Washington ran two opposite AI rulebooks in the same week: the FDA's TEMPO pilot puts four generative-AI devices into patients' homes without marketing authorization, and on Sept 15 CMS widened the ACCESS model they ride on to 160 organizations covering three in four Medicare beneficiaries — while the same technology, pointed at claims instead of patients, is approving less than Medicare Advantage in six WISeR states and Congress is trying to shut it down
Today's story is not a single rule taking effect; it is a structural swap. For generative medical AI, the gate that actually decides whether a product reaches a patient has moved from FDA premarket review to CMS payment design. The FDA's TEMPO pilot lets four companies hand products to patients with no marketing authorization, on the condition that they ride inside the CMS ACCESS model. On September 15 CMS widened ACCESS with four new tracks — heart failure, COPD, substance use disorder and tobacco cessation — across 160 participating organizations, and reckons three in four Medicare beneficiaries now qualify for at least one track. The other half of the same government is running AI prior authorization in six states, where early reports put approval rates below Medicare Advantage. Accelerating toward the patient, tightening at the till — same technology, same department.
01 — Top Stories
TEMPO: four generative-AI devices reach Medicare without authorization — and on Sept 15 CMS widened the channel they travel through to 160 organizations
The FDA's TEMPO pilot — Technology-Enabled Meaningful Patient Outcomes — lets digital health companies put generative-AI devices in patients' hands before marketing authorization, provided the product is delivered inside the CMS ACCESS chronic care model and the company collects real-world evidence. All four slots are filled: Dexcom, Cadence, and the two behavioral health firms added on August 24, Limbic and Sondermind (MedTech Dive, 2026-08-24). STAT on September 3 framed it plainly as a path to patients that precedes authorization. Nine days later, on September 15, CMS added four ACCESS tracks — heart failure, COPD, substance use disorder (opioid, alcohol and other), and tobacco cessation, plus extended bone, joint and mobility support — launching spring 2027 with 160 organizations at the gate.
Two agencies are posting each other's authority as collateral. The FDA can release unauthorized devices because ACCESS pays on results — Fierce Healthcare describes it as a ten-year value-based program that reimburses on patient outcomes rather than service volume, so "no effect, no money" stands in for premarket evidence. CMS can route the devices to Medicare because the FDA is nominally still watching. The trouble is that each safeguard is conditioned on the other: strip out the FDA's evidentiary threshold and the only remaining gatekeeper is a payment model that does not start until spring 2027 and will not produce outcome data until later still. Limbic's CEO Ross Harper stated the bargain without varnish: "Existing regulatory frameworks were not written for generative AI in healthcare — they predate that technology."
"Three in four Medicare beneficiaries qualify" is a CMS estimate of eligibility, not enrollment; the "34 million beneficiaries" figure is Limbic describing its addressable market, not its user base. TEMPO has published no safety-event reporting format or exit criteria, and "payment on outcomes" will not yield auditable data until the new tracks start in spring 2027. The STAT piece is paywalled; this report cites only its publicly visible headline and lede.
The FDA's generative-AI device discussion paper: a two-axis risk frame and "competency" evaluation, with comments closing October 19
On August 18 the CDRH Digital Health Center of Excellence published its discussion paper on regulating generative-AI medical devices, putting 26 questions to the field, with comments due October 19 under docket FDA-2026-N-7874. At its center is a two-axis risk frame: how independently the device acts on one axis, how severe the harm from relying on a wrong output on the other — a gradient that, as Arnold & Porter reads it, separates a non-directive risk score from an instruction to act. Premarket evaluation would move to a "competency" model borrowed from how clinicians are credentialed: non-clinical benchmarking of clinical knowledge, analytical capability, safety behavior and generalizability, followed by clinical confirmation. Postmarket, the paper floats periodic re-benchmarking, sample-based clinician review and performance-drift monitoring.
The paper's most consequential line is a question, not an answer: the FDA asks whether greater uncertainty could be tolerated at authorization if paired with robust postmarket monitoring. That sentence is TEMPO's legal theory, and the two documents only make sense read together — the paper asks in principle what the pilot is already doing in practice. FDA digital health associate director Rick Abramson put it less tentatively at a public event on September 10: "The evidentiary standard for FDA authorization of generative AI tools will change." Which makes the October 19 deadline less a technical consultation than the last public chance to object to the principle that evidence can arrive later.
A discussion paper carries no legal force — it is not draft guidance, still less a rule, and the FDA has made no commitment to revise along the lines of the comments. The document does not state how many generative-AI devices the agency has authorized to date, and this report could not source that figure from the primary text.
The other half of the same CMS: WISeR runs AI prior authorization in Medicare, approving less than Medicare Advantage
The WISeR model — Wasteful and Inappropriate Service Reduction — began in 2026 across Arizona, Ohio, Oklahoma, Texas, New Jersey and Washington, applying AI-assisted prior authorization to roughly a dozen procedures. Early implementation reports collected by the Center for Medicare Advocacy describe a system physicians find more cumbersome than what preceded it, care within coverage guidelines going unauthorized, decisions occasionally breaching federal deadlines, and approval rates below Medicare Advantage; Ohio physicians worry they may stop offering complex treatments that confuse the model. KFF's policy analysis sets it against the wider frame of prior authorization and consumer protection.
Lay WISeR beside TEMPO and the day's real subject appears. One agency: when AI is handed to the patient, evidence may come later; when AI is used to withhold payment from the patient, the evidentiary bar is just as low — but the person carrying the risk has changed. Congress's answer is the Doctors Not AI Act, introduced by Rep. Greg Landsman on September 1: AI may not make the final call on a claim, a licensed professional must make any medical-necessity denial, and insurers must disclose AI use to patients and staff. It is a statutory line insisting a human sits at the end — precisely the opposite of TEMPO's loosening on autonomy.
"Approval rates below Medicare Advantage" comes from an advocacy group's compilation of early reports and Washington Post coverage, not a CMS audit; no denominator or time window is published, and CMS has issued no formal WISeR evaluation. The Doctors Not AI Act is an introduced bill only, not through committee.
Five states legislated AI out of therapy this year; the same year, the federal government routed AI voice CBT into Medicare
By the Transparency Coalition's count (2026-07-27), eleven states enacted fourteen health-AI laws in 2026. Seven — Alabama, Colorado, Georgia, Illinois, Iowa, Utah, Washington — govern AI in utilization review, requiring decisions to rest on the patient's own history and clinical circumstances, with human professional review of denials. Five — Colorado, Maine, Rhode Island, Tennessee, Vermont — require that mental health services be delivered by licensed professionals rather than AI systems. Individual provisions include Iowa's HB 475, effective August 1, obliging clinicians to disclose verbally that AI is recording or transcribing before an appointment, and Utah's SB 150, effective March 24, barring AI from replacing direct practitioner–patient interaction. Becker's Payer Issues carries a parallel state-by-state list.
This is the sharpest regulatory collision on today's board, and nobody is acknowledging it in public. Limbic's Unpacked delivers 20-minute CBT sessions by AI voice agent over scheduled phone calls, with a clinician supervising — and Limbic already operates in a dozen US states and the UK. Any overlap between that footprint and the five states produces a concrete question: federal TEMPO permits an unauthorized AI therapy product, while state law says therapy must be delivered by a licensed professional. Which prevails is unresolved, and the density of clinician supervision is probably the only variable that decides where the line falls.
This report could not obtain a complete list of Limbic's operating states, so the "overlap" is an inference from both sides' public statements rather than a documented legal conflict. State definitions vary widely — whether "delivering mental health services" reaches a clinician-supervised AI session is not settled in identical language anywhere. The count of laws comes from an advocacy organization's tally, not a line-by-line check of state legislative records.
The 2027 Medicare physician fee schedule drew more than 40,000 comments; APM participants face a cut of about 1.19%
Holland & Knight's September 22 policy roundup records that the comment window on the 2027 MPFS proposed rule closed September 14 with more than 40,000 submissions. The proposal cuts payment roughly 1.19% for alternative payment model participants and 1.68% for other clinicians; organized medicine's objections concentrate on a proposed 50% reduction for same-day E/M services, new restrictions on remote monitoring (RPM and RTM), and the administrative weight of MIPS and mandatory prior authorization. The same roundup notes Medicare physician payment has fallen about 33% in real terms since 2001.
The remote monitoring restrictions are where this item actually touches AI. RPM and RTM are the codes through which nearly every home-based AI device currently gets paid; ACCESS's outcome payments do not begin until spring 2027, and until then the cash flow of AI chronic-care products still hangs on those legacy codes. CMS building a new outcome-based channel while tightening the old per-service ones pushes the whole sector toward ACCESS — not by accident, but along the line drawn in July when CMS floated a temporary "Software as a Medical Service" payment category, dissected in detail by Wilson Sonsini.
These are proposed figures; the final rule normally lands in early November, and both the cuts and the RPM/RTM provisions can move. The STAT article is paywalled; this report cites only its publicly visible headline and lede.
NIST and NIBIB issued a medical metrology and standards RFI on September 18, closing November 30
NIST, within the Commerce Department, together with NIBIB, published a Federal Register Request for Information on future needs for medical metrology and standards across medical imaging, devices, diagnostics and therapy, docket NIST-2026-0133, with comments due November 30, 2026 at 11:59 p.m. ET. Its background states flatly that "with the advent of rapidly advancing medical technology and artificial intelligence (AI), the future requirements for medical metrology and standards may change considerably," and it lists one topic directly on point: suggested changes to the metrology and standards process that would allow improved, cost-effective healthcare as technology changes rapidly and AI is incorporated.
There is no drama in this item, but it supplies the piece every other story here is missing. The FDA's competency model needs benchmarks; ACCESS's outcome payments need measured outcomes; the "human review" state laws require needs an auditable trace. All three presuppose an agreed measurement standard, and the existence of this RFI is itself the admission that none exists. Until that infrastructure is built, "payment on outcomes" is a cheque written in units nobody has defined.
An RFI solicits views; it carries no legal force and commits to no deliverable. Its treatment of AI is contextual rather than a set of specific questions about AI-enabled diagnostics.
The EU AI Act's medical-device obligations slip to August 2, 2028 — but the transparency article took effect on August 2
Per Gibson Dunn's reading of the Digital Omnibus agreement, provisional political agreement was reached on May 6, 2026 and confirmed by member state representatives on May 13. Annex III stand-alone high-risk systems — recruitment, credit scoring, law enforcement, education, borders — must comply by December 2, 2027; Annex I, meaning AI embedded in regulated products, which covers medical devices, machinery and vehicles, moves to August 2, 2028. The agreement also lets the Commission narrow AI Act requirements where sectoral legislation already imposes equivalents, reducing duplication. Article 50 transparency obligations, however, stayed on their original August 2, 2026 date, with a four-month watermarking grace period for existing systems to December 2, 2026; the deadline for establishing regulatory sandboxes slips to August 2, 2027.
For device makers shipping across the Atlantic this reads as a two-year cushion — but the cushion covers the quality management and technical documentation stack that comes with high-risk classification, not disclosure. Article 50 is already in force, meaning any generative-AI product interacting directly with patients in the EU must already tell people they are talking to an AI. Set beside TEMPO the contrast is instructive: what the US relaxed was premarket evidence; what the EU declined to relax was notice.
This is cited from a law firm's analysis rather than the EU's own text; the Commission's power to narrow application is an enabling provision with no worked examples yet.
ARPA-H puts $62.7 million into cardiovascular agentic AI, while the FDA commissioner nominee faces a Senate vote on September 24
On September 9, ARPA-H launched ADVOCATE — Agentic AI-EnableD CardioVascular CAre TransfOrmation — with up to $62.7 million over four years, aiming at an FDA-authorized clinical AI system spanning patient monitoring, medication adherence, deterioration detection and a hospital supervisory platform. The same roundup records that the Senate HELP Committee scheduled its hearing and vote on Heidi Overton as FDA Commissioner for September 24 (committee page), and that on September 15 the FDA opened an expedited IND pilot for eight to ten drug companies, applications closing October 30.
ADVOCATE's phrasing is worth noticing: the agentic AI the government is funding itself is explicitly aimed at FDA authorization. Set against TEMPO's exemption for private products, that is two standards inside one executive branch — publicly funded systems go the full route, private ones may start first. Meanwhile, if confirmed, Overton inherits a desk on which TEMPO and the generative-AI discussion paper are both out and neither is settled. The agency is currently run by acting commissioner Kyle Diamantas, whose August 18 framing was that "the United States must lead in shaping how this technology is developed and used safely and responsibly."
The $62.7 million is a ceiling, not money disbursed, and ADVOCATE has named no awardees. The September 24 agenda can move with committee scheduling.
02 — Product Analysis
Limbic Unpacked
AI voice agent delivering CBT · Limbic (UK)
Function and position. An AI voice agent delivers 20-minute CBT sessions over scheduled phone calls; a clinician oversees each patient, reviews progress and session content, and receives real-time safety alerts to adjust care. MedTech Dive records the labeled contraindications as non-English speakers and those without telephone access; the indications are anxiety and depression. The buyer is not a hospital but a Medicare care provider inside ACCESS, and the company waives copays for ACCESS patients.
- Strength : it did not walk into TEMPO empty-handed. The submission ran past 100 pages with nine peer-reviewed clinical studies plus data protection and quality assurance protocols. It operates in a dozen US states and the UK supporting 850,000 patients — the most substantial deployment of the four.
- Concern : the nine studies concern Limbic's earlier product line, not the efficacy of Unpacked's voice-CBT format — which is exactly what TEMPO exists to collect, so what is missing is simply missing. The unanswered question is how thin clinician supervision must get for the unit economics to work; thinned to alert review alone, the distance between this product and what five states just banned collapses into a definitional argument.
Sondermind Adjunctive Care Application
Adjunctive digital therapeutic app · Sondermind (Denver, US)
Function and position. A smartphone app to reduce anxiety and depression in adults, positioned explicitly as adjunctive — used alongside existing therapy or medication rather than in place of it. The contraindications MedTech Dive records cover individuals experiencing suicidality, manic episodes, psychosis or PTSD — an exclusion list that is itself the risk strategy.
- Strength : "adjunctive" is a far cheaper regulatory position. Against the FDA's two-axis risk frame it barely moves along the autonomy axis, so even if postmarket monitoring surfaces problems the attributable surface is much smaller — and it does not run head-on into the five states' requirement that mental health services come from a licensed professional.
- Concern : the cheap position buys a fragile payment case. ACCESS pays on outcomes, and an app used alongside therapy can almost never claim the improvement statistically — and once suicidality, mania, psychosis and PTSD are excluded, the remaining population starts milder with less room to improve. This report could not obtain public efficacy data or user numbers for the product.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Limbic AI voice CBT, London |
Selected for TEMPO in August; operating in a dozen US states and the UK, supporting 850,000 patients, waiving copays for ACCESS patients. CEO Ross Harper frames it as an answer to the shortage of mental health professionals. | The moat is an NHS-scale deployment history plus nine peer-reviewed studies — the hardest currency there is in a regime that trades premarket authorization for postmarket evidence. The weakness is that it sits at the dangerous end of the autonomy spectrum, where the five-state bans run straight across its expansion path. |
| Sondermind Adjunctive behavioral health app, Denver |
Joined TEMPO alongside Limbic on August 24, positioned as an adjunct to therapy or medication, with a contraindication list excluding high-risk psychiatric populations. | The low-friction regulatory route: replace no clinician, collide with no state law. The cost is that in an outcomes-paid model it struggles to prove its own contribution, leaving the long-run moat thin. |
| Cadence／Dexcom Remote chronic care / CGM |
The two earlier TEMPO selections, accepted before Limbic and Sondermind, mapping onto the existing ACCESS tracks for hypertension, diabetes, chronic musculoskeletal pain and depression. | Dexcom arrives with an existing regulatory function and distribution — compliance capability is the moat; Cadence's chip is its care delivery network. Both share one exposure: the MPFS tightening on RPM and RTM, which is their cash flow until ACCESS starts. |
| Aidoc Imaging AI, Israel |
Former AMA president Jesse Ehrenfeld now works there and argued publicly on September 10 that clinical judgment cannot be displaced, citing subtle signals an autonomous system may miss, such as a tremor or a shift in clinical context. | The standard-bearer for the full-authorization camp. TEMPO's shortcut raises the relative cost of that route — but the day postmarket monitoring produces a casualty, an accumulated compliance record converts into pricing power. |
Today's competitive structure is a wager: you bet oversight keeps loosening, or you bet it loosens and then snaps back. TEMPO's four slots make "ship first, prove later" a real option; the companies taking it trade speed for an undefined liability exposure, while those walking the full authorization route buy time against a policy that only pays out after an accident. What settles the question is not any FDA document but when the first serious adverse event traced to a TEMPO device occurs, and whether anyone can then say where responsibility sits. CMS deputy administrator and chief clinical AI officer Stephanie Carlton's formulation on reporting measures — "We want to have exactly what is necessary to make sure the technology is working" — has no published measures attached to it yet.
04 — Taiwan Angle
(1) Taiwan has already written the class of product TEMPO permits out of its guidance. On May 29, 2026 the Ministry of Health and Welfare issued the Guidelines for Medical Institutions' Application of Generative AI under reference Wei-Bu-Yi 1151663164, covering public and private hospitals and clinics. As Lee and Li's analysis sets out, the guidance reaches record assistance, clinical decision support, administrative documentation and patient communication — but expressly excludes AI agent systems with autonomous decision-making on patient safety grounds, while requiring that clinicians retain final responsibility for clinical decisions, that anything potentially qualifying as a medical device be separately assessed, and that patients be told AI is involved. Which is to say: the shape Limbic's Unpacked takes — an AI voice agent running sessions under clinician supervision — is precisely the category Taiwan's current guidance sets aside.
(2) But guidance without force of law means the real leverage sits at reimbursement and device review — and Taiwan is still building both. The TFDA has stood up a smart medical device project office and runs an AI/ML medical device information and matchmaking platform, while the NHIA continues case-by-case benefit assessment; earlier reporting records NHI coverage of AI-assisted blood pressure stabilization in high-risk surgery, with items such as CT interpretation for intracranial hemorrhage under evaluation. This week's American lesson is blunt: the moment the FDA stepped back from the premarket threshold, the gatekeeper became CMS payment design. To avoid the same structural drift, Taiwan needs the NHIA's eventual AI coverage conditions to carry the clinical evidence bar inside them — because however strict the guidance reads, the fee schedule will decide which AI actually reaches patients.
(3) One concrete action: until October 19, Taiwanese device makers can still comment on the FDA's generative-AI framework. The consultation under docket FDA-2026-N-7874 is open to international parties, and what goes into the "competency" benchmarks will directly determine the standing of non-English clinical settings and non-US population data in future authorizations. The same question recurs in NIST's metrology standards RFI, open to November 30. These are among the few windows in which Taiwan can put generalizability on the table before the American framework sets.
05 — Further Reading
-
Considerations for the Regulation of Generative AI-Enabled Medical Devices — FDA CDRH (2026-08-18)
The source document under everything here. Its 26 questions show more clearly than any analysis which parts the FDA has not settled — and reading it before October 19 is what makes commenting possible.
-
Health Dose: September 22, 2026 — Holland & Knight (2026-09-22)
The most complete single roundup of the American health policy week — ARPA-H, the MPFS comment count, the NIST RFI, the nominations — each dated, which makes it usable as your own fact-checking baseline.
-
Examining the Potential Impact of Medicare's New WISeR Model — KFF
If you want to reason about harm from AI on the payment side, KFF's methodology is a better analytic starting point than an advocacy compilation, and carries fewer priors.
-
RFI on Future Needs for Medical Metrology and Standards — NIST/NIBIB, Federal Register (2026-09-18)
Under ten minutes to read, and it fixes one point: in a world where payment-on-outcomes is becoming the main regulatory instrument, there is still no agreement on how outcomes get measured.
-
EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines and Other Key Changes — Gibson Dunn
The clearest account of the gap between Annex I and Annex III timelines. Any team shipping devices on both sides of the Atlantic needs this schedule rather than the headline's "delayed two years."
06 — References
- CMS Adds Medicare Options to Help People Manage Substance Use Disorder, Heart Failure, and Other Common Chronic Conditions. CMS Newsroom, 2026-09-15. cms.gov
- CMS extends ACCESS model to more chronic conditions as agency officials eye cross-payer alignment. Fierce Healthcare, 2026-09-16. fiercehealthcare.com
- ACCESS Model Participants. CMS Innovation Center. cms.gov
- FDA adds two behavioral health firms to TEMPO pilot. MedTech Dive, 2026-08-24. medtechdive.com
- Limbic selected to FDA pilot, will test voice AI therapy with Medicare enrollees. Fierce Healthcare, 2026-08-26. fiercehealthcare.com
- FDA pilot offers generative AI medical devices a path to patients before they are authorized. STAT News, 2026-09-03（付費牆／paywalled）. statnews.com
- 4 companies sign on to FDA's digital device pilot. Becker's Behavioral Health, 2026-09. beckersbehavioralhealth.com
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices. U.S. Food and Drug Administration, 2026-08-18. fda.gov
- Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. FDA Digital Health Center of Excellence, 2026-08-18. fda.gov
- FDA Seeks Public Feedback on Regulatory Approach for Generative AI-Enabled Medical Devices. Arnold & Porter Advisory, 2026-09. arnoldporter.com
- FDA in Flux — September 2026 Newsletter. Mintz, 2026-09-17. mintz.com
- Holland & Knight Health Dose: September 22, 2026. Holland & Knight, 2026-09-22. hklaw.com
- Early Reports on WISeR Model Are Troubling. Center for Medicare Advocacy, 2026-03-26. medicareadvocacy.org
- WISeR (Wasteful and Inappropriate Service Reduction) Model. CMS Innovation Center. cms.gov
- Examining the Potential Impact of Medicare's New WISeR Model. KFF. kff.org
- State lawmakers have passed new laws regulating the use of AI in health care. Transparency Coalition, 2026-07-27. transparencycoalition.ai
- 7 AI health insurance state laws passed in 2026. Becker's Payer Issues, 2026. beckerspayer.com
- States Continue Efforts to Regulate AI in Healthcare: A Review of Legislation Passed in 2026. Holland & Knight, 2026-05. hklaw.com
- Congress Returns With Sweeping AI Bills Targeting Superintelligence, Health Insurance. PYMNTS, 2026-09-08. pymnts.com
- Request for Information on Future Needs for Medical Metrology and Standards for Medical Imaging, Devices, Diagnostics and Therapy. Federal Register, 2026-09-18. federalregister.gov
- NIST/NIBIB Symposium on Medical Metrology and Standards for American Healthcare and Commerce. NIST, 2026-09. nist.gov
- EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes. Gibson Dunn, 2026. gibsondunn.com
- At CTA event, federal officials outline AI ambitions as clinicians debate risks. Fierce Healthcare, 2026-09-10. fiercehealthcare.com
- CMS signals intent to revamp how it pays for clinical software, AI. STAT News, 2026-07-16（付費牆／paywalled）. statnews.com
- CMS Proposes Payment Frameworks for "Software as a Medical Service". Wilson Sonsini, 2026. wsgr.com
- Nomination of Heidi Overton to be Commissioner of Food and Drugs. U.S. Senate Committee on Health, Education, Labor & Pensions, 2026-09. help.senate.gov
- 醫療機構應用生成式人工智慧指引（115 年 5 月 29 日衛部醫字第 1151663164 號函頒）. 衛生福利部, 2026-05-29. mohw.gov.tw
- 衛福部頒布「醫療機構應用生成式人工智慧指引」. 理律法律事務所, 2026. leeandli.com
- 食品藥物管理署智慧醫療器材專案辦公室成立. 衛生福利部. mohw.gov.tw
- 智慧醫療器材資訊暨媒合平台. 衛生福利部食品藥物管理署. aimd.fda.gov.tw
- AI 診斷工具擬納健保，年底完成效益評估. 台灣智慧醫療創新整合平台. hst.org.tw