Medicare now pays some providers up to $180 a year to treat a traditional Medicare beneficiary’s depression or anxiety, without setting a required number of visits. The payment comes through the ACCESS model, a voluntary Innovation Center program that began on July 5, 2026. Half of it is withheld and reconciled against how the provider’s patients score on standard symptom questionnaires. CMS’s own design documents describe “technology-led, clinician-supervised approaches” in which software delivers treatment directly under a clinician’s oversight.
One participant is already operating that way. Limbic Care P.C. is listed by CMS as an ACCESS behavioral health participant. Its clinicians direct care in Unpacked, a service from Limbic Inc. that delivers structured cognitive behavioral therapy through telephone sessions with an AI voice agent, and the FDA has admitted Unpacked to TEMPO, a pilot that lets selected digital devices be used inside ACCESS without premarket authorization. Licensed clinicians remain responsible for each patient. The FDA says it has not yet evaluated the device’s effectiveness.
Medicare is not paying an AI clinician. It is paying a licensed provider organization a fixed amount for a patient’s condition and outcome, on terms that do not ask whether a human or a model delivered the sessions. That is a narrower change than paying software directly, and a larger one than another digital-health billing code.
What ACCESS pays for
ACCESS, short for Advancing Chronic Care with Effective, Scalable Solutions, runs for ten years across four clinical tracks. The behavioral health track covers depression and anxiety in beneficiaries enrolled in Original Medicare Parts A and B; Medicare Advantage enrollees are not eligible. CMS’s participant list, last updated September 15, names 162 organizations. Ninety-seven take part in the behavioral health track, 19 of them in that track alone.
The unit of payment is a patient’s condition over a 12-month care period. For behavioral health, the maximum annual allowed amount is $180 in a patient’s initial period and $90 in a follow-on period of continued management. Those figures include the beneficiary’s 20 percent coinsurance, which participants may waive. CMS pays monthly, capped at half of the Medicare share, and withholds the other half for reconciliation after the care period ends.
Reconciliation is measured across the provider’s whole panel, not patient by patient. A patient with a baseline PHQ-9 depression score of 10 or higher meets the target with a five-point reduction; a baseline GAD-7 anxiety score of 10 or higher requires a four-point reduction. During the model’s first 18 months, a participant earns the full payment if at least half of its aligned beneficiaries complete the care period and meet every required target. Below that, payment is scaled down in proportion. In CMS’s own example, a participant at 40 percent attainment earns 80 percent of the full amount. The outcome reduction is capped at half of the full payment, and CMS says participants that fall short of the minimum threshold may be subject to termination. A separate reduction, capped at 25 percent, applies when aligned patients start certain substitute services with another provider, including a new psychiatric diagnostic evaluation, initial collaborative care management, or Medicare’s existing digital mental health treatment codes.
What the payment does not depend on matters as much. The RFA and CMS’s technical FAQs describe no required number of clinician encounters. Each monthly claim attests to “active care delivery,” which CMS defines as ongoing services including patient engagement, monitoring, and outcome reporting, not as a count of sessions. Services may be delivered “in-person, virtually, asynchronously, or through other technology-enabled modalities, when permitted under applicable law and clinically appropriate.” And the payment replaces billing instead of adding to it. A participant and its affiliates may not submit ordinary fee-for-service claims for an aligned beneficiary during the care period, and CMS says Medicare’s claims systems will suppress them automatically. In the FAQ’s words, ACCESS is “an alternative to fee-for-service (FFS) billing by an organization for a given beneficiary, not an add-on to it.”
That structure closes off the most obvious cost problem with paying for AI in healthcare: a provider billing for an AI agent while continuing to bill for the human service it was meant to replace. Under ACCESS there is one fixed amount per condition, and the provider keeps whatever it does not spend producing results.
The price assumes software
At the initial-period rate, the ceiling is $15 a month per patient. CMS’s RFA is explicit about the kind of care that rate anticipates. It describes technology-enabled organizations that range from clinician-led, technology-augmented practices to technology-led, clinician-supervised models, and offers as one example “a virtual health care provider group overseeing use of an FDA-cleared cognitive behavioral therapy application.” It says the outcome-based design “aligns with broader trends in AI-enabled services pricing, where low marginal costs make at-risk payment models more viable,” citing a 2024 Andreessen Horowitz essay on AI and outcome-based pricing.
The model was written with software-delivered treatment in view. What it requires is a licensed organization to stand behind that treatment.
Where the AI comes in
ACCESS requires that any medical device used in the model be legally marketed for that use, or otherwise be “subject to FDA enforcement discretion.” The last clause is the opening. Under the FDA’s TEMPO pilot, for Technology-Enabled Meaningful Patient Outcomes, manufacturers can ask the agency not to enforce premarket authorization, investigational device, and informed-consent and IRB requirements when their device is offered to or by ACCESS participants. In exchange, they collect and report real-world performance data. The FDA says it expects to select up to about ten U.S. manufacturers in each of the four ACCESS clinical areas.
As of the FDA’s August 21 update, four have been selected, two in behavioral health. SonderMind’s adjunctive care app is intended to reduce depression and anxiety alongside psychotherapy or medication. Limbic’s Unpacked goes further. The FDA lists its intended use as delivering “evidence-based psychological treatment programs for adults who are receiving a course of psychological talking therapy via an AI-voice agent,” specifically structured CBT for Medicare beneficiaries with clinically significant depression or anxiety, with real-time safety flagging and outcome monitoring to support clinical oversight.
The FDA’s participant page carries one sentence that covers every device on it: their effectiveness for the intended uses in the pilot has “not yet been evaluated by the FDA.”
Limbic said the service would begin rolling out on August 27. Its service page says Unpacked is covered by Medicare for eligible beneficiaries with the copay waived, and that licensed clinicians at Limbic Care direct each patient’s care, while the sessions themselves are with the AI agent. Patricia Taylor, a clinical psychologist who heads clinical operations at Limbic, stated the company’s position in its announcement: “This is not AI replacing a therapist. Unpacked is a clinical service, and a clinician remains responsible for every patient.”
The two pilots divide the work of evaluation in an unusual way. The FDA has deferred its judgment on whether the device is effective. CMS ties part of the provider’s payment to whether patients’ PHQ-9 and GAD-7 scores fall. For a TEMPO device used inside ACCESS, the payment model is carrying much of the effectiveness accounting that premarket review would otherwise have done first, though a panel-level threshold built on self-reported symptom scales is a coarser test than a controlled trial.
What this does not authorize
None of this is Medicare reimbursement for an autonomous AI clinician, and the gap is structural.
Medicare pays an organization, never software. Participants must be Medicare Part B-enrolled providers or suppliers and must designate a physician medical director accountable for the quality of care. Every physician and non-physician practitioner “furnishing or supervising care” must be individually Medicare-enrolled and practice within state licensure and scope of practice. There is no billing code for an AI and no route for a software company to enroll as the treating party.
State law still governs who may deliver therapy. ACCESS permits technology-enabled delivery only “when permitted under applicable law.” Illinois’s Wellness and Oversight for Psychological Resources Act, in effect since August 1, 2025, bars a licensed professional from allowing artificial intelligence to “directly interact with clients in any form of therapeutic communication.” Its exceptions cover religious counseling, peer support, and public self-help materials, not therapy services. The federal model does not override that law, one of several state approaches to AI in mental health care, so a service built around AI-delivered sessions cannot run nationwide on ACCESS terms alone. Limbic says availability “varies by state.”
The patient population is bounded. The behavioral health track excludes beneficiaries with suicidal or homicidal ideation, psychotic features, bipolar I disorder, moderate to severe dementia, or severe eating disorders, among others, and requires participants to refer patients whose needs exceed the model’s scope. Unpacked’s FDA contraindications are wider still, adding bipolar II, personality disorders, primary substance use disorder, and anyone who does not speak or read English or lacks a phone. The AI-delivered portion of this care sits, by design, at the mild-to-moderate, lower-risk end.
And the regulatory status is provisional. TEMPO is a pilot of enforcement discretion, not an approval or clearance, and the discretion applies only while a device is used inside ACCESS.
What changes when Medicare pays for the condition
Medicare behavioral health has been paid by the encounter: a psychotherapy session, a psychiatric evaluation, a month of care management with time thresholds attached. Under that structure, software that does the work of a session earns nothing unless a clinician’s billable time sits behind it. ACCESS removes that link for the patients it covers. A provider can take one fixed payment per condition, deliver most of the contact through software, keep licensed clinicians in supervisory and escalation roles, and earn the full amount if half its panel reaches target.
That changes what pays. More patients per clinician, asynchronous care, and software-delivered sessions all become ways to protect margin instead of ways to lose revenue. A company that sold a mental health app to employers, or billed Medicare session by session, now has a Medicare contract structure that rewards substituting for clinician hours. Whether that substitution happens in practice depends on staffing ratios no participant has disclosed.
The market response so far is narrow, and the companies in it are not alike. The behavioral-health-only participants on CMS’s list include Headspace’s California medical group, three SonderMind entities, Mindoula, Innerwell, and Limbic Care. Most are provider organizations with no device in TEMPO, and the broader group of 97 is mostly multi-track medical and telehealth practices, a reminder that ACCESS is open to conventional providers too. Two firms pair an ACCESS participant with a TEMPO device. SonderMind’s is an adjunct to therapy or medication delivered by clinicians. Only Limbic’s is one in which software carries the treatment itself. One company and one adjacent case are an early template, not yet a category and not a market.
Consolidation has reached the edge of it. Sword, which describes itself as an AI health company and already sells an AI mental health product, announced on September 16 that it would acquire Headspace, whose medical group is an ACCESS behavioral health participant. Headspace says the acquisition closed on October 1. The announcement did not mention Medicare or ACCESS.
The case that little has changed
The strongest counterargument is that ACCESS is another Innovation Center payment experiment, and such experiments have an uneven record of becoming permanent. It is voluntary, limited to traditional Medicare, and built with a control group for evaluation. It pays provider organizations, not technology, and leaves licensure, supervision, and medical director accountability where they were. Its outcome terms are softer than “paid only for results”: the outcome adjustment cannot cut a participant’s payment by more than half, however few patients reach target. Read that way, it resembles Medicare’s collaborative care codes more than a new market, a bundled payment to a supervised team in which software is one tool.
The money points the same way. Up to $180 a year per patient will not build a company on its own. The model’s significance is less the dollars than the terms. CMS has built track-specific billing codes and FHIR-based outcome reporting that Medicare Advantage, Medicaid, and commercial plans can adopt. In February, CMS said payers covering more than 165 million people, among them UnitedHealthcare, Humana, CVS Health, Cigna, and Centene, had pledged to offer payment arrangements aligned with the model by January 1, 2028. A pledge is not a contract, and the deadline is more than a year away. Whether those arrangements materialize is what would turn a pilot into a market.
The counterargument is right about the boundary. Medicare has not created a payment for AI clinicians, and nothing in ACCESS suggests it intends to. It has created a payment that does not depend on who or what ran the session, provided a licensed organization stands behind the care and its patients’ scores move. For mild-to-moderate depression and anxiety in traditional Medicare, that is now the route by which AI-delivered treatment gets paid.
The next tests are on the calendar. A substance use disorder track is set to begin on April 1, 2027, alongside new tracks for heart failure, chronic obstructive pulmonary disease, and tobacco cessation, which will show whether the same logic reaches a higher-risk condition. The 50 percent attainment threshold applies only to the model’s first 18 months; CMS says it will rise in later participation years and may be adjusted. The FDA has said it will add TEMPO participants as it selects them. And CMS plans to publish a directory of participants’ risk-adjusted outcomes, the first public evidence of whether supervised, software-delivered therapy improves symptoms at the rate the payment assumes.