David Brailer’s June 11, 2026, commentary in Health Affairs Forefront describes how US health care payers and providers use artificial intelligence (AI) to optimize revenue under current payment structures: providers documenting intensity, plans calibrating risk scores, and both sides automating the contest over denials and appeals. The description is accurate but narrow. AI also affects production of health: diagnosis, monitoring, triage, treatment selection, and the substitution of computation for clinical labor. A frame built on the first set of applications is likely to reach the wrong conclusions about AI’s effects on costs, prices, and value.

Brailer asserts that “both care improvement and extraction are inflationary,” and that “the right frame is to understand how fast AI will drive these forces and compound them, on top of an inflation baseline that was already unsustainable.” Three claims are packed in: That care improvement is inflationary, that the baseline is unsustainable, and that AI’s speed makes the moment exceptional. I take them in order, then turn to policy.

Defining And Measuring Health Care Inflation

Brailer does not define inflation and uses it interchangeably with spending. They differ. Inflation is growth in prices, holding the product constant; spending is prices times quantities. Spending can grow rapidly with zero or negative inflation when the growth comes from treating more patients, more intensively, or with better technology. The current data show exactly that: The national health expenditure accounts attribute recent growth predominantly to volume and intensity, with virtually no excess medical price inflation in 2023 or 2024.

Accounting for changes in quality deepens the difference further yet. Measuring the price of producing a health outcome rather than an encounter, Abe Dunn, Anne Hall, and Seidu Dauda found that quality-adjusted medical prices fell 1.33 percent per year from 2000 to 2017, even as the official index showed prices rising 0.53 percent per year faster than economywide inflation. Measured by the health it produces, medical care has been running a quiet deflation for two decades.

AI has real potential to accelerate this trend. Consider AI-assisted imaging that detects a cancer earlier. Aggregate spending may rise as the treated population expands, yet the quality-adjusted price of treating that cancer falls because treatment is less extensive and outcomes improve. A system with physician leaders who call that inflation and move to suppress it suffers an iatrogenic disease. The standard economic definitions and methods can help discern between AI applications that enhance value from those that enhance billing.

“Already Unsustainable” Was Richard Nixon’s 57-Year-Old Forecast

Brailer’s second claim, that the baseline “was already unsustainable,” echoes President Richard Nixon, who in 1969 warned, “We face a massive crisis in this area; and unless action is taken, both administratively and legislatively, to meet that crisis within the next 2 to 3 years, we will have a breakdown in our medical care system.” The breakdown is nearly six decades behind schedule.

Three demand-side findings explain the gap. As an empirical regularity, the income elasticity of health spending exceeds one, so health’s share of gross domestic product (GDP) rises with income; the Centers for Medicare and Medicaid Services’ (CMS’s) own projections treat disposable income as the primary driver of spending growth. Robert Hall and Charles Jones explain why that is optimal rather than pathological: As income rises, the marginal utility of more consumption falls while the value of healthy life does not, so the optimal health share rises with income, plausibly past 30 percent of GDP by mid-century. Kevin Murphy and Robert Topel size the benefit: US life-expectancy gains from 1970 to 2000 were worth roughly $3.2 trillion per year, value absent from any GDP account. Health complements all other consumption because longevity extends the horizon over which every other good is enjoyed. A boat is worth little to an owner too sick to use it.

I write to bury the forecast, not to praise US health spending. The health share of GDP is not, by itself, evidence of a problem. “Unsustainable” conflates fiscal financing, which concerns tax and transfer design and is a genuine constraint, with resource allocation, which AI is well-equipped to improve.

The Supply Side: What AI Does To Baumol’s Constraint

Demand explains why nations buy more care as they become richer. Supply-side factors help to explain why the relative cost of providing medical care has risen over time. Baumol’s cost disease highlights the role of differences in production functions across sectors: Where capital substitutes easily for labor, productivity rises and relative prices fall; in labor-intensive care it does not, so the relative cost of the staffed hour rises. By performing cognitive and administrative work within care delivery, AI represents a new ability to substitute capital for labor and cure the cost disease. This is a notable departure from Brailer: He treats AI as an accelerant of the existing cost structure, when its distinguishing property is that it relaxes the constraint that built that structure.

This diagnosis and prognosis counter Brailer’s claim that what AI “will not do, without deliberate intervention, is make care cheaper.” AI has meaningful ability to lower health care production costs. How this affects prices depends on pricing policies and market structure, addressed below.

Clinical AI Is A Targeting Technology, And Targeting Sets The Sign

Brailer’s third claim is, “The question is not whether clinical AI will raise costs. It will,” whereby “costs” he means spending. The pithy conclusion does not withstand scrutiny: Whether AI raises or lowers spending, and the normative implications of either, depends on how it influences the allocation of treatments. In recent work in NEJM AI, I show that predictive clinical AI can act as a return-on-investment multiplier on the therapies it guides, lowering the number needed to be treated to reach breakeven. For therapies being overused, or with poorly targeted deployment, this mechanism reduces volume. In clinical areas with underdetection, volume will rise and health will improve. Thus, we should expect that AI will increase volume of some treatments, lower volume of others, with a net effect on spending that depends on the relative magnitudes of these changes as well as how prices adjust.

Why Lower Production Costs Have Not Become Lower Prices

As Brailer notes, the electronic health record era saw building of billing infrastructure and not health infrastructure partly because the US health care system rewarded the first and not the second. At the core of that system is government price setting. Medicare pays administered prices set by formula, attached to codes describing inputs, procedures, and diagnoses. Providers invested in upcoding technology to arbitrage against sclerotic price-setting formulas.

Where prices adjust, arbitrage is short-lived as rivals chase the margin and buyers reprice. However, two features of the US system disable such adjustments. First, administered prices cannot adjust by construction; a committee revises them on regulatory timelines, so gaming persists until a late recalibration that is gamed in turn. This carries through to commercial prices, too: A provider selling to many payers faces a stepwise demand schedule in which Medicare’s administered fee forms a flat segment that anchors private negotiation. A $1.00 increase in Medicare’s fee raises corresponding private prices by about $1.16, and roughly 75 percent of physician services, accounting for 55 percent of commercial spending, are benchmarked contractually to Medicare’s relative-value menu. The benchmark comes from the government; market structure decides only who captures the markup. Second, with high market concentration of both payers and providers, the rents do not get competed away nor the savings passed through to consumers.

The Diagnosis Determines The Prescription

Brailer concludes that AI can only compound the cost problem. He reaches his verdict by examining only the mechanisms that raise spending—coding intensity, denial arms races, expanded detection—and presents evidence for each. That evidence is sound, and the arbitrage analysis above accounts for it: AI accelerates gaming against formulas that cannot reprice.

Overlooked, however, are three mechanisms that cut the other way. Two are discussed above: treatment of Baumol’s cost disease and improved targeting. The third is that consumer-facing AI tools can strengthen the quality competition that disciplines providers, the same competition that operates in any market where buyers can see and value what they are buying. Health care is not exempt from this economics. Where markets determine prices via competition, a firm that lowers its production costs must pass the savings on to buyers as lower prices, which is why no one warns that logistics AI is dangerous when it works as intended. Thus, Brailer’s distinction between clinical and extractive AI by “what those processes optimize for” misdiagnoses the issue. Optimizing for profit is constant across industries: The variables are the pricing mechanism and market structure.

Brailer proposes three regulatory remedies. First, required disclosure of an AI system’s objectives. Second, documentation of AI on par with that required of humans. Third, restricting payment for AI-assisted care to only those establishing improved outcomes. These regulations risk precluding the very mechanisms by which AI can lower costs. Disclosure and parity obligations are fixed compliance costs, lightest on the incumbents who already run them and heaviest on the would-be entrants who would otherwise enhance competition. Likewise, conditioning participation on demonstrated outcomes favors incumbents who already hold the data.

With proper measurements and comprehensive consideration, AI represents a substantial technological advancement with the capacity to lower medical inflation and increase value. Whether it does turns on reforms that span AI’s health care applications: data governance that facilitates development of AI tools; payment methods that allow innovators to capture some of the value they create; and regulatory pathways that don’t saddle AI with constraints built for humans or older, static technologies.