Medicare Tested AI to Pre-Approve Care — and Medical Services Dropped Immediately

Medicare has been quietly running one of the most consequential artificial intelligence experiments in American healthcare — and the first results arrived at a speed that startled even seasoned policy watchers. The federal program is testing the use of AI to conduct prior authorizations, the advance approvals that determine which medical services get covered, in an effort to target wasteful spending, MarketWatch reported on Thursday, October 29, 2026. The outcome was immediate: a large drop in medical services followed the deployment. That single finding upends years of assumptions about how slowly bureaucratic AI adoption moves, and it lands squarely in the middle of a fierce debate about whether algorithms should ever sit between doctors and patients. Beneath the headline sits a deeper story — about what "waste" actually means, how quickly algorithmic gatekeeping reshapes behavior, and why every regulated industry, from insurance to permitting, just acquired a new reference case for what AI does to a system's throughput. The experiment is young. Its implications are not.

What Medicare actually tested

Per MarketWatch, the facts are these: Medicare is testing the use of AI to conduct prior authorizations, with the explicit goal of targeting wasteful medical spending — and the result was an immediate, large drop in medical services. Each clause deserves attention, because together they describe something healthcare policy almost never produces: an intervention whose effect was visible at once.

Prior authorization is the process by which payers require advance approval before agreeing to cover a procedure, test, or treatment course. It exists because payers, public and private alike, have long wanted a mechanism to screen out care they deem unnecessary before it happens, rather than argue about it afterward. Traditionally that screening has been performed by human staff working through submissions — a process famous for delay, inconsistency, and frustration on every side.

What Medicare has now done, on a test basis, is place an AI system at that gate. The logic is seductive. Prior authorization is rules-heavy, document-intensive, and high-volume — precisely the kind of structured drudgery modern AI handles readily. If the technology can sort defensible care from questionable care faster and more consistently than human reviewers, the savings could be enormous. The early returns, per MarketWatch, show the gate working exactly as gates do: change how the gate operates, and what passes through it changes too — immediately.

Why prior authorization is the perfect AI test case

Every industry has bureaucratic chokepoints, but prior authorization is a near-perfect laboratory for automation, and it is worth understanding why. The process involves high volume, repetitive structure, and formal criteria: a request arrives, is checked against coverage rules and clinical policy, and is approved, denied, or returned for more information. There is a paper trail, a defined vocabulary, and a decision at the end. These are the conditions under which algorithmic systems excel — even before the current AI wave, payers automated whatever parts of the process the technology of the day allowed.

Prior authorization also sits at a rare confluence of incentives. Payers want to control spending. Policymakers want to target waste. Employers want premiums contained. And patients, at least in theory, want protection from unnecessary — and potentially harmful — interventions. An AI system that could genuinely distinguish necessary care from unnecessary care would serve nearly everyone at once. That is the promise that made this experiment worth running.

The catch is that the task's apparent tidiness conceals its hardest edge: the difference between wasteful and merely uncertain is often a clinical judgment, not a clerical one. A system can be flawless at processing rules and still systematically wrong at the boundary cases where medicine stops being a checklist. That tension — between the clerical surface and the clinical depth — is what makes prior authorization both the perfect test case and the perfect cautionary tale.

The speed is the story

Plenty of AI deployments produce effects that analysts argue about for quarters. This one, per MarketWatch, produced a large drop in medical services immediately. That immediacy is the single most important fact in the story, and it deserves more attention than the size of the drop itself.

Bureaucracies are supposed to have inertia. Review boards, staffing constraints, and human queue times impose friction that changes only gradually. An AI-operated gate apparently does not. When a reviewer's capacity becomes effectively unlimited and its latency approaches zero, the throughput of the entire system changes in a single step: requests get screened faster, marginal requests get caught more often, and the volume of services surviving the filter adjusts within days rather than years.

The behavioral lesson generalizes far beyond healthcare. Utilization in any system responds to the width of the gate more than to the price of admission. Providers who might once have submitted a borderline request — knowing a harried human reviewer might wave it through — face a different calculus against an algorithm that applies criteria uniformly and instantly. Some of that change is the intended elimination of waste. Some of it may be providers self-censoring legitimate requests to avoid friction. The immediate drop contains both effects braided together, and telling them apart is the urgent analytical task that begins now.

Cost savings or access risk? The question the data can't settle yet

The drop in medical services will be read two ways, and both readings describe the same number. Read one: the AI gate is doing exactly what it was built to do — filtering out spending that adds cost without adding health. Read two: the gate is filtering out care, some fraction of which patients actually needed, with the savings purchased through deferred diagnoses, abandoned treatments, and quiet harm that claims data records as efficiency.

Which reading dominates depends entirely on composition — which services disappeared, for which patients, under which criteria. A drop composed mostly of low-value, duplicative services is a triumph of policy engineering. A drop that includes necessary procedures discouraged by aggressive algorithmic scrutiny is a failure wearing a spreadsheet. As of today, composition is the missing variable, and it is the one that will decide whether this experiment is remembered as a milestone or a warning.

There is sober precedent for concern in how gatekeeping behaves at the margin: screening systems tend to suppress some legitimate care alongside the illegitimate, because the friction that deters wasteful requests deters all requests. The question for Medicare's test is whether algorithmic precision can beat that historical pattern — whether an AI gate can be more surgical than a blunt one. Answering requires clinical outcome data, not just utilization counts. Until that data exists, the honest summary is uncomfortable: we know the gate closed fast. We do not yet know what it closed on.

The measurement problem: waste is easier to cut than to define

Buried inside this experiment is one of the hardest measurement problems in health policy: nobody can precisely define the waste being targeted. Estimates of unnecessary spending in American healthcare have circulated for years, but they rest on judgment calls about what counts as appropriate care — judgments that vary by specialty, region, guideline, and philosophy. Waste is not a line item in any database. It is an inference, and reasonable people infer different amounts.

This matters because an AI system cannot target what it cannot operationalize. To build the system, someone had to translate the fuzzy concept of waste into concrete decision criteria — coverage rules, documentation requirements, evidence thresholds. That translation is where the policy actually gets made, quietly, in configuration rather than legislation. The immediate drop in services tells us the criteria bite. It does not tell us whether they bite in the right places.

There is also an asymmetry every regulator should internalize: the two kinds of error are invisible in different ways. A denied unnecessary service produces a saving that shows up in the accounts. A denied necessary service produces an absence — a screening that never happened, a treatment quietly foregone — that shows up, if at all, in someone's worsened health years later. Systems optimized on visible metrics will systematically overvalue the savings and undervalue the harm. Any serious evaluation of this test must therefore measure what did not happen — the hardest thing in the world to measure, and the most important.

What this signals for AI ROI across regulated industries

If the early results hold, Medicare's test is not just a healthcare story. It is a proof of concept for AI's most valuable and most controversial role in the economy: the automated gatekeeper. Every regulated industry has prior-authorization analogues — insurance underwriting, building permits, tax audits, benefits eligibility, customs clearance, licensing — all high-volume, rules-based approval processes that consume administrative labor and shape economic behavior at scale. A demonstrated case of AI moving utilization immediately in a program as consequential as Medicare will be studied by every agency and every compliance department that operates a gate.

For companies, the signal cuts two ways. As operators of gated processes, they now have evidence that AI deployment can move core metrics fast — a finding consistent with what we described in Enterprise AI: From Pilot Projects to Profit Centers: the winners of this era are those who move from experiments to production systems embedded in core operations. But as the subjects of other people's gates, the same evidence is a warning. An AI-run approval process at a supplier, insurer, or regulator can change a company's throughput overnight, in ways no vendor strategy anticipates.

That is why finance chiefs and strategy teams are likely to treat today's news as an ROI datapoint of unusual clarity, alongside the spending patterns we mapped in Corporate AI Budgets: Where the Money Actually Goes. Utilization that moves this fast turns AI from a cost-line experiment into a first-order business variable — the kind that belongs on the same dashboard as pricing, headcount, and capital allocation.

The guardrails this experiment still needs

An AI system that gates medical care is a system that needs more than a success metric. Whatever one makes of the results so far, the oversight architecture around the test will determine whether it becomes a model or a cautionary tale. Several guardrails are non-negotiable in any deployment of this kind.

First, appeal rights that are real. A gate that decides instantly must be contestable promptly, with human review available on timelines that respect clinical urgency — an appeal process slower than the disease it concerns is not a safeguard but a fig leaf. Second, transparency about criteria: providers and patients cannot comply with rules they cannot see, and secret standards breed defensive under-utilization, where providers avoid entire categories of legitimate requests to escape friction. Public agencies may even find that inspectable, open models carry a governance premium, a dynamic we explored in Open-Source AI Models: What They Mean for Business Adoption. Third, outcome monitoring rather than utilization monitoring alone: the test must track what happens to patients whose services disappeared, because savings that convert into sickness are transfers, not efficiencies. Fourth, disparity auditing, because gatekeeping tends to fall hardest on those with the least capacity to fight it. Fifth, sunset and review provisions — a test must remain allowed to fail.

None of this is anti-AI. It is the difference between automation and abdication, and the presence or absence of these safeguards — together with the hidden costs we cataloged in The Real Cost of AI: What Enterprises Underestimate — will say more about the experiment's meaning than the size of the initial drop ever will.

Why Washington is watching utilization, not just technology

Medicare's test did not happen in a vacuum. The program operates under relentless fiscal pressure, with spending trajectories that trouble budget officials across the political spectrum, and waste has long been the politically palatable target — the promise of painless savings that lets everyone defer the harder conversations about benefits, taxes, and reimbursement. An AI system that targets waste is therefore not merely a technology story; it is a political instrument, offering officials a way to claim savings without voting to cut anything specific.

That framing explains both the enthusiasm and the danger. If the technology is seen to work — in the crude sense that services drop and spending eases — the temptation to scale it will be immense, and the pressure to ask about composition will be correspondingly weak. Political incentives reward the visible number, and the visible number is the drop. The invisible numbers — the foregone care and its delayed costs — do not appear on any slide in any hearing, which is precisely why they require deliberate institutional effort to see.

There is a deeper significance, too. Economists continue to debate AI's aggregate economic footprint, but here is a domain where the effect arrived overnight and is measurable: services rendered, approvals processed, spending deferred. A technology whose macro impact remains contested just moved a concrete, consequential number — immediately. That contrast between macro ambiguity and micro specificity is exactly why this story will echo far beyond health policy, and why both AI's advocates and its skeptics will be citing Thursday's news for a long time.

The precedent problem: fast results invite fast copying

Perhaps the most underappreciated risk in Medicare's experiment lies not within the program but around it: imitation. Health policy is a copycat industry, and private insurers, state programs, and health systems are reading the same headlines. If AI prior authorization produced an immediate, large drop in services in Medicare, the commercial logic for every other payer to deploy something similar is obvious — similar technology, similar vendors, similar incentives. Adoption could cascade through the insurance system far faster than evaluation can keep up.

This is the classic pattern of technology diffusion in regulated markets: pilots run in relative obscurity, results land, and replication outruns understanding. The problem is that each copy inherits the original's unknowns — the unresolved question of what the drop actually consists of — while adding local variations in criteria, oversight, and patient populations. An error embedded in a widely copied template becomes a systemic error, distributed across vast numbers of approval decisions before anyone completes a proper evaluation.

For regulators, this creates a sequencing dilemma. Clamp down early and they stand accused of blocking savings and innovation; wait for evidence and the evidence may arrive only after deployment is effectively irreversible. The second failure mode is historically the more common one, because the constituency for measured caution is always smaller than the constituency for savings. Whether Medicare's test becomes a carefully evaluated pilot or the first domino in a wave of unexamined adoption will be decided not by the technology, but by whether evaluation infrastructure gets built faster than imitation spreads.

What happens next

The immediate agenda is set by the missing data. Watch for details on the test's scope: which service categories the AI reviews, what criteria it applies, and how edge cases and appeals are handled. Watch for composition analysis of the drop — whether independent clinicians and researchers can establish whether the disappearing services were predominantly low-value, or not. And watch for patient-level outcome tracking, the only evidence that can settle the savings-versus-access question honestly.

On the policy side, scrutiny will arrive from multiple directions. Patient advocates will demand transparency and appeal guarantees; fiscal hawks will push to expand whatever appears to save money; and other payers will move toward replication regardless of what Medicare concludes, because competitive pressure does not wait for peer review.

Per MarketWatch, the core facts are simple: Medicare tested AI-driven prior authorization to target wasteful spending, and medical services dropped immediately. What that means — efficiency or exclusion, innovation or abdication — depends on details that have not yet emerged. One thing is already clear: the debate over whether AI belongs in consequential bureaucratic decisions is over. The questions now concern rules, evidence, and accountability — and they arrive at a pace that, as of Thursday, everyone finally understands.