Q3 Health AI Tracker: Federal Agencies, California and Utah’s Sandbox Take Center Stage

Introduction:

Q3 2026 marked a shift from broad health AI policymaking to implementation. Federal agencies took concrete steps toward payment and regulatory pathways for clinical care, and California enacted several new laws that push the boundaries of state AI regulation across child safety protections for AI chatbots, health data privacy, clinical decision support, and independent AI audits. Of the more than 280 bills introduced this year, covering topics from AI chatbots to payer use of AI to transparency, 40 bills have been enacted across 24 states. Meanwhile, state AI regulatory relief sandboxes continue to evolve: Utah’s Office of AI Policy (OAIP) evidentiary expectations for participants, announced three new mitigation agreements and two “master agreements,” and signed six agreements with independent organizations serving as “third-party evaluators.”

At the federal level, activity picked up in Q3, driven by agencies maturing AI-related programs and seeking input to shape payment and regulatory pathways for AI in clinical care. The Advanced Research Projects Agency for Health (ARPA-H) for its Agentic AI-Enabled Cardiovascular Care Transformation (ADVOCATE) program, which aims to produce an FDA-authorized agentic AI system for cardiovascular care, and the Centers for Medicare and Medicaid Services (CMS) launched the first cohort of the Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) model in July before four new tracks beginning in April 2027. At the same time, agencies sought input on how AI in clinical care should be paid for and regulated. CMS’s ) and proposed rules include a new Medicare payment category for algorithmic software (Software as a Medical Service (SaMS)) and several Requests for Information (RFIs) on how Medicare should pay for technology and AI-enabled care. The Food and Drug Administration (FDA)’s Center for Devices and Radiological Health (CDRH) a discussion paper and request for feedback on the regulation of generative AI-enabled medical devices, proposing a two-axis risk framework and pre- and post-market monitoring approach.

Congressional action on AI remains limited, though a bipartisan proposal is advancing. The Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting (FRONTIER) Act, introduced by Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) and developed from the Great American AI Act discussion draft published this summer, would set a national framework for frontier AI developers that preempts related state requirements on developers, while expressly preserving states’ authority to regulate how AI is used.

This activity is unfolding within the Administration’s broader deregulatory AI strategy. In late September, President Trump signed an to establish an AI-powered government portal (America.gov) and convened tech leaders to sign the White House Accord on Super Intelligence, emphasizing a self-regulatory approach to governance.

The following sections provide additional detail on key developments from Q3 2026. For a full description of themes and activity, see our ; we will publish a full analysis of action in 2026 at the end of the calendar year.

Deep Dive: State Activity in Q3 2026:

California drove nearly all of the quarter’s meaningful state legislative activity:, referred to as “Adam’s Law,” was signed by Governor Newsom on September 10, 2026. It establishes one of the nation’s most expansive chatbot safety frameworks, updating child-safety provisions in California’s existing chatbot requirements (California , effective on January 1, 2026), and imposing more stringent age-assurance, auditing and parental-control requirements. In contrast to many other state chatbot laws, which impose enhanced protections when an operator knows or has reason to believe a user is a minor, SB 1119 requires operators either to determine a user’s age under California’s age-assurance framework or apply the statute’s child-protection requirements to all users.The law requires companion chatbot operators to conduct a comprehensive child risk assessment and child safety audits, implement a child safety policy and crisis response protocols, prohibit the companion chatbot from providing certain responses to children, and establish robust parental controls, among other safeguards. SB 1119 contains two versions of its child safety audit provision. The law aligns with the auditor requirements outlined in AB 1405, also enacted this session (described below). SB 1119 takes effect on January 1, 2027, including its age-assurance provisions; most other core provisions become effective on July 1, 2027.

• (signed September 30, 2026, effective January 1, 2027) makes two amendments to existing laws: 
  • First, it deems certain deployers of consumer-facing health care chatbots to be “providers of health care” under California’s Confidentiality of Medical Information Act (CMIA), the state of California’s mini-HIPAA (Health Insurance Portability and Accountability Act (of 1996)), if the chatbot helps the individual to manage their health information or support diagnosis, treatment, or management of a medical condition. A health care chatbot is “a generative artificial intelligence system with a natural language interface that” provides adaptive, human-like responses to user inputs; is marketed as facilitating or supporting health services to a consumer; and uses health care chatbot information to facilitate or support health service to a consumer. The law covers information about a consumer’s physical or mental health or wellness that the consumer provides directly to the chatbot, makes accessible to it, or that the chatbot collects, generates, or infers. This significantly expands the scope of CMIA, making many AI chatbots that interface between a user and their personal health records subject to certain privacy restrictions.
  • Second, the law requires health facilities, clinics, physician offices, and group practices to take reasonable steps to preserve clinicians’ independent professional judgment when clinical decision support systems, broadly defined, are used in patient care, essentially requiring a human in the loop at all times. In addition, these entities may not use AI to direct, guide, supervise or instruct unlicensed personnel in performing licensed clinical functions, or to perform any functions requiring a professional medical license.
• (signed September 30, 2026, effective January 1, 2027) establishes requirements for developers and deployers of AI-enabled clinical decision support systems to identify, mitigate and monitor the risk of biased impacts in health care programs and activities. The law requires developers to provide deployers with documentation regarding a system’s intended uses, training data, evaluation methods, known limitations, and foreseeable bias risks; it requires deployers to regularly monitor systems and take reasonable and proportionate steps to mitigate those risks. Notably, SB 503’s focus on mitigating biased impacts resembles some provisions of Colorado’s original AI Act (), which the Trump Administration targeted in a December 2025 executive order and which Colorado repealed and replaced this year (). While SB 503 is narrower, applying only to AI-enabled clinical decision support, we will watch to see if this law becomes subject to federal interest. (signed September 9, 2026) is the second state law enacted this year establishing a framework for Independent Verification Organizations (IVOs), following Connecticut ‘s mandate to establish a pilot IVO program in the state. SB 813 directs the California Government Operations Agency (GovOps) to establish a framework for IVOs, including application requirements and designation criteria for IVOs, criteria for determining whether an AI auditor qualifies as an IVO, procedures for suspending or terminating a designation, and stakeholder working groups composed of competitor AI engineers and AI safety experts to inform those standards. Designated IVOs must report annually to GovOps and the Legislature on their standards and methodologies.

On September 18, Governor Newsom issued , which accelerates the laws’ implementation by requiring GovOps to develop IVO application requirements, procedures, and criteria by May 1, 2027. It also requires GovOps to submit recommendations to the Governor by November 16, 2026 on whether to: require large frontier developers to embed designated IVOs onsite; require independent verification of required safety frameworks, transparency reports, and risk assessments; require frontier models to include an independently verified “kill switch,” and expand the definition of critical safety incidents to include loss-of-control incidents.

• (signed September 14, 2026) directs GovOps to establish an AI Auditor Registry on its website by January 1, 2029 and prohibits offering, selling, or conducting a “covered AI audit” without registration. accelerates the registry timeline, requiring GovOps to establish the Registry and begin the related agency actions by December 1, 2027.

Notably, Governor Newsom vetoed two bills that would have imposed significant restrictions on the clinical use of AI:

• , which was expected to pass, would have significantly restricted the use of AI tools in psychotherapy services. Most notably, it would have barred using AI to record or transcribe psychotherapeutic communications and sessions, or perform triage or screening unless the patient (or their legal representative) was informed of the AI use and its purpose and gave consent. In his veto message, the Governor said the bill was overly broad and would drastically limit clinicians’ use of AI tools that benefit care today and that its definition of “psychotherapy services” would capture general-purpose AI systems not designed to deliver that care.
• would have barred AI and clinical decision support developers and deployers from claiming that a clinician’s failure to override AI output cuts off their liability for harm. It also would have protected direct-care workers from retaliation based solely on overriding, or relying on, a clinical decision support system’s output.

Two other states enacted relevant laws this quarter:

• Hawaii (signed and effective July 14, 2026) joined many other states in legislating AI chatbots by providing protections for mental health and minor users. The law requires conversational AI operators to disclose to users that are interacting with AI, adopt protocols for responding to suicidal ideation and self-harm, including crisis referrals and evidence-based measurement, and refrain from representing the chatbot as a professional mental or behavioral health care providers. For users known to be minors, the law adds additional protections, including hourly break reminders, restrictions on engagement-maximizing rewards, and parental account-management tools, with annual reporting to the state beginning January 1, 2027.
• Illinois (signed July 10, 2026, effective January 1, 2028) became the second state law this year (after Indiana ) regulating payer downcoding. The law prohibits payers from using algorithms or automated processes to bypass the billing professional’s submitted information when downcoding a claim, requires all downcoding determinations to be made or reviewed by a person following American Medical Association Current Procedural Terminology guidelines, and prohibits downcoding based solely on diagnosis codes.

In August 2026, Utah’s OAIP evidentiary expectations for participants in its health care AI sandbox. The framework requires participants to demonstrate that their AI tools have the potential to support evidence of: (1) non-inferior quality of care, (2) bodily and psychological safety, and (3) public health benefits for Utah residents. To reduce barriers to market entry, the framework applies a lower evidentiary threshold at initial deployment, coupled with ongoing post-deployment monitoring of safety, performance, and adverse events and continuous quality improvement.

OAIP also three new regulatory mitigation agreements this quarter:

• Expect Fitness offers online, AI-driven pelvic floor physical therapy and is taking a phased approach. In the first phase, a licensed physical therapist reviews every treatment plan.
• August AI offers an AI health companion that may refill maintenance medications from a fixed list of non-controlled drugs, but cannot change doses, make substitutions or issue new prescriptions.
• Nolla Health, which is now operating, assesses mild to moderate acne from patient photos. It may issue first-time prescriptions and refills for topical treatments from a predetermined list of medications, which centers on Nolla’s own compounded formulations.

OAIP also signed “master agreements” with University of Utah Health and Intermountain Health that establish standing terms for systemwide AI pilots. OAIP will approve each proposed pilot through an addendum rather than a separate agreement. No pilots are active yet. The framework extends OAIP’s agreements beyond private companies to health systems.

Utah also announced agreements with six independent . This group includes four clinical and technical evaluators (Stanford Clinical Excellence Research Center, mpathic AI and Vega Health), two runtime attestation partners (Clarion AI Partners and Glacis Technologies), and the Coalition for Health AI, which will convene an evaluator network. Before OAIP approves a regulatory mitigation agreement, clinical and technical evaluators may test the AI tool for clinical errors and unsafe behavior, arrange expert clinical review, and assess privacy safeguards. Once a pilot begins, the evaluator may audit real-world activity against the participant’s reports. Runtime attestation partners produce tamper-evident digital receipts each time a specific safety check runs. Participating companies pay the evaluators, but evaluators report to OAIP, which retains authority to approve, continue, or terminate a pilot. This independent review supplements the participant self-reporting on which OAIP has relied to date.

Looking ahead, we expect Doctronic to seek a renewal of its regulatory mitigation agreement, which expires at the end of October. We will also watch for Legion Health, Expect Fitness, and August AI to launch, as well as University of Utah Health and Intermountain Health to propose pilot programs.

With most state legislative sessions concluded, we do not expect significant additional activity this year, except in states still in session or awaiting gubernatorial action. We will return next quarter with a roundup of the key themes across state-level AI activity in 2026; an updated list of all health AI laws enacted to date is available .

Deep Dive: Federal Activity in Q3 2026:

While state activity slowed considerably in Q3 as most legislatures adjourned, federal activity picked up. Agencies took a range of actions, including steps toward building payment and regulatory pathways for AI use in clinical care. Congressional action on frontier AI legislation remains limited. Federal highlights from Q3, organized by originating agency, are discussed in greater detail in the sections that follow.

Centers for Medicare and Medicaid Services (CMS)

ACCESS Model: On September 15, CMS that ACCESS will add four new clinical tracks, beginning on April 1, 2027: heart failure, chronic obstructive pulmonary disease (COPD), substance use disorders and tobacco cessation. The model will also extend longer-term support for chronic musculoskeletal conditions beyond the current 12-month period. With the expansion, CMS estimates that roughly three in four Medicare enrollees will qualify for at least one track. The expansion is a sign of the federal government’s ongoing interest in addressing chronic disease, paying for outcomes and encouraging technology-enabled care. We will be monitoring the qualifying conditions, outcome measures and substitute spend lists for the new tracks. For additional analysis on the ACCESS Model, please see Manatt Health client alert .

New Medicare Payment Category for Algorithmic Software: CMS released the on July 2 and the on July 14. In each, CMS proposed to create SaMS, a new Medicare payment category for software that supports clinical decision-making through algorithmic analysis. The OPPS rule would designate 36 codes as SaMS, including AI-based retinal imaging analysis and CT-derived fractional flow reserve. The PFS rule also seeks input on AI-enabled primary care and proposes crediting clinician use of AI in the Merit-based Incentive Payment System. This is CMS’s first effort to pay for clinical software as its own category. Payment would still be interim and set code by code, but the proposal is an early step toward a Medicare reimbursement pathway for AI. The proposed rules include several RFIs on how Medicare should pay for technology and AI-enabled care, including better understanding current use of clinical AI tools, considering appropriate payment, and coverage of clinical AI tools, including outcomes-based reimbursement methodologies and reporting and monitoring.

WISeR Records Release: In September, the Electronic Frontier Foundation (EFF) roughly 1,000 pages of records on the Wasteful and Inappropriate Service Reduction (WISeR) model, obtained through a Freedom of Information Act filed in March 2026. According to EFF, the records show authorizations or denials were provided well beyond CMS’s 72-hour response standard; vendor technical issues; thousands of denials in the model’s first three months; and a vendor payment structure tied to denials. The release adds to the scrutiny that has followed CMS’s launch of WISeR in January 2026.

CMS AI Strategy: On September 9, CMS Deputy Administrator and Chief Clinical AI Officer at a Consumer Technology Association health care event. She outlined a four-pillar AI strategy:

1. Building public trust.
2. Expanding data sharing and interoperability.
3. Clarifying pathways for AI market access and regulation.
4. Developing reimbursement frameworks for AI-enabled technologies.

Carlton described ACCESS as a first step toward defining AI’s role in managing total cost of care. Her remarks signal that CMS intends to use its payment models to test the value of clinical AI.

Medicaid Technology Procurement: CMS and the General Services Administration (GSA) their partnership fast-tracking Medicaid technology vendors onto the GSA Schedule. The number of vendors on the Schedule grew from 12 in February to 33 as of July. Several of them have pledged AI-enabled tools to states, including Conduent’s fraud detection tool and GDIT’s Medicaid policy enforcement tool. The partnership began as support for implementing Medicaid community engagement requirements. It now gives states a faster route to procure technology, including AI tools, ahead of the January 1, 2027 deadline for those requirements.

Food and Drug Administration (FDA)

Technology-Enabled Meaningful Patient Outcomes (TEMPO): In August, FDA added participants to its TEMPO pilot (which was launched in partnership with CMS’s ACCESS model). Two of the new participants use generative AI. Cadense Solutions offers a hypertension medication management tool, and Limbic offers an AI voice agent that delivers cognitive behavioral therapy. Their participation in the TEMPO pilots will generate real-world evidence that can be used to support their application for premarket authorization. FDA has not yet said whether it will expand the TEMPO pilot to include the new ACCESS tracks. More participants are expected.

Generative AI Device Discussion Paper: Following UpDoc’s Software as a Medical Device (SaMD) clearance in June (noted in Q2), FDA continues to signal how it may regulate generative AI-enabled devices. Its August proposes a two-axis framework that assesses both the degree and independence of a device’s activity, ranging from providing information to taking autonomous action, and the potential harm if a user relies on an incorrect output. Under this approach, regulatory scrutiny and evidence expectations would increase as a device becomes more autonomous and the consequences of error become more severe. For premarket review, FDA outlines a competency-based model that pairs high-volume benchmarking with clinical confirmation, which could range from retrospective testing and simulated or shadow deployment to clinician adjudication and prospective studies, depending on the device’s risk. The paper also considers whether FDA could accept greater premarket uncertainty in exchange for stronger postmarket oversight, including periodic re-benchmarking, clinician review of real-world outputs, and monitoring for performance drift. It separately seeks input on oversight of third-party foundation models and agentic systems, including how manufacturers should manage changes to underlying models they may not control. The paper is not guidance and does not establish new requirements, but it is FDA’s most detailed signal yet of how the agency may adapt its total product life cycle approach to generative AI-enabled devices. Comments were due October 19.

FDA Deputy Commissioner for Technology and AI: On September 8, the Department of Health and Human Services (HHS) as FDA’s first Deputy Commissioner for Technology and Artificial Intelligence. He will set agency-wide strategy for technology, software, and AI. Together with Carlton’s role at CMS, both agencies now have dedicated senior AI leadership.

ARPA-H

ADVOCATE Awards: On September 9, ARPA-H awards under its ADVOCATE program, worth up to $62.7 million over four years:

• Patient-facing AI agents: Atman Health, Tempus AI and UpDoc.
• Supervisory agent: A Stanford University-led team with the University of Washington, Tuple Health, and OpenAI, and a University of California San Diego Health, Harvard University, and Anthropic team.
• Health system implementation: Kaiser Permanente and a Duke University-led consortium including the American Heart Association.

ARPA-H’s ADVOCATE program aims to develop the first FDA-authorized agentic AI system capable of providing continuous cardiovascular care, including autonomously adjusting appointments, medications, diet and exercise while escalating to clinicians when needed. The program will pair patient-facing AI agents with a supervisory AI system that monitors their safety and effectiveness in real-world use, with deployment across diverse health care settings.

The success of ADVOCATE may have broad implications for how the FDA regulates autonomous clinical AI.

Congress

FRONTIER Act: On July 23, Reps. Obernolte and Trahan introduced the bipartisan FRONTIER Act (). It builds on the covered in our Q2 Tracker and would set tiered obligations for developers of the most advanced AI models, including transparency reports, critical safety incident reporting and catastrophic-risk frameworks. For the largest developers, the bill would require audits by federally licensed IVOs. The bill would establish a new Undersecretary for AI Security at Commerce to enforce these obligations and would preempt state laws on developer risk transparency, auditing and incident reporting. However, it explicitly preserves state regulation of how AI is used; as a result, most state health AI laws regulating how providers and payers use AI are likely to be unaffected. In September, OpenAI the bill’s third-party audit requirement, its first backing of a federal audit mandate, though it has not endorsed the full bill.

In the Senate, Sens. John Thune (R-SD), Ted Cruz (R-TX), Amy Klobuchar (D-MN) and Maria Cantwell (D-WA) are negotiating a separate frontier AI bill built around a “duty of care.”

Taken together, federal AI legislation this quarter has focused on developers of the most advanced AI models. There is bipartisan interest, and some industry support, in requiring transparency, incident reporting and independent audits. Preemption remains the key sticking point. However, the most developed proposal, the FRONTIER Act, would preempt only state requirements on AI developers related to subjects covered by the FRONTIER Act and preserve state authority to enforce generally applicable laws that do not target AI developers; regulate deployers’ and users’ use of AI, including under consumer protection, civil rights, contract, criminal, and privacy laws, so long as those laws do not impose substantive obligations on developers’ model development, training, evaluation, or release; protect minors from AI-related harms; and govern state procurement and use of AI systems. Accordingly, states would remain the primary regulators of AI use in health care for now.

Litigation on Ambient AI Tools in Clinical Practice:

A Washington state trial court recently issued one of the first rulings in the country addressing patient rights related to ambient AI tools, including how those rights apply to information captured by ambient AI tools during clinical encounters. In , a patient sought a preliminary injunction to obtain the ambient recording of her telehealth visit with her provider, arguing that the recording was part of the health information she was entitled to access under Washington’s Uniform Healthcare Information Act. The court, however, disagreed, and denied the injunction, holding that the recording fell within the law’s exception for information compiled for administrative purposes because it was used to assist the provider in preparing the written record of the encounter. Though only a state trial-court ruling, Raphael provides an early indication of how courts may approach patients’ access to information captured by increasingly common ambient AI tools: approximately one-third of providers are now estimated to have access to the technology.

Conclusions and What to Watch:

The key question for 2027 will not be whether health care AI adoption continues, but whether policy, payment, and accountability frameworks can keep pace as AI moves from assisting clinicians and administrative teams to performing more sophisticated  functions and, in some cases, acting with greater autonomy. We are watching six developments:

• From pilots to durable payment pathways. CMS has begun creating pathways for technology-enabled care through ACCESS,  the proposed SaMS payment category, and the RFI related to reimaging and modernizing primary care payments. The next question is whether these initiatives result in improvements in clinical outcomes and save money, and support a movement to more tech-enabled, value based care, rather than simply adding another reimbursable layer to the system.
• FDA’s approach to increasingly autonomous clinical AI. FDA’s generative AI discussion paper, the TEMPO program, and ARPA-H’s ADVOCATE program indicate that federal policy is beginning to contemplate how to test innovative models in the current federal regulatory environment, and in particular, evaluate AI that does more than provide recommendations. We will watch how FDA calibrates its oversight system to gather and evaluate evidence, perform pre- and post-deployment testing and monitoring, human oversight and incident or safety event reporting requirements. We also would not be surprised if FDA launched a new TEMPO-like sandbox available to technologies not serving ACCESS Model members.
• Liability and the emerging standard of care. Existing liability doctrines may not clearly allocate responsibility among developers, health care organizations, and clinicians when AI contributes to patient harm, particularly when no clinician reviews each decision. As evidence and adoption increase, the standard of care may shift such that  question becomes whether failing to use  AI creates liability.
• States as the primary regulators of AI use. States will continue to regulate how AI is used in health care, with particular attention to chatbots, minors, clinical decision support, payer practices, privacy, bias and professional licensure. California’s 2026 laws are likely to influence bills introduced in other states in 2027.
• Sandboxes and independent verification as alternative oversight models. Utah’s expanded sandbox with independent evaluators and California’s IVO framework reflect growing interest in supervised deployment, generating real-world evidence, and independent evaluation rather than relying exclusively on prescriptive regulation and relying on the AI tools own data and verification. We will watch whether these models produce scalable pathways to broader deployment and whether other states or federal agencies adopt similar approaches.
• Governance moving from best practice to expectation. Most health care organizations have basic AI governance in place. As the list of AI use cases grows and organizations become more facile with AI use, they will advance towards risk-based governance models that reduce friction for low risk use cases and increase the rigor applied to higher-risk use cases, such as more autonomous AI. Organizations will also invest much more in post-deployment monitoring and incorporate oversight of in-production models into their governance processes. 

For questions on the above, please reach out to or .

SB 243 requires operators to disclose that a chatbot is artificially generated, maintain protocols to prevent suicide and self-harm content, and provide added protections to users known to be minors.

If AB 1405 had not been enacted this session, SB 1119 included auditor-independence standards and sworn auditor certification standards that would have been enacted.

Defined as “an artificial intelligence system that produces a prediction, classification, recommendation, evaluation, or analysis that aids clinical decisionmaking related to timing of care, diagnosis, or treatment. It does not include systems that provide appointment management such as booking, canceling and rescheduling appointments, appointment reminders, patient education and pre-visit materials and preparation, and payment processing, to the extent the independent performance of these activities by the system does not require a professional license.”

Clinical decision supports are defined in the same manner as in AB1979.

Michelle Andrews, Doctors Increasingly See AI Scribes in a Positive Light. But Hiccups Persist., KFF Health News (Jan. 27, 2026), .