primary care. Radiology. Community health clinics. Independent imaging centers. All of them are looking at 2027 and running the same math: revenue flat or down, costs up 4-6%, margins shrinking.

This is not a new story. But the cumulative weight of it is. Since 2001, Medicare physician payments have fallen approximately 29% in inflation-adjusted terms. Practice costs per FTE physician rose more than 63% from 2013 to 2022. The conversion…

Read more