The question every reader asks after seeing the evidence in Part 1 is: “Why don’t we just fix it?” Why not raise Medicaid reimbursement? Why not fix EHRs? Why not stop the claim denials? The answer, after years of tracing every link in this chain, is that U.S. healthcare operates inside a political-economic system that, through lobbying, regulatory capture, corporate consolidation, political cycling, and institutional lock-in, resists exactly these fixes. The entities profiting from the $5.3 trillion system have spent decades constructing barriers to reform that have proven difficult for legislative action to overcome.

The $751 Million Lobbying Wall

The healthcare industry is the single largest lobbying sector in the United States, and it has held that position for over two decades. The health sector spent about $750.8 million on federal lobbying in 2024, according to OpenSecrets data analyzed in JAMA Health Forum, the only sector exceeding $700 million. The pharmaceuticals and health products industry alone spent $387.5 million, more than any single industry in any sector. Health services and HMOs added $118 million. Over the past two decades, the health sector has spent billions of dollars on federal lobbying alone, not including state-level lobbying, campaign contributions, or dark money spending.

The spending is bipartisan by design. PhRMA, AMA, and AHA all ranked among the top 10 overall lobbying spenders in the first half of 2024 and contribute to candidates on both sides of the aisle. UnitedHealth Group alone reported $12.54 million in federal lobbying for 2025, up from $7.52 million for all of 2024, according to lobbying disclosures. News reporting has noted that large insurers’ lobbying rosters routinely include former congressional staff, although the specific headcounts circulated publicly vary by source and year; readers should consult OpenSecrets registrant-level disclosures for current named staff.

Throughout 2024, the pharmaceutical, hospital, and insurance industries lobbied heavily against several reforms under debate. Some measures were narrowed, delayed, or dropped; others were ultimately included. Readers should consult the final December 2024 continuing resolution and subsequent reconciliation bills for the current status of PBM reform, site-neutral payment reform, and Medicare physician payment fixes, as those provisions have shifted repeatedly. The No UPCODE Act (S.1105), targeting Medicare Advantage overbilling, had not advanced in the 2025 reconciliation process as of mid-2025. Physicians received another year of Medicare payment cuts.

Sources: OpenSecrets (2024-2025 lobbying totals), via JAMA Health Forum (2026); Roll Call (2024); Becker’s Hospital Review (2025) on the No UPCODE Act; Axios, American Prospect, Star Tribune, Washington Post reporting on 2024 end-of-year legislative activity

The Private-Public Symbiosis: Why the Spending Is About Preserving the Split

Part 1 of this paper laid out the argument in detail. Private insurers, whose economics structurally disfavor enrolling lower-income and chronically ill patients, depend on government programs to absorb exactly those populations. Medicaid, CHIP, DSH payments, FQHC grants, Disproportionate Share Hospital funding, Medicare’s traditional fee-for-service system, and the EMTALA-funded emergency department safety net together cover the people the private market is least profitable serving. This is not a moral judgment. It is a description of a stable equilibrium that every major actor has an economic incentive to preserve.

That equilibrium is what the lobbying spend defends. Every dollar spent to block Medicaid parity with Medicare defends the access gap that routes Medicaid patients into FQHCs at higher per-visit payment rates, where the federal government pays. Every dollar spent to preserve the current Medicare Advantage risk-adjustment methodology defends the $43 billion that MedPAC estimates coding intensity added to Medicare Advantage payments in 2023, which federal taxpayers cover. Every dollar spent to block prior-authorization reform defends the denial machine that lets private insurers extract revenue retention at provider expense. The lobbying is not, principally, about blocking universal coverage. It is about preserving the specific split between what the private system covers profitably and what the public system absorbs at a loss.

Any reform strategy that does not understand this symbiosis underestimates the resistance it will face. A reform that would genuinely reduce private insurer revenue (universal Medicare expansion, a strong public option, national prior-authorization reform with teeth) will trigger coordinated industry opposition on a scale the last four decades of American politics have repeatedly demonstrated. A reform that would genuinely reduce federal healthcare spending by forcing the private system to cover the populations it now offloads would trigger the same response. This is why the most durable reforms in recent history have been the ones that increased total spending without altering the split. The equilibrium is resilient because every powerful actor profits from some part of it.

The implication for any serious reform proposal is unambiguous. If the goal is to bend the cascade this paper describes, the reform must operate at a point the lobbying apparatus cannot easily contest. That point is the encounter note. A provider cannot be lobbied out of using a tool that writes a better note. A biller cannot be lobbied out of filing an auto-generated appeal. An independent practice cannot be lobbied out of reopening its Medicaid panel when the arithmetic finally works. The political economy makes the encounter-note intervention attractive precisely because it is the one reform path that does not require the private-public split to be legislatively renegotiated.

Sources: OpenSecrets federal lobbying data; MedPAC (March 2024, ch. 12) and HHS OIG on Medicare Advantage coding intensity; MACPAC supplemental payment and FQHC workaround analysis; Part 1 of this white paper on the wealth-health-insurance sorting machine

The Corporate Profit Motive: $5.3 Trillion Is Someone’s Revenue

The fundamental economic reality is that every dollar of waste in U.S. healthcare is someone’s income. The $812 billion in administrative costs (Himmelstein, Woolhandler & Campbell, using 2017 data), the $55.6 billion (in 2008 dollars) in medical liability system costs including defensive medicine (Mello et al.), the $32 billion that UnitedHealth Group estimates could be saved on avoidable ER visits, and the tens of billions of dollars in FQHC and supplemental payment workarounds all represent revenue streams for corporations with powerful incentives and resources to preserve them.

Industry-advocacy analyses indicate the seven largest health insurers’ combined revenue roughly tripled in a decade, from approximately $511 billion in 2014 to $1.52 trillion in 2024, per EPIC for America’s compilation of insurer annual reports. UnitedHealth Group’s revenue reached approximately $400 billion in 2024, roughly 4.6 times its 2009 level of about $87 billion. Over a longer span, 2001 to 2025, physician Medicare payment fell approximately 33 percent in real terms, according to the AMA. Roy et al. (JAMA Internal Medicine, 2025) documented $2.6 trillion in cumulative shareholder payouts (dividends plus stock buybacks) from publicly traded U.S. healthcare companies from 2001 to 2022, rising from $54 billion in 2001 to $170 billion in 2022, during a period of rising physician burnout, declining access, and worsening outcomes.

The Partnership for America’s Health Care Future (PAHCF), an alliance of hospital, insurance, and pharmaceutical interests, formed in 2018 with a stated objective of preventing single-payer healthcare, Medicare expansion, or a public option. Leaked planning documents describe objectives to “change the conversation around Medicare for All” and “minimize the potential for this option in health care from becoming part of a national political party’s platform in 2020.” Its members spent a combined $143 million on lobbying in 2018 alone, according to Center for Responsive Politics data reported by BenefitsPro.

Sources: EPIC for America (December 2025, secondary compilation of insurer annual reports); Roy et al., JAMA Internal Medicine (2025); AMA (2025) on Medicare physician payment, 2001 to 2025; Mello et al., Health Affairs (2010); UnitedHealth Group on avoidable ER visits; The Intercept; BenefitsPro, citing the Center for Responsive Politics; Jacobin (2024)

Regulatory Capture: The Regulators Often Come From and Return to the Regulated

A Health Affairs study (2023) documented the revolving door at HHS. Among people appointed between 2004 and 2020, roughly 15 percent came directly from private industry and 32 percent exited to industry after their tenure. The greatest net exits were from the CDC and CMS, the two agencies most central to healthcare regulation. More than half of appointees at CMS, CDC, and the Office of the Deputy Secretary went to private industry after government employment. CMS Administrator Seema Verma went to Oracle. CMS Administrator Marilyn Tavenner came from 25 years at for-profit hospital giant HCA. The pattern is bipartisan and persistent.

The AMA’s Relative Value Scale Update Committee (RUC), which effectively sets Medicare physician payment rates, has been described as perhaps the most consequential example of regulatory capture in healthcare. CMS accepts RUC recommendations roughly 90 percent of the time, and the committee has systematically favored procedure-heavy specialties over primary care, contributing to the primary care crisis that drives the cascade.

Sources: Health Affairs (2023); Healthcare Dive (2023); Milbank Memorial Fund (2022)

Political Cycling: Healthcare Reform’s Short Half-Life

The history of healthcare reform includes many ambitious proposals that were killed or narrowed by similar forces: Truman’s universal health insurance (1948, opposed by the AMA), Clinton’s Health Security Act (1993, defeated after a sustained industry-wide opposition campaign widely estimated to have cost hundreds of millions of dollars across industry coalitions, including the Health Insurance Association of America’s “Harry and Louise” advertising), the ACA’s public option (2009, dropped amid intense insurance industry pressure), and Medicare for All (2019-2020, opposed by PAHCF despite majority public support in polls and more than 100 Congressional co-sponsors).

The most instructive case is the ACA’s Medicaid primary care fee bump. In 2013 and 2014, the ACA temporarily raised Medicaid primary care rates to Medicare parity. The evidence was unambiguous: appointment availability increased by 7.7 percentage points (Polsky et al.), and a later study of the fee cuts that followed found that each $10 change in fees moved appointment availability by about 1.7 percentage points (Candon et al.). Then 34 states let the increase expire in January 2015, which the Urban Institute had projected would cut Medicaid primary care fees by an average of 42.8 percent. The policy demonstrated that the fix worked. Then it was allowed to expire.

Sources: PMC / Obama JAMA article (2016); KFF (2025); Urban Institute; Polsky et al., NEJM (2015); Candon et al., JAMA Internal Medicine (2018)

The Complexity Barrier

Perhaps the most insidious barrier is epistemological. The healthcare system is so complex that most people, including many policymakers, have difficulty identifying the root cause. Current political debates focus on drug pricing (approximately 10 percent of total spending), PBM reform (which CBO has scored at roughly $0.7 to $3 billion in savings over ten years per bill), and hospital consolidation (real but downstream of practice economics). None addresses the encounter note.

The CPT code set changes every year; the 2025 edition alone carried 420 updates. The billing and documentation requirements are so complex that administrative errors (including incorrect coding, missing information, and eligibility issues) account for over 80 percent of preventable claim denials across specialties. This complexity is not an unfortunate byproduct. It functions as a mechanism through which administrative costs flow disproportionately to providers while payers automate denials; ProPublica reported that Cigna’s medical directors signed off on denials in batches averaging 1.2 seconds per case, a characterization Cigna disputes.

When someone says “just raise Medicaid reimbursement,” they are proposing a policy that must survive the Senate filibuster (60 votes), overcome hospital, FQHC, and insurer lobbying, sustain funding across multiple administrations, and survive CMS implementation by revolving-door officials. When someone says “just fix the EHRs,” they are proposing replacing $35 billion in HITECH-subsidized infrastructure against vendor lobbying. When someone says “just stop the claim denials,” they are proposing reducing insurer revenue by billions against an industry that has successfully blocked prior authorization reform for decades. Each “simple fix” triggers opposition from multiple powerful industries at once.

Sources: Health Affairs; CAQH Administrative Cost Report; CBO scoring data; AMA (CPT 2025); ProPublica on Cigna claim review

The $35 Billion Subsidy That Built Proprietary Silos

The HITECH Act of 2009 drove EHR adoption with incentive payments that reached $34.7 billion by May 2016, according to CMS. Hospital adoption of even a basic EHR stood at 9 percent in 2008 and is now nearly universal; physician adoption of a basic system stood at 17 percent in 2008. But the program’s fatal design flaw was mandating that providers digitize records without specifying how systems should be built. No unified architecture. No enforceable data standards. No interoperability requirements. Billions in public funds effectively subsidized the entrenchment of proprietary, incompatible systems.

An HHS Office of Inspector General audit (A-05-14-00047, June 2017) estimated that $729.4 million in Meaningful Use incentive payments, about 12 percent of those payments, went to professionals who did not meet program requirements. Academic research has also suggested that the HITECH subsidies accelerated EHR adoption by only a modest margin above what market forces would have achieved on their own.

The Repository That Was Never Mandated

The deeper design failure was structural. HITECH did not mandate a single longitudinal repository of patient records. It did not require EHR vendors to interoperate. It did not specify who would hold a patient’s unified chart, how to address the same patient across systems, or what must happen when a patient crossed from one organization’s EHR to another’s. The consensus decision at the time was to let competitive markets produce interoperability. They did not. It took the 21st Century Cures Act of 2016, and its final rule in 2020, for the federal government to formally name information blocking as something providers and vendors could not do, and enforcement has lagged well behind complaints filed.

The operating condition this produced is not abstract. Consider a clinical arrangement that is typical rather than exotic. A private clinic runs its own EHR. The same providers round at two local hospitals that are both owned by the same Healthcare company and both run Cerner, but each hospital maintains its own siloed Cerner instance that does not share data with the other. The same providers also see patients at a partnered FQHC that runs a different EHR again. Four systems. The same patient can appear in all four. A treating clinician at any one site, at the point of care, sees only the fraction of that patient’s record that lives in the system they happen to be logged into. There is no unified place to go look.

Walk a single patient through the workflow and the danger becomes concrete. The patient is seen in the clinic and told to go to the ED. They are admitted by the hospitalist at Hospital A, whose EHR instance does not automatically pull in the clinic note. The patient clinically deteriorates and is transferred to Hospital B for a higher level of care. The transferring hospital is responsible for sending the chart with the patient. Sometimes it arrives. Often it does not, or arrives only partially, or arrives as a fax the receiving clinician will not read because the patient is already in front of them with an urgent presentation. The receiving clinician, in practice, is working from what the patient can tell them, which has been measured in peer-reviewed studies as incomplete or inaccurate in a material share of cases. This is not a corner-case of American medicine. It is a standard operating condition.

Interoperability Remains a Promise, Not a Reality

Despite more than a decade and a half of regulatory effort, ONC survey data has repeatedly found that a minority of independent hospitals exchange data routinely, and that clinicians rate outside-organization integration as one of the weakest aspects of their EHR experience. Patient data fragmentation is severe: Medicare beneficiaries’ records are commonly spread across multiple EHR vendors over a patient’s lifetime, and patient matching accuracy drops sharply across organizations (widely reported at roughly 50 percent in cross-system exchanges). TEFCA took seven years from authorization to launch and remains voluntary. Enforcement actions for information blocking have lagged far behind complaints filed.

ONC Data Brief 71 (May 2024) reported that only 42 percent of non-federal acute care hospitals send patient health information to most or all other hospitals, and only 38 percent send to most or all ambulatory providers outside their own organization. Those are the rates more than a decade into HITECH and four years into the Cures Act final rule. The other side of the exchange matters just as much: even when a record is sent, the receiving clinician has to find it, trust the patient match, and read it in time to act on it.

What Fragmentation Costs in Lives and Limbs

The safety literature on fragmented records is consistent and difficult to read dispassionately. The Joint Commission estimated in 2010 that roughly 80 percent of serious medical errors involve, at least in part, communication failures during handoffs. AHRQ’s Patient Safety Network has cited figures in the same range, with analyses pointing to communication breakdowns during transitions of care as a contributing factor in a large majority of preventable adverse events. Makary and Daniel estimated in BMJ (2016) that medical error was associated with roughly 251,000 deaths per year in the United States; later analyses have debated the point estimate but not the underlying finding that the count is large enough to place medical error among the leading causes of death.

Interhospital transfer is the sharp end of the problem. A frequently cited study by Usher et al. (Journal of Critical Care, 2016) examined patients transferred into tertiary ICUs and reported an adverse event rate of approximately 42 percent within the first 24 hours of arrival, in-hospital mortality in the range of 17 percent for the transferred cohort, and documentation completeness below 60 percent for records arriving with the patient. Medication reconciliation studies have documented frequent errors on admission, with undocumented allergies and omitted home medications being the most common categories. Patient self-report, which is what the receiving clinician falls back on when the chart does not arrive, has been measured in peer-reviewed work as an imperfect match for the record, for both medications and allergies. This is not dishonesty. It is memory, stress, language, cognitive status, and the simple difficulty of listing every medication a complex patient takes.

Malpractice data is the downstream mirror. CRICO Strategies reported in 2015 that approximately 30 percent of malpractice claims filed from 2009 to 2013 involved communication failures, with associated costs of $1.7 billion and 1,744 deaths. Candello’s more recent analyses, through 2024-2025, have reported the communication-failure share of claims closer to 40 percent, suggesting the problem has not improved and may be worsening as care becomes more distributed and time-pressured.

None of these statistics live only on a dashboard. They are what the workflow described above produces, every day, in every market where patients legitimately receive care across more than one organization.

Epic’s Inescapable Gravity

Epic controls a plurality of U.S. acute care hospitals (about 42 percent of acute care hospitals and about 55 percent of beds at the end of 2024, per KLAS) and generated an estimated $5.7 billion in revenue in 2024. Large implementations routinely exceed $1 billion, including Mass General Brigham at approximately $1.2 billion and Mayo Clinic at approximately $1.5 billion (figures per public reporting from the respective health systems). Antitrust complaints filed in 2025 against Epic describe the EHR platform as one that is, in practice, extremely difficult for hospital customers to leave once implemented, reflecting high switching costs rather than a genuinely competitive marketplace. Physician satisfaction scores on EHR usability remain low in many surveys, yet customer retention remains very high because migration away from an entrenched EHR is prohibitively expensive.

The VA’s Oracle / Cerner implementation is a definitive cautionary tale. The VA’s current lifecycle cost estimate is $37.2 billion (September 2025), and an independent estimate by the Institute for Defense Analyses, cited by GAO, put the cost at $49.8 billion over 28 years. When deployment was paused in April 2023, only 5 of 171 VA medical centers had gone live, after roughly $5 billion in spending.

The Note Is the Real Problem, Not the EHR

U.S. clinical notes are about four times as long as notes in other countries using the same EHR software (Downing et al., 2018). U.S. clinicians spend 90.2 minutes per day actively using the EHR, versus 59.1 minutes for non-U.S. clinicians (approximately 53 percent more time) using the same systems (Holmgren, Downing et al., 2020). This suggests the dysfunction is not a pure technology problem. It is a documentation-for-reimbursement problem rooted in the U.S. payment architecture.

Wang et al. found that only 18 percent of progress note text was manually entered by physicians, with 46 percent copied from prior notes and 36 percent imported from other structured sources. Rule et al. found median outpatient note redundancy reached 58.8 percent by 2018. AI ambient scribes, in their current form, largely automate the data entry layer without challenging the billing architecture that demands voluminous documentation.

Sources: CMS (2016) HITECH incentive payment data; ONC EHR adoption data; HHS OIG audit A-05-14-00047 (2017); 21st Century Cures Act (2016) and Final Rule (2020); ONC (2023); ONC Data Brief 71 (May 2024); KLAS Research; Epic Research (2023); GAO-25-106874 and House Veterans’ Affairs Committee (2023) on the VA EHR program; Joint Commission (2010) on handoff communication; AHRQ Patient Safety Network; Makary and Daniel, BMJ (2016); Usher et al., Journal of Critical Care (2016); medication reconciliation literature; CRICO Strategies (2015) and Candello (2024-2025); Downing et al. (2018); Holmgren, Downing et al. (2020); Wang et al., JAMA Internal Medicine (2017); Rule et al., JAMA Network Open (2021)

AI ambient scribes, among the fastest-adopted technologies in healthcare history, rest on a premise that we believe is flawed: that converting spoken words to written notes solves the documentation crisis. The evidence from a growing body of peer-reviewed studies suggests otherwise.

Hallucinations, Omissions, and Fabrications

In one evaluation, 31 percent of AI-generated ambient notes contained at least one detected hallucination versus 20 percent of physician-written notes (Palm et al., Frontiers in Artificial Intelligence, 2025). A large manual evaluation (Asgari et al., npj Digital Medicine, 2025) found a hallucination rate of 1.47 percent at the sentence level, with 44 percent classified as “major,” meaning they could impact patient diagnosis or management. The most dangerous hallucinations appeared in the Plan section (21 percent), the section containing direct instructions for patient care.

An OHSU study using ChatGPT-4 found an average of 23.6 errors per clinical case. Only 52.9 percent of data elements were correctly reported in all three replicates of the same encounter, suggesting that AI scribe output can be fundamentally unreproducible. OpenAI’s Whisper model, used through Nabla by more than 30,000 clinicians at 40 health systems, was found by Koenecke et al. (2024) to fabricate sentences in approximately 1.4 percent of transcriptions tested (187 of more than 13,000 audio snippets), including violent and racially charged fabrications triggered by speech pauses.

The Editing Burden: Trading Typing for Proofreading

A UCI Health study (medRxiv preprint, 2026) of 23,760 AI-drafted notes found that 84.4 percent were edited by clinicians before signing. A randomized trial of AI scribes (Lukac et al., NEJM AI, 2025) found that Nabla users cut note-writing time by only about 9.5 percent relative to the control group, roughly 23 seconds per note once the control group’s own improvement is netted out, while DAX Copilot savings did not reach statistical significance. Kaiser Permanente’s 63-week deployment, the largest real-world evaluation at 7,260 physicians and 2.5 million encounters, reported that among non-users, barriers included “the perception that editing AI-generated notes took more time than typing from scratch.”

Physician satisfaction with AI scribes has been mixed in reported deployments, with health systems reporting roughly balanced shares of enthusiastic adopters and frustrated non-users. Readers should look to each institution’s own published data for specific Net Promoter Scores and retention rates.

A 30-Year-Old Workflow in New Packaging

The concept of converting spoken words to written clinical documentation goes back more than a century. Whether the intermediary is a stenographer (1900s), a transcriptionist (1960s-1990s), Dragon NaturallySpeaking (1997), or a GPT-4 ambient listener (2023), the workflow is largely unchanged: the provider speaks, a machine converts speech to text, the provider reviews the output. Microsoft acquired Nuance for $19.7 billion in 2022, and DAX Copilot is marketed as “the latest addition to the Dragon family.”

The core limitation is irreducible. Ambient scribes can only document what is spoken aloud. Clinicians who have piloted these tools have repeatedly noted that physical-exam findings and cognitive clinical reasoning often go uncaptured unless the physician narrates them out loud. The Assessment and Plan section, the most clinically important part of the note, is systematically under-documented by pure transcription because clinical reasoning is cognitive work that rarely gets spoken during the visit.

Coding Naive and Billing Blind

An AAPC article by Stephani Scott, “Why coding-aware AI scribes are essential,” draws the line between coding-aware scribes and coding-naive ones and argues that the coding-aware kind is essential. The Peterson Health Technology Institute’s March 2025 independent assessment, Adoption of AI in Healthcare Delivery Systems: Early Applications and Impacts, concluded that AI scribes have not yet achieved significant return on investment, and one study it reviewed found no statistically significant financial improvement. Reports in the field have also raised concerns that generative AI scribing, absent billing optimization, may increase note length rather than reduce bloat. Institutions evaluating AI scribes should assess note length changes directly in their own deployments.

AI ambient scribes today represent an incremental improvement on a decades-old dictation workflow. They still require the provider to verbalize everything, still generate errors requiring physician review, still often produce verbose notes, still struggle in noisy or complex clinical environments, still largely ignore billing optimization, and still lack FDA oversight while placing full legal liability on the signing physician.

Sources: Palm et al., Frontiers in Artificial Intelligence (2025); Asgari et al., npj Digital Medicine (2025); OHSU, JMIR (2024); Koenecke et al. (2024); UCI Health, medRxiv (2026); Lukac et al., NEJM AI (2025); Kaiser Permanente, NEJM Catalyst (2025); AAPC (Scott, undated); Peterson Health Technology Institute (March 2025)

If the encounter note is the root cause of healthcare’s dysfunction, and if traditional reform is structurally difficult, and if current technology largely automates the problem rather than solving it, then the solution must be a technology that operates directly on the encounter note, within the existing system, without requiring permission from any of the entities profiting from the status quo.

AHI has built two interlocking platforms designed for exactly this purpose. MDMai (Medical Decision Making AI) analyzes existing encounter notes for billing accuracy and documentation quality. ENCOUNTERai generates complete encounter notes predictively from minimal structured input, trained on each provider’s own clinical documentation patterns.

MDMai: What It Does and Why Nothing Else Does It This Way

MDMai is designed around a single objective: to make sure every encounter note it processes is billed at the most accurate and most defensible level the documentation will support. It is not a scribe, a claims scrubber, or a clearinghouse. It sits between the signed encounter note and the superbill, in the narrow window where the clinical record has already been created but billing decisions have not yet been made. That position is the one place in the revenue cycle where documentation can still drive coding rather than the other way around.

The problem MDMai is designed to solve has two compounding structural drivers. The first is that many notes are not signed at peak clinician attention. Arndt et al. (2017, Annals of Family Medicine) found that family physicians in one Wisconsin health system spent an average of 86 minutes per weekday on EHR work after clinic hours, the figure widely cited as “pajama time,” and subsequent JGIM and JAMA analyses have confirmed the pattern across specialties. A note signed at ten o’clock at night by a physician who has already seen 25 patients is rarely a note whose every billable element has been fully refined. The clinician wants to finish, sign, and go to bed. The downstream consequences of that understandable decision are significant, and they almost always favor the payer.

The second structural driver is that the U.S. medical coding system has grown into a thicket no single clinician or biller can be expected to hold in working memory. CPT alone spans thousands of codes. ICD-10-CM contains roughly 74,000 to 75,000 billable codes. HCPCS Level II adds several thousand more. Each set ships with sub-rules, modifiers, specialty carveouts, payer-specific overlays, and per-code documentation requirements. The reference libraries used by full-time trained coders literally fill shelves. Two concrete examples show how that complexity translates into invisible revenue loss at the note level.

The first example is sepsis specificity. ICD-10-CM distinguishes sepsis (A41.x), severe sepsis without septic shock (R65.20), and severe sepsis with septic shock (R65.21), and the grouping rules turn on exactly what the note says. When sepsis is the principal diagnosis, the admission groups to the MS-DRG 870 to 872 family, and the difference between DRG 871 (with a Major Complication or Comorbidity, or MCC) and DRG 872 (without one) depends on whether the record documents a separate secondary diagnosis that carries MCC status. When sepsis is coded as a secondary diagnosis, A41.9 is itself an MCC. A physician whose note describes the clinical picture without naming the condition it supports, whether that is the sepsis itself or a serious secondary condition, has unintentionally downcoded their own documentation. The note will be coded at the less specific level because that is what it literally says, and legitimate DRG weight can walk out the door on a single admission. ACDIS cautions that the severe sepsis codes, although they carry MCC status, may not change the MS-DRG once another MCC has been captured, which is why specificity has to be reviewed against the whole record rather than code by code. AHIMA, AAPC, and ACDIS have documented this pattern in clinical documentation integrity practice for years. The same under-specificity dynamic plays out in pneumonia (organism-specific versus unspecified), diabetes (with versus without named complications), heart failure (staged versus unstaged), and many others.

The second example is critical care. CPT 99291 and 99292 are criteria-based codes, not location-based codes. The governing criteria are a patient with critical illness or injury having a high probability of imminent or life-threatening deterioration, direct high-complexity decision-making by a physician or qualified healthcare professional, and total critical care time on the date of service of at least 30 minutes for 99291, with 99292 added for further time (under CMS rules, starting at 104 total minutes; under CPT, at 75). Nothing in the CPT or CMS definition restricts the code to an ICU or a critical care ward. Critical care can be billed in the emergency department, a step-down unit, a telemetry floor, a regular medical-surgical floor, or during a rapid response or code event, as long as the clinical criteria and time thresholds are met. (CMS Evaluation and Management Services Guide, MLN006764; AMA CPT Professional Edition.) Most clinicians who have not been trained as billers assume the opposite. Because they assume the opposite, they do not document the critical care time they actually provided, and the code never gets filed.

These are two examples of many. The point is not that providers are bad clinicians. The point is that the coding system rewards full-time trained billers and penalizes everyone else, and most independent practices cannot afford full-time trained billers. One could almost read the shelf-filling reference libraries and their overlapping rule sets as a system engineered to ensure no single practitioner could master the whole thing. MDMai is built to close that gap at the note level, before the superbill is created.

The Workflow MDMai Applies to Every Note

MDMai is a multi-agent AI system. When a signed encounter note enters the pipeline, it proceeds through five stages.

Stage One: Path Selection. For office and outpatient E/M visits (99202 to 99215), the 2021 E/M revisions adopted by AMA and CMS (effective January 1, 2021 and still in effect) allow the provider to select the billing level based on either total time on the date of encounter or medical decision-making complexity. Both paths remain valid, and they often yield different codes. MDMai evaluates each path against the documentation and recommends whichever yields the higher defensible code. This is not upcoding. It is the allowable election the CPT rules created specifically so that clinicians would not be penalized for spending more time with complex patients than the MDM structure alone rewards.

Stage Two: Time Reconstruction and Prolonged Services. When the time-based path is preferred, MDMai applies algorithms built specifically to reconstruct total clinician time on the date of service (face-to-face plus non-face-to-face time, as the 2021 rules define it) from the clinical evidence in the note. If the reconstructed time crosses the prolonged-services threshold, the appropriate add-on code is identified: CPT 99417 for commercial payers, HCPCS G2212 for Medicare, and the analogous inpatient codes (CPT 99418 commercial, HCPCS G0316 Medicare inpatient, G0317 nursing facility). Prolonged services are among the most commonly missed add-ons in outpatient E/M, and the revenue impact is material.

Stage Three: Billable-Element Refinement. Before any billing codes are generated, MDMai refines every billable element against the clinical data in the note. Diagnoses are pushed to the most specific ICD-10-CM level the documentation actually supports (the sepsis case is one instance; under-specificity in pneumonia, diabetes with complications, and heart failure staging are others). Procedures are checked for the most specific CPT code and correct modifiers. HCPCS Level II codes are scanned for separately billable supplies, drugs, and services. Critical care criteria are evaluated independently of encounter location. Chronic care management, transitional care, and other add-on elements are checked.

Stage Four: Code Generation. Only after Stage Three is complete does MDMai generate the final recommended billing codes. Because the upstream refinement has already happened, the codes are generated against refined clinical elements rather than against whatever the note happened to say in its raw form. The output is a structured report that lists each recommended code, the specific clinical language in the note that supports it, the CMS or payer rule that governs it, and any ambiguity flags where the documentation is close to the line.

Stage Five: The Biller Review. The MDMai report goes to the biller or practice manager. The biller decides what goes on the superbill. This is the architectural core of the system. MDMai does not write the superbill, does not submit claims, and does not remove the human from the billing decision. The report is a tool. The human has the last word on every claim that leaves the practice.

Why Human-in-the-Loop Is Deliberate, Not Incidental

The decision to position MDMai between the signed encounter note and the superbill, rather than automating the superbill itself, is not just a design preference. It is a compliance architecture.

That architecture is AHI’s own design choice, not a requirement imposed by federal guidance. HHS OIG’s General Compliance Program Guidance (November 2023) sets out the elements of an effective compliance program and does not address artificial intelligence. OIG’s Medicare Advantage Industry Compliance Program Guidance (February 2026) is addressed to Medicare Advantage organizations, and it raises AI in two places: payers’ use of AI in prior authorization, and risk adjustment, where it flags AI-generated prompts to physicians as potentially fraudulent and discusses AI platforms that analyze provider diagnosis reporting. Neither document requires human review of a provider’s coding suggestions before a claim is submitted. The enforcement backdrop is the False Claims Act: DOJ reported FCA settlements and judgments exceeding $6.8 billion in FY 2025, of which health care accounted for more than $5.7 billion, with managed care, prescription drugs, and medically unnecessary care among the areas it named.

MDMai is built so that every code it recommends can withstand that kind of scrutiny. Every recommended code is traceable to specific passages of the note. The biller’s review is the human step the design requires. If the biller accepts the recommendation, the clinical documentation supporting it is already attached. If the biller rejects it, the reasoning sits in the audit trail. A practice using MDMai is positioned to demonstrate exactly what auditors and enforcement reviewers are looking for. A fully autonomous AI biller would not just be clinically risky. It would invite precisely the False Claims Act exposure that unsupported codes create.

The Seven Functions MDMai Combines

Put the stages above together and MDMai performs, in a single integrated pass, seven functions that to our knowledge no other single system in healthcare combines:

Ingests encounter notes from any EHR. EHR-agnostic by design, with no requirement for API access, IT cooperation, or vendor approval.

Selects the most defensible allowable billing path (time-based or MDM-based) for each eligible E/M encounter under the 2021 and current guidelines.

Reconstructs clinician time on the date of service and identifies prolonged-services and add-on code opportunities that are routinely missed.

Refines every billable element (diagnoses, procedures, HCPCS, critical care criteria, chronic care management) against the clinical data in the note before any billing codes are generated.

Generates a structured, auditable report linking every recommended code to the specific documentation that supports it, giving the biller everything a compliance review would ask for.

Produces the annotated training dataset that ENCOUNTERai requires, paired note by note with MDMai’s analysis.

Generates a payer-specific first-pass appeal automatically the moment a claim is denied or downcoded. The biller uploads the denial, MDMai maps the denied elements to the documentation already in the note and to the specific CMS or payer rule the denial appears to violate, and produces a fully assembled appeal letter in seconds. This inverts the appeal economics described in Part 1: instead of 30 to 60 minutes of biller labor per denied claim, the first-pass appeal is generated in seconds, which makes it economically rational for billers and practices to contest every improper denial rather than letting the intimidation loop force them to stop coding accurately.

To our knowledge, no EHR performs this kind of upstream MDM and element-specificity analysis out of the box, no AI ambient scribe analyzes existing notes for billing optimization, no CDI tool handles outpatient E/M MDM analysis systematically across payer rule sets, and no revenue cycle platform processes notes before coding occurs. The gap is as much structural as it is technical. We are not aware of another organization that has built a system designed to simultaneously serve as a billing optimization tool, a documentation quality analyzer, an AI training data pipeline, a compliance audit trail, and an automated appeal generator, and to do all of it while leaving the human biller firmly in charge of the superbill.

The Accidental Single Source of Truth

There is a second, quieter consequence of MDMai’s architecture that was not a primary design goal but becomes one of the platform’s most important properties in practice. Because MDMai is EHR-agnostic, because it ingests every signed encounter note a provider produces regardless of which system the note was signed in, and because every analysis output is written back to AHI’s Healthcare Analytics Intelligence (HAL) platform under appropriate consent and privacy controls, the provider who uses MDMai across every site they practice is, without intending to, assembling the unified longitudinal record the government declined to mandate in 2009.

In the four-EHR scenario described in Chapter 2, the same patient’s clinic note, two separate Cerner hospital notes, and the FQHC note would each pass through MDMai. The analysis output for each encounter, along with the underlying note, would sit in one place. For the first time in that patient’s relationship with that provider group, there is a single store of record the treating clinician can consult at the point of care, keyed to the patient rather than to the facility that owns the EHR. MDMai was built to make notes billable. The side effect is that it also makes them findable.

This is not a claim that MDMai replaces an EHR, a Health Information Exchange, or TEFCA. It does not. The providers still chart in their hospital’s Cerner or their clinic’s system of record, and the legal medical record remains where the institution keeps it. What MDMai produces is a working, patient-centered, cross-facility view of every encounter the provider has signed, which is what is clinically useful at the bedside and what has been structurally missing from American medicine since digitization began. In a market where interoperability has remained a promise for seventeen years, a unified view that emerges as a byproduct of solving the billing problem is an advantage worth being explicit about.

ENCOUNTERai: The Predictive Documentation Engine

ENCOUNTERai represents a fundamentally different approach to clinical documentation. Rather than transcribing what a physician says (ambient scribes) or templating what they click (EHRs), ENCOUNTERai is designed to generate complete encounter notes predictively from minimal structured input (diagnosis codes, key findings, treatment decisions) in each provider’s clinical voice, optimized for billing from the moment of creation.

The underlying insight: research suggests that a substantial share of clinical note text is not originally composed at each visit. Wang et al. (JAMA Internal Medicine, 2017) found only 18 percent of progress note text was manually entered; 46 percent was copied and 36 percent imported from structured data. Rule et al. (JAMA Network Open, 2021) found 58.8 percent median note redundancy in an analysis of 2.7 million outpatient notes. Holmgren, Sinsky, Rotenstein and Apathy (JGIM, 2024) analyzed EHR use by 200,081 ambulatory physicians and documented large, systematic differences in EHR and documentation time across specialties. This predictability is precisely the condition under which a provider-specific AI model can be most useful.

The design goal is to approach approximately one minute of provider review-and-sign time per encounter, compared with the 16 to 36 minutes currently consumed by full documentation. This one-minute figure is a design target based on AHI’s internal modeling, not a measured outcome in a clinical trial, and will be validated as the platform is deployed. The physician’s review time is intended to be spent on clinical accuracy verification rather than billing compliance work.

The Training Pipeline

The MDMai-to-ENCOUNTERai pipeline operates in four stages.

Stage 1. Provider uses MDMai. Encounter notes are processed, providing immediate billing optimization and documentation feedback. The provider benefits from Day 1 through recovered revenue from undercoding and reduced denial risk.

Stage 2. MDMai generates annotated training data. Each processed note produces a paired dataset of raw note plus MDMai analysis (billing codes, MDM level, optimization recommendations, gap identification), stored in AHI’s Healthcare Analytics Intelligence (HAL) platform under appropriate consent and privacy controls.

Stage 3. Once a provider has accumulated a sufficient corpus of annotated notes, ENCOUNTERai training begins. The accumulated dataset allows per-provider fine-tuning. The model is intended to learn the provider’s documentation patterns, clinical reasoning, terminology, examination conventions, and the billing-to-documentation relationship. AHI’s current working threshold is on the order of 1,000 to several thousand labeled notes per provider, calibrated per specialty. These thresholds reflect internal design targets rather than published outcome data, and will be refined as real-world evaluation proceeds. At typical outpatient volumes of 20 to 25 patients per day, providers can be expected to reach this threshold in roughly 60 to 90 working days.

Stage 4. ENCOUNTERai generates predictive documentation from minimal input, optimized for billing from inception. The provider reviews and signs.

This pipeline is why MDMai is offered free to physician practices. It is not philanthropy. It is the data acquisition mechanism that makes ENCOUNTERai possible. The value proposition is aligned. The provider benefits immediately, and the accumulated data creates an asset that would be difficult for competitors to replicate.

The Competitive Moat: Why This Pipeline Is Difficult to Replicate

EHR vendors have limited incentive to build a comparable pipeline. Their business model is platform lock-in, and a vendor-neutral analysis tool would threaten their position. AI scribe companies generate notes from voice but do not analyze existing notes or perform MDM analysis. Revenue cycle companies operate post-hoc. They process claims after submission, not clinical notes at creation. Academic institutions typically lack an EHR-agnostic processing engine deployable to any practice. The per-provider training threshold creates a time-based moat. Even a competitor who built a comparable engine tomorrow would need months of data per provider before their system could function.

The U.S. generates on the order of a billion clinical encounter notes annually; CDC counts about 1.0 billion physician office visits a year. We believe AHI’s MDMai platform is among the only EHR-agnostic mechanisms that can aggregate these notes with billing annotation at scale. The resulting database (encounter notes paired with MDM analysis, billing codes, and optimization data) is intended to become one of the more complete clinical documentation datasets available, and it grows with every provider who uses the free MDMai tool.

Sources: AMA CPT 2021 E/M Revisions and FAQs; CMS Medicare Claims Processing Manual Transmittal R11842CP; CMS Evaluation and Management Services Guide (MLN006764); Arndt et al., Annals of Family Medicine (2017); AHIMA and ACDIS clinical documentation integrity guidance on sepsis and under-specificity coding; AAPC and Find-A-Code guidance on 99417, G2212, 99418, G0316, G0317 prolonged services; HHS OIG General Compliance Program Guidance (November 2023); HHS OIG Medicare Advantage Industry Compliance Program Guidance (February 2026); DOJ False Claims Act enforcement statistics (FY 2025); CMS MS-DRG definitions manual; Wang et al. (2017); Rule et al. (2021); Holmgren, Sinsky, Rotenstein and Apathy, JGIM (2024); CDC National Ambulatory Medical Care Survey

From 1,400 Hours to 100 Hours

A national analysis of EHR use by 200,081 ambulatory physicians (Holmgren, Sinsky, Rotenstein and Apathy, JGIM, 2024) found that physicians spend 5.8 hours on the EHR for every 8 hours of scheduled patient time across all specialties, 7.3 hours in primary care, and 8.4 hours in infectious disease, the highest of any specialty. At an assumed 240 clinical days of 8 scheduled patient hours each, the all-specialty and primary care figures translate to approximately 1,400 to 1,750 hours a year consumed by EHR work. Family physicians in the Arndt study additionally logged an average of 86 minutes of after-hours “pajama time” per weekday, roughly 340 hours over 240 working days, stolen from evenings and families.

Target figures above reflect AHI’s internal modeling of ENCOUNTERai once trained for a given provider and are not outcomes from a completed clinical trial.

A time dividend of 1,000 to 1,500 hours would create meaningful physician-directed optionality. At 18 minutes of saved documentation time per encounter, recaptured hours could in principle support additional patient visits per year. As an illustrative projection only, reallocating even part of that time to additional patient visits could add meaningful annual revenue per physician; actual outcomes will vary substantially with specialty, payer mix, local fee schedules, and physician capacity, and have not been measured in a controlled study. Across one million physicians, even 500 additional visits per physician equals 500 million additional patient encounters nationally, a substantial offset to the AAMC’s projected shortage of up to 86,000 physicians by 2036, with HRSA’s current projection showing a shortage of 141,160 physicians in 2038.

In one survey reported by Becker’s, 77 percent of physicians said they would take a pay cut for more autonomy and about half for better work-life balance, supporting the view that working conditions, documentation burden among them, are a binding constraint, not just compensation level.

The Medicaid Access Unlock

The economic barrier to seeing Medicaid patients is not low reimbursement alone. It is low reimbursement combined with high documentation time. Across state Medicaid fee schedules, a 99213 visit pays a median of about $71, with rates ranging from about $35 to $165. When that visit requires 20 to 30 minutes of documentation time, the median fee works out to roughly $140 to $215 per hour of documentation time alone. With ENCOUNTERai, AHI’s design target is for the same visit to require roughly one minute of review (again, a design target rather than a measured outcome). The economic friction behind the roughly 26 percent of physicians accepting new patients who did not accept new Medicaid patients (MACPAC, 2017 data) is significantly reduced.

The downstream cascade: more physicians accept Medicaid patients, FQHC demand decreases (easing the cost of a parallel system whose health centers took in $42.9 billion in revenue in 2022, according to GAO), avoidable ER visits decrease (a savings opportunity UnitedHealth Group puts at $32 billion a year), wait times decrease (currently around 31 days on average, up significantly since 2004 according to AMN Healthcare’s 2025 physician wait time survey), and patient outcomes improve through earlier access to primary care.

The Documentation Quality Paradox: Less Time, Better Notes

Olson et al. (JAMA Network Open, 2025) reported that when AI documentation tools partially fix the burden, burnout drops from 51.9 percent to 38.8 percent in 30 days, after-hours documentation time drops approximately 0.90 hours per week, and note-related cognitive task load falls 2.64 points on a 10-point scale. The study was an uncontrolled before-and-after survey and did not measure note quality.

ENCOUNTERai separates clinical quality from billing optimization. Today, physicians must simultaneously think about what is clinically relevant, what is legally protective, and what supports billing. This triple cognitive load can degrade all three. ENCOUNTERai is designed to handle the legal and billing dimensions automatically, allowing physicians to focus on clinical accuracy during their brief review. The goal is notes that are better clinically despite requiring much less physician time.

Sources: Holmgren, Sinsky, Rotenstein and Apathy, JGIM (2024); Arndt et al. (2017); AMN Healthcare Physician Wait Time Survey (2025); Olson et al., JAMA Network Open (2025); MACPAC (2021), from NEHRS data; payerprice.com compilation of state Medicaid fee schedules; GAO-24-106815; UnitedHealth Group on avoidable ER visits; Becker’s Hospital Review on physician pay and autonomy; AAMC (2024); HRSA National Center for Health Workforce Analysis projections

A natural objection arises. If MDMai reduces undercoding and practices suddenly bill at accurate levels, will insurers respond with more aggressive denials, audits, and downcoding? The evidence shows this concern is real but strategically manageable. Payers are already retaliating against accurate coding, making MDMai’s defense architecture not optional but essential.

Part 1 of this white paper described the closed loop that the payer retaliation exploits: accurate coding triggers denials, appeals are individually unprofitable for billers working on commission (30 to 60 minutes of labor per appeal against a 7 to 10 percent commission on a recovered claim), and practices therefore come under pressure to stop coding accurately to preserve the biller’s economics. We call that closed loop pay by intimidation. MDMai’s defense architecture is designed to neutralize the loop by automating the appeal itself, which inverts the appeal economics that the payer is relying on.

The Cigna / Aetna Downcoding Offensive

Cigna’s Evaluation and Management Coding Accuracy Policy (R49), effective October 1, 2025, introduced automated review of E/M codes 99204 and 99205, 99214 and 99215, and 99244 and 99245, in which payment is adjusted downward when the submitted diagnosis code is deemed not to support the billed complexity level without a review of the underlying medical record. Aetna’s Claim and Code Review Program, piloted in 12 states, expanded in March 2025 to all commercial states except Louisiana. (Cigna paused R49 for its California fully insured HMO and PPO plans after a challenge from the California Medical Association; readers should consult primary Cigna policy bulletins for its current status elsewhere.) To receive the originally billed payment, providers must file a post-payment appeal with full medical records. E/M code 99214 is the most frequently billed E/M code in Medicare Part B, representing approximately 25 percent of allowed E/M claims. ASIPP concluded that “overcoding of CPT codes 99214 and 99215 is de minimis,” suggesting that these policies target legitimate coding rather than fraud.

ASIPP characterized the policies as a “deny first, pay later” system that “presumes physicians are guilty of ‘upcoding’ until proven innocent.” The strategic logic is transparent. Insurers expect that busy practices will not have the time or resources to appeal every improperly downcoded claim, resulting in net savings. ASIPP also noted that independent practices lack the armies of coders and administrators that hospital systems have, and often lack the capacity to engage the insurers’ bureaucratic processes.

40-Plus Health Systems Have Called the Bluff

Becker’s counted 40 health systems dropping Medicare Advantage plans in 2025, and Mayo Clinic, NewYork-Presbyterian, and UNC Health have announced exits for 2026. Providers often have more leverage than they realize. HealthPartners alleged UnitedHealthcare denial rates were “up to 10 times higher than other insurers.” A KFF report found that MA enrollees had access to only about 48 percent of the physicians available to Traditional Medicare beneficiaries.

CMS has been widely reported to have never imposed intermediate sanctions on Medicare Advantage plans for network adequacy violations. MedPAC’s June 2024 report noted that CMS has the authority to impose intermediate sanctions or civil monetary penalties for noncompliance with network adequacy standards, but it has never done so. Provider exits force payers either to renegotiate or to risk losing members when health systems grant patients special enrollment periods to switch plans.

Why Documentation Beats Algorithms

The evidence on claims appeal success rates is central to MDMai’s defense:

These figures come from different datasets and denial types and should be read individually rather than combined into a single range. They support a consistent conclusion: most of the denial system’s economic advantage rests on low appeal rates, not on the strength of individual denials.

The system’s most exploitable vulnerability is that so few denials are appealed: fewer than 1 percent of denied Marketplace claims (KFF) and about 1 percent of Medicare Advantage denials (HHS OIG, 2018). The payer strategy depends on this: deny at scale, count on providers not appealing, and retain revenue on the overwhelming majority of denials that go unchallenged. MDMai is designed to disrupt this equation. When documentation is airtight from creation and auto-generated appeal letters map specific MDM elements to CMS guidelines, AHI’s design goal is to materially reduce the provider’s cost of appealing; this is a design target rather than a measured outcome to date. If achieved, the payer’s “deny first” strategy becomes economically less attractive at typical denial volumes.

The Auto-Appeal Generator: Neutralizing Pay by Intimidation

Part 1 of this white paper named the closed loop that keeps providers undercoding: payers deny accurate claims, billers lose money appealing at standard commission rates, and practices stop coding accurately to protect the biller’s economics. Pay by intimidation. The auto-appeal generator inside MDMai is the technical response to that loop.

When a claim is denied or downcoded, the biller uploads the denial notice. MDMai already has the underlying encounter note in memory, along with the MDM analysis it produced when the claim was first coded. The system generates a payer-specific appeal letter in seconds, citing the exact documented elements that support the originally billed level, the specific CMS policy or payer medical policy language that contradicts the denial rationale, and the precedent-style argument appropriate to that denial reason. The biller or practice reviews, approves, and submits. What previously cost 30 to 60 minutes of labor per claim now costs a few minutes of review.

The economic math inverts. At 30 to 60 minutes of labor per claim, appeals were individually unprofitable for billers working on commission, which is why providers were pressured to stop coding accurately. With automated first-pass appeal generation, the labor cost per appeal drops sharply, which means appeals become rational for the biller and the practice to pursue. At that point, the payer’s denial strategy stops being a cheap revenue-retention mechanism and starts being a cost center, because each denial now draws a well-documented appeal that the payer has to actually adjudicate. When the cost of denying accurately coded claims exceeds the expected recovery from providers who give up, the denial machine loses its economic rationale.

The auto-appeal generator is not a workaround for the payment architecture. It is a direct counter to the enforcement mechanism that the denial system uses to keep the payment architecture broken. It is specifically the reason MDMai is designed as both an offensive tool (coding accuracy) and a defensive tool (appeal automation) in a single platform.

Policy Lever: States That Raise the Wage Floor Must Also Raise Reimbursement

There is also a state-policy corollary. States that raise the minimum wage for healthcare workers (California’s SB 525 is the clearest current example) create a cost-reimbursement scissor that accelerates the closure of independent practices unless they pair the wage increase with an equivalent update to Medicaid physician fee schedules and commercial parity protections. The scissor is not an argument against wage increases for healthcare workers. It is an argument that wage policy and reimbursement policy must move together, because the note is the only point where the combined cost side and revenue side of the practice meet. Technology alone cannot absorb a permanent scissor if the state raises cost while the payer holds reimbursement flat.

Sources: Cigna R49 policy documentation; AAFP Family Practice Management on Cigna R49; ISMA (June 2025) on Aetna’s Claim and Code Review Program; ASIPP (September 2025); Becker’s Hospital Review (2025); KFF Health News (2025); MedPAC (2024); KFF (2025), 2023 data; KFF (2024 data) on Medicare Advantage prior authorization appeals; Health Affairs (2025); HHS OIG (2018); California SB 525 (2023); U.S. Department of Labor state wage data; MGMA benchmarks on appeal labor cost