Legion Health (Arthur MacWaters)

Response from Legion Health

Legion Health FDA RFI Response
Considerations for the Regulation of Generative Al-Enabled Medical Devices: Discussion Paper
and Request for Feedback

Monday, October 19, 2026​
​
For any questions, please reach out to:​
​
Arthur MacWaters​
Co-founder, Legion Health ​
founders@legion.health

IV. Risk assessment
1. Does the two-axis risk framework, organized around device activity and the consequence of
relying on an incorrect output, appropriately capture the dimensions most relevant to the risk of
a GenAI-enabled software function? If there are additional dimensions—such as the reversibility
of a resulting action, the availability of downstream safeguards, the time pressure of the
deployment setting, or the traceability of the output (i.e., to primary source materials)—that
should be represented in a risk framework, please describe and provide examples of how they
should be represented.

Legion Health is an AI-native, fully-integrated psychiatry clinic. We deliver clinician-led care with
AI-powered automations (intake summarization, population risk screening, operational
automation, revenue cycle automation) and maintain documented safety protocols and
auditability for clinical operations. Our providers have completed nearly 50,000 telemedicine
visits in Texas, supported by our AI platform, with exceptional clinical outcomes and patient
satisfaction. On this foundation, and with the recognition that properly designed generative AI
can meet the need for mental health care at scale without sacrificing patient safety, we built a
tool that received a first-of-its-kind regulatory authorization in Utah for AI-powered renewals of
eligible psychiatric maintenance medications. The authorization permits our AI agent to transmit
eligible renewal prescriptions directly to pharmacies, subject to the agreement’s phased review
requirements and escalation protocols.

Legion Health supports efforts to expand clinical uses of healthcare AI through a risk-based
framework that creates a common path for gen-AI products to advance toward greater
autonomy upon evidence of safe outcomes and compliant design. We are generally very happy
with this risk framework; in several important respects it aligns with our approach and
commitment to the responsible deployment of our clinical AI system in the Utah AI Sandbox.
Based on our real-world experience, we broadly endorse this approach because it creates
incentives for thoughtful, risk-based deployment of clinical AI rather than one-size-fits-all
regulation.

We want to ensure that FDA regulation is calibrated to the level of risk while also recognizing
that design of a system with appropriate guardrails and oversight can maintain patient safety,
clinical integrity, and public trust while concurrently creating an environment that is supportive of
innovations that can expand access to high-quality care. The proposed matrix rightly recognizes
the critical role of human oversight in the healthcare ecosystem and its effect on regulatory
discretion.

A graduated framework recognizes that Gen-AI tools in a clinical setting function not only as
products but as practitioners. Oversight reflecting this reality can come both from direct human
supervision of the tool in operation and initial regulatory review ensuring its performance meets
the applicable standard of care. By focusing on the specific application, context, functionality,
and degree of human oversight associated with a tool, approaching regulation of Gen-AI
through this framework provides a regulatory foundation that can evolve alongside technological
advancements and emerging evidence.

At the same time, we believe there is an opportunity to further strengthen the framework through
increased clarity about how gen-AI tools graduate towards greater levels of autonomy over time.
We offer recommendations that preserve the matrix's risk-based structure while enhancing its
applicability to real-world care delivery environments.

However, there is not a continuous slope of risk amongst all medical interventions. It’s easy to
see that surgical interventions are higher risk than medication management, but hard to tell
exactly the risk amount from one medication class to another, or a refill of one medication vs. a
new diagnosis and initiation of another medication. The model is heuristically very sound but
would become difficult to implement across different categories of care without explicit guidance.
As an example of how this could be applied, any gen-AI approval of medications or treatments
would all progress in the following manner, recognizing that refilling an existing,
clinician-prescribed treatment is generally lower risk than initiating treatment or changing
treatment. The sensible first step would be to use supervised refills as a starting point with
graduation to autonomous refills, and with supervised initiation of prescriptions permitted only
after evidence is gathered to understand performance of the underlying model. This approach
has allowed Legion Health to manage our risk profile and make innovative progress
simultaneously!

Our Utah agreement provides a concrete example: 250 renewals reviewed before issuance,
with greater than 98% concordance to advance; then 1,000 retrospectively reviewed renewals,
with greater than 99% concordance to advance to periodic sampling and incident-driven audits.
The broader principle is to agree on case counts, performance thresholds, and safeguards in
advance so companies know what evidence unlocks the next stage.

The laddered graduation path illustrated above also provides protection against patient harms
by platforms that have not been designed to align with the standard of care expected of human
practitioners. We propose that all companies should be held to the same high standard of care
and all companies should only be allowed to progress towards greater autonomy or treat more
complex conditions based upon a demonstration of efficacy and safety measured upon shared,
known evidentiary requirements agreed upon by the developer and FDA. Additional
considerations may also include whether the developer has been found to deviate from
standards of care, have improper privacy protections, etc.

We would like clarity on how a company like ours can move up the risk and automation axes,
e.g., graduate to supervised and then autonomous prescription initiation or expanding our
medication formulary. We believe that AI will only be truly beneficial and deflationary if it can
actually provide important parts of the care journey. Therefore, we propose a standardized
protocol to enable deployment of tools addressing higher-impact use cases with greater
autonomy in a clinically robust way in the real world. ​
​
The remaining question is who defines the risk category and number of patients in each stage
gate. To create clarity for industry innovators, we would like to request published, usable
categories, clinical examples, and default advancement criteria based on the intervention,
medication class, patient population, potential harm, reversibility, and safeguards. Companies
should be able to apply those rules directly, with prompt regulator confirmation for ambiguous or
novel cases. This would preserve the matrix’s flexibility while making both the starting scope
and the path to greater autonomy predictable. FDA should use the flexibility available within its
existing authorities to support this approach.​
​
We encourage the CDRH to continue to develop its approach around a theme of right-touch
regulation based on least burdensome provisions and avoid unnecessarily punitive or overly
prescriptive requirements that could impede the development and deployment of beneficial
technologies, with the recognition that there may be factors in how a tool is developed and
deployed that provide levels of oversight that ensure clinical safety. We believe that FDA has
within its existing regulatory authorities the ability to apply this risk-based framework. We would
encourage the FDA to leverage its existing authorities to incorporate this risk-based framework
into its approval processes. We appreciate the thoughtfulness of the proposal.

2. Directiveness. CDRH seeks input on how to account for the spectrum of GenAI-enabled
informational functions that vary in the degree to which they direct a user to a particular action
(i.e., between “non-directive” and “action-directing”). What characteristics of an output—such as
its wording, specificity, personalization, or context—could be considered as modifiers of the risk
associated with an informational function, after accounting for the device’s overall functionality
and intended use? What additional information would provide manufacturers with sufficient
clarity and predictability around risk assessment for informational functions while recognizing
that directiveness may exist along a continuum rather than as a binary distinction?

We agree that directiveness is a spectrum, but don’t feel that directiveness alone is indicative of
the quality or safety of an answer. Accurate, well-supported medical guidance with human
review as needed can be incredibly helpful to patients. The goal here should be a predictable
path for companies to take on more clinical responsibility as they demonstrate competence. Any
future guidance should avoid creating an incentive for companies to give vague, unhelpful
answers simply because actionable care is harder to classify.

3. Patient-facing information. CDRH seeks input on whether and when a GenAI-enabled
function that results in delivery of clinical information to patients, as opposed to HCPs, could
present different or higher risks, while also recognizing the potential benefits associated with
improved patient empowerment, engagement, and access to clinical information. What device
characteristics, output features, or safeguards might mitigate risks that could arise when a user
lacks the domain knowledge to independently evaluate an output, without unnecessarily
underestimating patient capability?

Some characteristics that would make patient facing AI communications higher quality and more
useful to patients:

1.​ Cited external sources that patients can use to learn more and verify answers (e.g.,
JAMA/NEJM links)
2.​ Cited internal resources (e.g., chart notes or other medical records)
3.​ Clear persistent indication for a patient to know when they are interacting with AI vs.
human
4.​ Clearly labeled, simple off-ramp for the patient to escalate to a human HCP (including a
telehealth visit as needed)

4. Generalists and specialists. CDRH seeks input on whether and how the distinction between
generalist and specialist physicians might inform the assessment of risk for HCP-facing
GenAI-enabled functions. Under what circumstances, if any, might risk be affected when an
HCP who lacks the relevant clinical specialist knowledge receives information that falls within a
specialist area of practice? What device characteristics or safeguards might mitigate such risks?

Clinicians informing the delivery of care should have the requisite knowledge of the standards of
care for the particular condition treated and should inform how the tool interacts with the patient.
Tools designed around a specialized standard of care, with the appropriate level of engagement
by such specialists, meet conditions that may lead to a shift in the matrix of regulatory oversight
as presented. Over time, evidence that the combined system safely handles a task should allow
generalists to extend access to specialty knowledge within their scope of practice.

5. Multi-turn conversations. For multi-turn conversational GenAI-enabled devices that may
migrate from providing “non-directive” information to “action-directing” information over the
course of an exchange, how should risk be assessed across realistic conversational
trajectories? How could the intended use of such a device be characterized when its behavior is
emergent across a conversation?​
​
For action-taking systems (like renewing a medication) we recommend including deterministic
(non-AI) checks on either end of an interaction that parallel the appropriate standards of care
and ensure that there are multiple independent methods of ensuring a good outcome.

6. Under- and over-escalation. For GenAI-enabled care escalation functions, CDRH is
considering how the evaluation may account for both under-escalation and over-escalation. How
could manufacturers characterize and weigh these two directions of error, given that they may
not be commensurable and that acceptable trade-offs may vary by clinical context?

We strongly support measuring both directions of error, but they should not be treated as
equally harmful. For an early medication-renewal system, referring an otherwise eligible patient
to a clinician is generally preferable to renewing when clinical review was needed. Our Utah
agreement reflects this distinction: cases where the AI escalates too cautiously and the
reviewing clinician would have renewed do not count against the initial greater-than-98%
concordance requirement. That is a useful incentive to be conservative while gathering
evidence.

We would still report unnecessary escalation separately, including whether patients actually
receive care and how long they wait. An escalation can become harmful if it causes a
medication interruption or sends a patient unnecessarily to emergency care. Review of a routine
renewal and emergency triage therefore need different thresholds.

V. Competency and clinical confirmation
7. Competency-based evaluation. Is the competency-based approach described above, i.e.,
device benchmarking followed by clinical confirmation, a useful and appropriate framework for
evaluating GenAI-enabled devices?

We are very supportive of this approach. It resembles the practical process that Legion has
taken as a developer: demonstrate competence in realistic simulations, validate it with human
clinicians in actual care, and then progressively reduce the amount of review as the evidence
supports it. The responsible organization remains accountable for meeting the standard of care
throughout.

8. Evidence proportional to risk. How might the two-axis risk framework described in Section
IV be considered within the competency-based approach described above to help determine the
level of evidence needed for a premarket submission?

If the clinical risk is low (especially when buttressed by downstream safeguards) then the
evidentiary threshold for premarket submission should be commensurately lowered in order to
allow innovation. Developers should be held to the same standard of care as traditional
providers and progress towards greater autonomy in treating complex conditions based upon
shared, known evidentiary requirements agreed upon by the developer and FDA.

We encourage the CDRH to continue to develop its approach around a theme of right-touch
regulation and avoid unnecessarily punitive or overly prescriptive requirements that could
impede the development and deployment of beneficial technologies with the recognition that
there may be factors in how a tool is developed and deployed that provide levels of oversight
that ensure clinical safety.

9. Benchmark content. Would a benchmarking structure such as the one described above be
likely to provide adequate evidence of clinical knowledge, analytic capabilities, safety behavior,
communication, and generalizability to support a reasonable assurance of safety and
effectiveness? Are there elements that are missing, redundant, or inappropriately categorized?
Are there externally developed standards that could be leveraged?
We are supportive of shared applied medical benchmarks built from de-identified real cases,
however we are cautious of regulation that relies upon the broad application of standardized
FDA-generated benchmarks due to the fact that benchmarks quickly become outdated and
models can be overfit to them.

10. Benchmark validity. CDRH recognizes that publicly available benchmarking assets may be
subject to data contamination, saturation, and limited real-world representativeness. How should
a sponsor establish that performance on a given benchmark predicts safe and effective
real-world behavior for the device’s intended use? What evidence should support the construct
validity of a benchmark used to gate device evaluation, and what role should sponsor-developed
benchmarks play given potential concerns around independence and optimization to the test?

While we appreciate the FDA’s interest in benchmarking, we believe individual companies are
best-positioned to develop and conduct benchmarking either internally or with an independent
3rd party adjudicator. Sponsor-developed benchmarks are valuable because care providers
understand their own workflows and failure modes and can facilitate constructive discussion
with regulators on whether a specifically-tailored benchmark provides sufficient evidence to
support shared goals of patient safety. We would combine them with independent adjudication
and external challenge cases.

11. Clinical confirmation. CDRH is considering that clinical confirmation for a GenAI-enabled
device might not require a prospective clinical study in every case, and has described above a
range of approaches of increasing rigor and patient exposure. How might a sponsor select and
justify a confirmation approach tailored to a device’s intended use and proportionate to the
device’s risk profile? Are there device types or risk profiles for which one or more of these
approaches would be insufficient or inappropriate? How might the anticipated distribution of
real-world inputs be taken into consideration? Are there other methods of clinical confirmation
that might help inform the evaluation of GenAI-enabled devices?

We find it very important to emphasize that it would be detrimental to innovation broadly and the
quality of clinical outcomes if companies who are developing AI tools that can provide clinical
interventions were required to seek approval through the traditional approval pathway for
software-as-a-medical device. Treating AI-led care as a medical device does not match the way
in which these tools work, nor the premise of including both humans and AI together in the
practice of care which meets or exceeds the standard of care.

As AI models become more competent, they will behave far more like humans and far less like
deterministic pieces of software. Our strong position is that this means that they should be
incorporated into the medical world as part of the care team, with supervision by other
providers, and judged based on their ability to meet the standard of care for patients.

Evaluating competency based upon the ability of a tool to meet a standard of care, as confirmed
by appropriate human oversight of the clinical encounter, is a more appropriate approach to
ensuring safety and efficacy. Such an approach is also supported by prior commentary
regarding clinical decision support tools which have led to software being regulated not as a
medical device per se, but as a tool that falls within a much lower risk tier.

As the FDA refines its approach, the competency-based framework being used in Utah provides
a promising reference point. Underpinning the approach in Utah is the recognition that AI tools
will ultimately function more like members of a clinical care team than a traditional piece of
software. When viewed through that lens, regulatory oversight focuses on whether the tool
operates within the applicable standard of care and whether clear accountability exists for
patient outcomes. This framework also acknowledges that, as in traditional healthcare delivery,
complementary layers of human review and supervision can serve to ensure competent delivery
of care.

Viewing AI systems as participants within a supervised care team and held to a standard of care
encourages innovative regulation that checks for competence and protects patients from harm
while avoiding unnecessarily burdensome requirements that could impede innovation and delay
the deployment of beneficial technologies.

12. Statistical measurement. How should sponsors achieve statistically meaningful
performance measurement for GenAI-enabled devices? Where synthetically generated inputs
supplement real patient data, how should sponsors account for differences between the
synthetic and real-world distributions when estimating performance, and under what conditions,
if any, is it appropriate to combine benchmarking evidence and clinical confirmation evidence to
support a single performance estimate?

In developing concordance measurements with the Utah Office of AI Policy, Legion reached a
shared understanding with the Office about which data was meaningful for evaluating
performance. We recognize that generative AI tools will take many forms, and that different
measurement methods may therefore yield data sufficient for a regulator to conclude that the
tool is appropriately designed to avoid harmful outcomes and functions to provide care in a safe
manner. For that reason, we support a framework that allows developers to present evidence
addressing the core regulatory concerns of safety and effectiveness to progress through the
regulatory approval process.

In general we also think it’s important to define the clinical decision being measured and its
denominator. For example, passing through a stage gate of deployment in the Utah AI Sandbox
requires 98% concordance between the AI system and the human providers reviewing the
outputs. Without clearly defined, common evidentiary standards (like those which were agreed
to by the Utah OAIP working with Legion Health) a claim of “98% concordance” is difficult to
interpret without knowing which disagreements count, how cases were selected, and how many
independent patients were evaluated. ​
​
The Utah AI Sandbox has very clearly defined concordance measures, and as a consequence,
we are supportive of extending this approach more broadly. In addition, successful participation
and outcomes within a state regulatory sandbox should count toward the statistical
requirements for broader deployment, allowing a faster path to authorization.
13. Synthetic data. For which clinical domains, device functions, or subpopulations is synthetic
data particularly well-suited, or particularly inadequate, as a supplement to real-world evidence?
What safeguards would mitigate the risk that synthetic data generated by models of the same
class as the device under evaluation reproduces the very performance gaps the evaluation is
intended to detect, particularly for underrepresented subgroups?

We see substantial value in simulation. In preparing our Utah refill workflow, we ran 100,000
simulated patient interactions. Simulation lets a team systematically test situations that would be
slow or inappropriate to assemble through live care: contradictory medication histories, unusual
combinations of symptoms, pressure to go beyond scope, and rare scenarios requiring urgent
escalation.

Simulation should help us find weaknesses and expand coverage. Real patient evidence
remains necessary to establish how the system performs in practice.

14. Performance standards. For open-ended device outputs, how should performance
comparators and acceptance criteria be selected? When a panel of qualified clinicians serves
as the comparator, how should the applicable standard (for example, the standard of care
versus the performance of a median clinician in practice) be defined and justified? Should
generalist or specialist physicians be used as a performance standard? When should human-AI
team performance, rather than the device operating alone, serve as the basis for evaluation?

We believe the applicable standard of care is the right paradigm. The ultimate goal is to
establish whether a system with both humans and AI makes clinically acceptable decisions for
its intended patients, and ultimately whether patients get better. There may be several
appropriate answers, so an evaluation should not penalize a reasonable alternative simply
because it differs from one clinician's preference.

15. The real comparator. Are there ways in which performance might be assessed relative to
the care, technology, or course of action likely to occur in the absence of the device, rather than
to the comparators described in this section? What approaches might be used to identify and
justify a comparator such as unaided clinical judgment, delayed specialist review, or no
intervention?

The shortage of doctors and other providers in this country is a massive problem that affects
every patient. In Utah, for example, 27 of 29 counties are designated as mental health shortage
areas. The real comparator to the AI-driven care we have been authorized to provide is not
simply human care, it’s human care that might never be available, or is prohibitively far,
expensive, and delayed. We believe that it is a moral imperative to allow AI to help expand
high-quality care to more patients, at substantially lower cost and with greater personalization.

16. Independent assessment. Should independent third parties be involved in some or all
aspects of a competency-based approach, including device benchmarking and clinical
confirmation? If so, in what ways might qualified, independent third-party participation contribute
to this approach, and what qualifications and independence criteria should apply? Are there
aspects of the assessment for which third-party involvement would be impractical or
inadvisable? What safeguards or program design features would be critical to prevent
third-party participation from limiting competition or preventing innovation?

We support independent participation where it makes the evidence more credible: maintaining
private test sets, adjudicating difficult clinical cases, and auditing whether a company's
evaluation process works as claimed. Evaluators should have relevant clinical and technical
competence, disclose conflicts, and be independent of the development work they assess. A
sponsor can pay for an evaluation without controlling its findings.

Designating specific organizations as gatekeepers for participation risks limiting innovation in
both care delivery and scientific inquiry. While third-party validation by traditional incumbents in
health care can play an important role in establishing trust, the primary stakeholders in the
regulatory approval process should be the FDA and developer of the AI tool.

Any framework should recognize that there are multiple credible methods for assessing the
safety, performance, and reliability of AI tools. For example, successful participation in a
regulatory sandbox or specific clinical setting can generate meaningful real-world evidence
demonstrating that a tool operates within the applicable standard of care, functions safely in
clinical settings, and has not resulted in adverse patient outcomes. Such evidence should be
considered along with assessments by designated third-party validators. Flexibility in evaluation
of outcomes and performance can encourage innovation while maintaining appropriate
safeguards for patients and the public.

17. Different architectures. CDRH is exploring whether a competency-based approach could
be applied to devices that incorporate a variety of underlying model architectures, such as
multimodal vision-language models and generative or predictive world models. Are there
aspects of the described approach that might be ineffective or inapplicable to these devices?
What additional or distinct considerations might need to be addressed?

We think a competency-based approach becomes more valuable as architectures diversify. A
care provider should be able to choose a frontier model, an open-source model with its own
supporting software, or a combination of models if the complete system demonstrates the
required clinical performance.

VI. Monitoring and change control
18. Greater postmarket reliance. CDRH is considering whether it may be appropriate to
accept greater premarket uncertainty regarding a GenAI-enabled device’s benefit-risk profile
through greater reliance on postmarket monitoring. Under what conditions might such an
approach be appropriate, and what characteristics of a monitoring program would need to be in
place to justify reduced premarket evidence? Are there device types or risk profiles for which
this approach would not be appropriate?
Particularly in the case of lower-risk applications of AI tools, an appropriate regulatory approach
will place a primary emphasis upon postmarket monitoring. We believe rigorous postmarket
monitoring is a credible way to learn safely in actual care environments and continue to gather
valuable real-world evidence. The safeguard of human providers in the loop (and, ideally, an
entire provider organization with operational experience operating at the standard of care) helps
ensure quality of care within the context of ongoing positive iteration.

19. Monitoring design. Please comment on the potential approaches to postmarket
performance evaluation, including periodic re-benchmarking, sample-based clinician review, and
performance degradation monitoring. What additional approaches should CDRH consider, and
how should the cadence and triggering events for reassessment be determined?

We strongly favor sample-based clinician review as the best and most empirically sound
methodology proposed. This supports ongoing assessment of whether the standard of care is
being met. We would advocate for review frequency to be set by the FDA and the developer
according to risk, volume, and how quickly harm could emerge.

20. Supervisory agents. Please comment on whether the proposed approaches to postmarket
monitoring can be facilitated by machine-based supervisory agents. What considerations,
including the evaluation and reliability of the supervisory agent itself, should CDRH take into
account for such an approach?

Supervisory agents are an additive approach and could become part of the standard of care.

21. Accountability. What roles might clinicians, healthcare institutions, professional societies,
standards-setting bodies, and other stakeholders appropriately play in postmarket monitoring,
and how can these roles be structured without diffusing manufacturer accountability?

For a vertically integrated provider such as Legion, clinical oversight, technology development,
and the ability to bring in a clinician can be coordinated within the same operating model. As a
generally applicable policy principle, the entity designing the tool should remain accountable for
meeting the standard of care even when that requires using more humans. This has been a
great benefit to our progress, and aligns incentives in our organization to maximize good
outcomes. This structure should be considered a positive factor when evaluating industry
partners.

22. Proportionate changes. CDRH envisions that a manufacturer’s premarket
competency-based assessment might serve as a baseline against which post-deployment
modifications could be re-evaluated. How might the extent of re-benchmarking or other
evidence be scaled to the nature and expected impact of a given modification? For example,
are there categories of change that might not significantly affect safety or effectiveness of a
GenAI-enabled device, or might be appropriately managed within a sponsor’s quality
management system versus requiring FDA premarket review and authorization before
implementation? Are there categories of changes that might be appropriate for inclusion in a
PCCP?
Our strongest preference is to evaluate changes according to their clinical effect. Given the
rapid pace of improvement within the underlying AI models, it is critical to avoid a detrimental
scenario in which patients receive care from an outdated system simply because approval was
tied to one model and one set of supporting software.

23. Unforeseeable improvements. How might PCCP concepts or other change-control
approaches be adapted for GenAI-enabled devices when the nature or scope of future
modifications cannot be fully prespecified?

We think the practical answer is to specify what must remain true about the care, while allowing
flexibility in how the technology achieves it. It is difficult to predict which model or architecture
will be best a year from now. It is much easier to define eligible patients, permitted interventions,
required escalation behavior, and acceptable clinical performance.

24. Third-party updates. For devices built on third-party foundation models, changes to the
underlying model may be initiated by the third-party model developer rather than the device
manufacturer. How can a manufacturer detect, evaluate, and respond to such changes in a
timely manner? What mechanisms—for example, contractual, technical, or through a
PCCP—could provide reasonable assurance that third-party developer-initiated changes do not
compromise the safety or effectiveness of the device?

The care provider needs to know when the technology it relies on changes. Contracts with
model suppliers should support version identification, advance notice, testing access, and a
reasonable transition period where possible. Pinning a version can help, but cannot be the only
protection if the supplier changes other service behavior or withdraws that version.

VII. Foundation models and agents
25. Foundation Model MAFs. Would voluntary Foundation Model MAFs be practical for
foundation model developers to provide and to keep current, and what content might they
include? Given that participation would be voluntary and that model developers may have
limited incentive to disclose safety-relevant information, what would make such a program
sufficiently useful for premarket review? Are there alternative mechanisms CDRH should
consider for obtaining information about underlying foundation models?

n/a

26. Agentic care. Are there additional considerations that inform the premarket and postmarket
evaluation of agentic GenAI-enabled devices, beyond those applicable to non-agentic
GenAI-enabled devices? How should the elevated risk associated with autonomous multi-step
action, tool use, and reduced opportunity for human review be reflected in acceptance criteria
and oversight?

n/a