CMS Tests Direct Tech Payments for Chronic Care

More than 150 organizations have entered a Medicare experiment that can pay technology-enabled care companies directly when patients with chronic conditions reach defined health targets, a structure federal officials now describe as a fundamental shift in reimbursement. At a Washington health-AI event last week, Centers for Medicare & Medicaid Services Deputy Administrator Stephanie Carlton said the agency is trying to connect trusted technology, interoperable data, market access and payment rather than treating each as a separate policy problem.

The model, called Advancing Chronic Care with Effective, Scalable Solutions, or ACCESS, is not an across-the-board Medicare benefit and does not establish that any particular artificial-intelligence product improves health. It is a test of whether recurring, outcomes-linked payments can support technology-enabled care for conditions including diabetes, hypertension, chronic kidney disease, obesity, musculoskeletal problems, depression and anxiety. The distinction matters: CMS is purchasing accountable care processes and measured results, not endorsing a class of algorithms.

The timing also puts two federal experiments on parallel tracks. CMS is building a market for digital chronic-care services while the Food and Drug Administration is reconsidering how generative and agentic medical software should be evaluated. Together, the initiatives could determine whether healthcare AI moves from isolated pilots into routine care—and what evidence, data infrastructure and human oversight will be required when it does.

A payment model built around measured results

CMS has divided ACCESS into four clinical tracks: early cardio-kidney-metabolic disease, more advanced cardio-kidney-metabolic disease, musculoskeletal care and behavioral health. Its official payment schedule sets annual maximums that vary by track and by whether a beneficiary is in an initial or follow-on care period. Early cardio-kidney-metabolic care is capped at $360 initially and $180 later; the advanced track is $420 and $210; musculoskeletal care is $180; and behavioral-health care is $180 initially and $90 later. A $15 rural device supplement is available in the two cardio-kidney-metabolic tracks.

The mechanics place substantial risk on participants. CMS pays one-twelfth of the annual amount monthly during a care period, but those installments total no more than half of the maximum. The agency holds the other half for reconciliation after performance is measured. Full payment requires at least half of aligned beneficiaries to meet every required target, and the final amount can be reduced for weaker outcomes or substitute spending. The design therefore differs from simply reimbursing a device, app subscription or clinician encounter.

That structure creates a difficult balance. Fixed payments can give organizations room to combine remote monitoring, coaching, messaging and clinical review without billing for each interaction. But the maximums are modest once hardware, connectivity, licensed clinicians, patient support and reporting are included. Participating companies will have to show that software and automation reduce operating cost without allowing engagement or clinical quality to become superficial.

Data becomes the operating layer

ACCESS turns data collection into part of the service itself. Participants must establish baselines, report measures on a schedule and document outcomes within defined validity windows. Blood pressure, weight, body-mass index and patient-reported measures may need recent observations, while laboratory measures such as hemoglobin A1c can use longer windows depending on the track. Health information exchanges and CMS-aligned networks may help supply measures, reducing repeated data entry if records can be matched reliably.

CMS already operates several claims-data services that show what this infrastructure can look like. The BCDA interface supplies standardized Medicare claims to organizations in alternative payment models, generally updating data at least weekly. The beneficiary-facing Blue Button interface lets people authorize third-party applications to receive their Parts A, B and D claims. CMS also groups these and related resources in its developer tools catalog. ACCESS adds a different requirement: combining administrative data with timely clinical observations that can prove whether an individual target was reached.

That raises questions that ordinary app analytics cannot answer. A blood-pressure value must be tied to the correct beneficiary, device, date and care period; a missing value must not automatically be treated as clinical failure; and a software-generated trend should remain traceable to the underlying measurements. If those details are weak, a model can appear to reward health improvement while actually rewarding superior data capture. Interoperability, provenance and correction processes are therefore payment controls as much as technical conveniences.

The economics may favor automation

The payment levels explain why AI is central to the policy discussion even though ACCESS is technology-neutral. A company paid a few hundred dollars per beneficiary per year cannot build its model around frequent, high-cost clinician encounters. Automated outreach, risk stratification, documentation and measurement can extend human teams, especially when patients do not need an office visit. The policy test is whether that leverage produces better care or merely cheaper contact.

Market response has already exposed the tension. A detailed review of the participant slate found that several established digital-care companies did not join, with some executives saying the rates were too low to sustain their current service models. The same review identified concerns that a separate technology organization could fragment care unless information and treatment plans flow back to the patient’s primary-care team. CMS offers limited co-management payments for clinicians, but a fee cannot by itself guarantee coordination.

Scale could change the calculation. Commercial insurers representing a large share of covered Americans have signaled interest in aligning with the model, which may reduce the cost of operating different measurement and payment systems for every payer. That commitment is not the same as a binding nationwide contract, however. Companies will still face different benefit designs, state rules and data-access arrangements. ACCESS can demonstrate a common structure; it cannot make the surrounding market uniform.

FDA rethinks the evidentiary threshold

While CMS tests payment, the FDA is asking what proof should be required when medical software generates new content or takes actions. In an August discussion paper, the agency proposed assessing generative-AI devices along two dimensions: how autonomously the system operates and how much harm an incorrect output could cause. It also outlined competency testing that combines nonclinical benchmarking with clinical confirmation, plus postmarket monitoring for systems that may change or encounter new conditions after deployment. Public comments are due Oct. 19.

The agency’s current device list includes AI-enabled products that met applicable premarket requirements, but the FDA warns that the list is not comprehensive. It plans to identify products that use foundation models or large language models more explicitly. That work is important because a model embedded in a diagnostic device presents different risk than software drafting a message, prioritizing outreach or recommending a next step in a chronic-care workflow. A single “AI-enabled” label cannot substitute for intended use and clinical context.

Evidence from ambient documentation illustrates both the opportunity and the limits. A five-center JAMA study of 8,581 clinicians found that adoption of AI scribes was associated with about 13 fewer electronic-record minutes and 16 fewer documentation minutes per eight scheduled hours, along with a small increase in weekly visits. It found no significant change in after-hours record use, and the observational design could not establish causation. Workflow gains are measurable, but they are not evidence that patients’ blood pressure, kidney function or depression improved.

What the test can actually establish

Federal policy will also develop alongside state law. A current policy tracker counted dozens of healthcare-AI bills introduced in 2026, with enacted measures ranging from behavioral-health restrictions to patient-notification requirements. Those laws address legitimate local concerns, but they can create different disclosure, consent and professional-practice obligations for a service operating across several states. CMS payment eligibility and FDA device review will not automatically resolve those differences.

The most meaningful ACCESS results will therefore go beyond enrollment and technology adoption. CMS will need to show how many beneficiaries complete a care period, how often targets are met, whether improvements persist, how total Medicare spending changes and whether outcomes differ by geography, race, disability, income or digital access. It should also track patient exits, adverse events, complaints, clinician workload and the share of records that require correction. Strong average results could otherwise conceal a model that works mainly for people already equipped to use connected tools.

For technology organizations, the test is equally concrete. They must produce clinical evidence under real payment constraints, exchange trustworthy data with clinicians and patients, and explain where automation influences a decision. For regulators, the challenge is to protect beneficiaries without freezing software at the moment it is cleared. For CMS, success means demonstrating that a direct technology payment can improve health rather than simply create another vendor channel.

The federal government is moving from debating whether digital chronic-care tools belong in Medicare to specifying how they will be paid, measured and supervised. ACCESS can supply evidence that the market has often lacked, but only if its data distinguish engagement from health improvement and its oversight distinguishes useful automation from unverified clinical judgment. The decisive result will not be how much AI the program uses. It will be whether beneficiaries receive safer, more coordinated and measurably better care.