The prevention claim has been tested once, and the same trial measured the cascade. In a randomized study, flagging coronary calcium found incidentally on prior chest imaging raised statin initiation at six months from 6.9 percent to 51.2 percent. That is a large effect on prescribing. In the same trial, downstream cardiac testing rose from 2.3 percent to 15.1 percent, a more than sixfold increase. The investigators state explicitly that further research is needed to determine whether the approach reduces cardiovascular events.

So the mechanism and the cascade were measured in the same study, and only the cascade was measured in hard clinical units.

One program shows the whole sequence in a single American dataset. A health system deployed artificial intelligence to ensure that recommended imaging follow-up actually happened. It worked. Recommendations rose from 75.4 percent to 95.7 percent of relevant reports, and completion of recommended follow-up rose from 29.6 percent to 50.7 percent. Program volume went from roughly 6,000 examinations a year to roughly 19,000 after natural language processing was added.

The share of followed findings that turned out to be malignant, suspicious, or clinically significant was 33.3 percent before and 35.3 percent after. Unchanged.

And the program’s published description states that the added imaging covers the cost of the staff running it, and that an internal business plan shows the program is profitable.

Read that sequence carefully, because it is the article in miniature. The technology worked exactly as designed. It found more. The proportion that mattered did not change. And it paid for itself on the additional imaging.

Which brings us to how the return on imaging artificial intelligence is actually calculated. The most cited return analysis in the peer-reviewed radiology literature reports 451 percent over five years, rising to 791 percent when radiologist time is counted. It is a model rather than a measurement, built from expert interviews, using one practice’s scan volumes and assuming ten percent annual visit growth. Representatives of the vendor whose platform it evaluates are co-authors.

Set all of that aside and look only at how the model earns its return. Its revenue sources include additional revenue from follow-up scans, hospitalizations, and treatments triggered by artificial intelligence findings. The single most influential variable in the model is the number of additional treatments attributed to artificial intelligence identification.

The headline business case for imaging artificial intelligence is driven by how much additional care the technology generates. That is not an accusation. It is an accurate description of where the money comes from, published openly, and almost never quoted.

Two measured per-patient figures, pointing in opposite directions. Artificial intelligence applied to lung cancer screening saved up to 242 dollars per patient. Applied to mammography, it raised cost by up to 19 dollars per patient, attributed to lower specificity combined with per-use pricing. Both are measurements. The second is the kind of finding that does not reach a vendor deck.

And the payment picture explains the rest. Several hundred imaging algorithms have been cleared. As of January 2026 there are two Category 1 payment codes for imaging artificial intelligence, both cardiac. The first, a measure of blood flow derived from cardiac computed tomography, became the first artificial intelligence enabled clinical software reimbursed by Medicare in 2018. Across five years it generated 18.2 million dollars in total Medicare spending, reached 2.2 percent of patients receiving the underlying scan, and was billed by only 17 percent of hospitals performing it.

Eighteen million dollars over five years is a rounding error in a program that spends roughly a trillion. The distance between what has been cleared and what has been paid for is the whole story of imaging artificial intelligence economics.

One coda, for anyone building governance around this. Medicare’s Appropriate Use Criteria program, intended to curb inappropriate advanced imaging through mandatory decision support, was paused and its regulations rescinded effective 2024. The stated reason was that the agency’s own claims processing system could not automate compliance monitoring. The program designed to constrain imaging volume ended because it could not be administered.

The pattern underneath all of it

Lay the quality evidence out side by side and something becomes obvious.

Three applications have randomized evidence of improved patient outcomes: the retinopathy screening trial, the deterioration model, and genotype-guided dosing. Every one of them changes what happens next. A referral step is removed. A clinical team escalates. A dose is different.

The applications with only accuracy or efficiency evidence share the opposite property. The note is written faster, and the same visit occurred. The alert fires, and it is overridden. More adenomas are found, and the same proportion of patients has advanced disease.

Information that does not change an action does not change an outcome. An outcome that does not change does not change cost.

There is a third category, and imaging is where it lives. Sometimes the information does change an action, and the action it changes is another test, another procedure, another prescription. The pathway moves. It just does not move toward a better outcome. It moves toward more care.

That is the version of this problem operators should be most alert to, because it looks like success in every report that matters to a vendor. Detection is up. Follow-up completion is up. Utilization is up. Revenue is up. The only number that did not move is whether patients are better off.

So the question to put to any clinical artificial intelligence vendor is not how accurate the model is. It is: which step in our pathway disappears, and which step gets added. If the honest answer is that nothing disappears and a workup gets added, the operator has just been shown the business case, and it is not a cost reduction.

Cost: the honest answer

Here the evidence is not ambiguous, and the most persuasive sources are not critics. They are the organizations writing the checks.

What actually drives spending

Across the commercially insured population, per-person spending rose 18.8 percent between 2019 and 2023, an increase of 1,068 dollars. Over that same five years, average prices rose 17.7 percent. Utilization rose 1.8 percent.

Read those three numbers again with a vendor business case in hand. Essentially all of the spending growth was price. Almost none of it was people using more care. A tool that lowers the cost of producing care therefore changes almost nothing for the purchaser unless it changes a negotiated price, and nothing in the technology touches that negotiation.

The payers and employers now name artificial intelligence as an inflator

This is the finding that should reset the conversation.

In the 2027 medical cost trend survey, health plans rank artificial intelligence tools that help providers capture more revenue as the single largest inflator of trend. Seventy percent of plans place it in their top three. The survey covers 27 plans and more than 103 million employer-sponsored members. Projected group trend for 2027 is 9.0 percent.

Employers say the same thing independently. In a survey of 127 employers covering 11 million lives, 64 percent report that artificial intelligence is driving their costs up through revenue optimization and upcoding. Their projected 2027 trend is a median of 9.2 percent.

The mechanism both describe is specific and it is not fraud. Documentation and coding tools record greater clinical specificity and greater reimbursable severity for care that was already delivered. The encounter does not change. The claim does. The payer pays more.

This is not a new behavior. Federal auditors found that between 2014 and 2019, hospital stays billed at the highest severity level rose by almost 20 percent while every other severity tier fell and total stays declined. Average length of stay for those highest-severity cases fell by half a day. More than half had only one qualifying condition. The conclusion was that this reflected changes in billing rather than sicker patients.

Artificial intelligence did not create that practice. It lowered the marginal cost of doing it to nearly zero.

Where the money is actually going

Across surveyed health systems and plans, the pattern of deployment is consistent and it is not clinical. Health systems report 71 percent some deployment of ambient documentation, with 41 percent enterprise-wide, against 18 percent enterprise-scale clinical decision support. Health plans report 67 percent some deployment of clinical documentation review.

Both sides of the payment relationship automated the documentation contest first.

Industry surveys of return on investment point the same direction, with the caveat that they are self-reported and at least one is published by an investor in the category. In those surveys, 67 percent of provider artificial intelligence solutions run as semi-autonomous or fully autonomous agents, while only 4 percent of clinical solutions do. Reported returns are led by revenue cycle, claims, and back-office work. Clinical returns are among the lowest reported.

The number everyone quotes incorrectly

The most cited savings estimate in this field is 200 to 360 billion dollars annually, or 5 to 10 percent of United States health spending. It is quoted as a total almost every time it appears.

Read the paper. The authors deliberately excluded the zero-sum domains, because revenue one organization gains is revenue taken from another. The estimate is already net of redistribution. The authors also state plainly that their figures rest on consulting experience rather than randomized or quasi-experimental evidence, and the paper does not specify whether the savings would accrue to payers, providers, or patients.

Alongside it, note what the waste literature actually concluded. Administrative complexity is the largest single category of waste in United States healthcare at 265.6 billion dollars annually. The authors assigned it no savings estimate at all, because no generalizable intervention studies existed to support one.

Artificial intelligence is not paid for as care

As of early 2026 there are 26 Current Procedural Terminology codes for clinical artificial intelligence, 23 of them Category III, which generally carry no Medicare pricing. There are 19 software-as-a-service codes. In 2023, no such code was billed more than 3,600 times or paid more than 200,000 dollars. The imaging figures above are representative rather than exceptional.

Between 2018 and 2023, only three Medicare-covered artificial intelligence tools reached more than 10,000 beneficiaries.

The Medicare Payment Advisory Commission’s position is that separate payment for software could encourage overuse and raise costs without demonstrated value.

The operational consequence is straightforward. Artificial intelligence is an operating expense that buys no new reimbursable service on the care side. That is precisely why the return chases the revenue cycle. It is the only place the money is.

The clawback, which almost no one has priced

This is the most important near-term fact in this article and it has had very little attention.

In the calendar year 2026 physician fee schedule, the Centers for Medicare and Medicaid Services finalized a 2.5 percent efficiency adjustment reducing work relative value units and the intraservice portion of physician time for services that are not time-based. The stated rationale includes that digital tools and automation accelerate efficiency gains. Time-based codes are exempt.

Consider what that means for a business case. A practice buys a tool, becomes more efficient, and keeps the gain until the next repricing cycle, at which point the payer captures it by regulation. There is also a long-standing statutory productivity adjustment reducing annual hospital payment updates by a rolling average of economy-wide productivity.

So there is exactly one line of business in which artificial intelligence’s production savings reliably reach the purchaser, and it is Medicare fee for service. The mechanism is not competition. It is a regulation that assumes the savings exist and reduces payment accordingly, whether or not the practice ever bought the tool.

That is an uncomfortable finding for providers and it belongs in every multiyear financial model that includes an artificial intelligence investment.

The precedent

The federal government spent roughly 35 billion dollars on electronic health record adoption incentives. Adoption became near-universal. Documentation burden went up. The lesson is not that the investment was wrong. The lesson is that near-universal adoption of a technology is not the same event as a reduction in cost, and the two were conflated for a decade.

The strongest counterargument, stated fairly

Artificial intelligence demonstrably lowers the cost of producing administrative work inside a single organization. Coding, claims handling, prior authorization submission, appeal drafting, documentation. That is real, it is measured, and it is where essentially all of the current return is concentrated. Anyone arguing otherwise is not reading the deployment data.

The distinction is this. Reducing one organization’s cost of contesting payment with another organization is not reducing the cost of care. The money moves faster and at lower administrative expense. Nothing becomes cheaper to deliver.

What the payment architecture rewards

The technology is nearly identical across every line of business. What differs is what each payment architecture pays it to do, and that single observation explains most of the variation in where artificial intelligence has been deployed.

Medicare fee for service cannot raise price, so the return sits in volume and documented intensity, and the program has now begun reclaiming assumed efficiency through the fee schedule. Medicare Advantage pays for diagnosis capture, which is why roughly 22 billion dollars of the projected 76 billion dollar gap between Medicare Advantage and fee-for-service payment in 2026 is attributed to residual coding intensity. The marketplace is budget neutral within each state, so coding harder than your neighbors is the only advantage available and it vanishes the moment they match you. Self-funded employers have no subsidy to mine, so provider revenue capture lands directly on the plan sponsor. And Medicaid is contracting by statute rather than by technology, with artificial intelligence deployed to administer the contraction.

That last line carries the governance lesson that transfers everywhere else. Four states have lost litigation over algorithmic reduction of benefits, and in every case they lost on notice and explainability rather than on accuracy. A plan or a state can lose on an algorithm nobody ever proved was wrong, because it could not explain what the algorithm did. Most artificial intelligence governance frameworks underweight this badly.

The companion resource sets out each line of business, what its architecture rewards, who bears the cost, and who captures the saving.