Click Therapeutics (Austin C. Speier)

See attached file

Via Electronic Filing System (www.regulations.gov)

October 19, 2026

Division of Dockets Management (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

Docket No. FDA-2026-N-7874 for “Considerations for the Regulation of Generative AIEnabled Medical Devices: Discussion Paper and Request for Feedback”

To the Digital Health Center of Excellence,

Click Therapeutics, Inc. (“Click”) appreciates the opportunity to comment on CDRH’s discussion
paper on generative AI (GenAI)-enabled medical devices.1 These comments build on our prior
submissions on GenAI, AI policy, and digital health.2 We support the Discussion Paper’s riskbased, total product life cycle approach and offer recommendations toward three objectives:

(1) Updating software-as-a-medical-device (SaMD) authorization pathways, and the
criteria for when new premarket submissions are required, so that SaMD such as digital
therapeutics (DTx) can integrate GenAI functions that improve on-market, based on an
initial validated performance baseline, GenAI-specific change control, and real-world
performance monitoring, without new premarket clinical studies.
(2) Creating a pathway for “agentic therapeutics”: agentic devices that treat one or more
conditions over an extended period of time by selecting and delivering validated
interventions, based on GenAI-supported autonomous and semi-autonomous clinical
reasoning, with risk-appropriate controls (e.g., under deterministic constraints and
clinician escalation, where warranted).
(3) Delivering authorized SaMD within general-purpose GenAI interfaces and personal
agents, where patients increasingly seek health information.

I. About Click Therapeutics

Click is an ISO 13485- and MDSAP-certified SaMD manufacturer of patient-facing software in
which the software is the treatment. Four FDA-authorized DTx have been developed on Click’s
platform or acquired by Click: Rejoyn™ (K231209), developed with Otsuka, for adjunctive
treatment of major depressive disorder symptoms; CT-132 (DEN240064), for preventive
FDA CDRH, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback (August 2026).
Click Therapeutics comments to Docket No. FDA-2024-N-3924 (January 21, 2025); Docket No. 2025-02305 (AI Action
Plan, March 15, 2025); CMS-0042-NC (June 16, 2025); and Docket No. FDA-2025-N-2338 (October 17, 2025).

T. 1-332-222-4177 80 White St 3rd floor,
info@clicktherapeutics.com New York, NY 10013 1
treatment of episodic migraine; Motivista™ (K254038), developed in collaboration with
Boehringer Ingelheim for the treatment of negative symptoms of schizophrenia;3 and
AspyreRx™, for type 2 diabetes. Click is also developing Software-Enhanced Drug™ (SE™)
treatments that combine software with pharmacotherapy. Our products are therapeutic, used
over weeks or months, and often used by people with serious mental illness, for whom GenAI
failure modes such as sycophancy and reinforcement of delusions are clinically significant.

II. GenAI Features in Digital Therapeutics

A DTx delivers a validated intervention, such as digitized cognitive behavioral therapy, on a
protocolized schedule. It is authorized by FDA for a treatment intended use, e.g., for the
preventive treatment of episodic migraine. We expect GenAI to offer many benefits to DTx: for
example, to personalize the presentation of authorized content, support conversational skill
practice, provide adaptive coaching within a program, or support new content generation to
enable long-term real-world patient engagement, amongst other functions.

A. Enabling GenAI content within a “Bounded Generation Specification (BGS)”
Many of the benefits listed above depend on custom content generation within the DTx. As
noted by FDA, not all possible outputs of these systems can be evaluated or validated, unlike in
a deterministic system. In addition to the competency-based approach FDA is considering, we
propose that sponsors set out the limits of generation in a Bounded Generation Specification
(BGS), submitted with the device’s 510(k) or De Novo request, or added to an authorized nonGenAI device through a 510(k) supported by non-clinical data. A BGS could contain, for example:

• Therapeutic content specification: the active elements generated content must contain,
permitted and prohibited content, a directiveness ceiling, and readability requirements.
• Safety behaviors: states that trigger deterministic escalation regardless of generative
output.
• Conformance test suite: benchmarks drawn from S.1–S.3, E.4, R.1, R.2, Therapeutic
Fidelity (generated content preserves the intervention’s active components) and
Relational Safety (sycophancy, dependency, and reinforcement of delusions), with
acceptance criteria that every configuration must meet (e.g., benchmarked guardrails on
over-refusal and over-escalation).
• Configuration criteria: qualification criteria for foundation models, prompts, retrieval
sources, and guardrails, rather than a fixed list of components.
• Monitoring and fallback: real-world performance monitoring and a validated safe-fallback
mode that reverts to authorized non-generative content.
A BGS defines what the device is authorized to generate, not which future changes it may
undergo. A new configuration that meets the configuration criteria and passes the conformance
suite produces output within the authorized space, so it is not a change that could significantly
Pratap A, et al. A Digital Therapeutic Intervention for Negative Symptoms of Schizophrenia: A Randomized Clinical Trial.
JAMA Netw Open. 2026;9(9):e2635753.
affect safety or effectiveness under 21 CFR 807.81(a)(3); it is documented in the sponsor’s quality
system. This parallels the design space concept in ICH Q8(R2), under which working within an
approved design space is not considered a change. Expanding the BGS, for example to new
content domains, a higher directiveness ceiling, or relaxed acceptance criteria, may trigger a
new 510(k), potentially a Special 510(k) based on non-clinical data where it stays within a
validated therapeutic baseline (see Part II.B below). Table 1 compares a BGS with a predetermined change control plan (PCCP).

PCCP Bounded Generation Specification

What is authorized Specific future modifications to the The space of outputs the device may
device generate

Legal basis Section 515C of the FD&C Act Existing 510(k) and change framework (21
CFR 807.81(a)(3))

What is prespecified Each modification, with a modification Content limits, configuration criteria,
protocol and impact assessment conformance tests, and acceptance
criteria

Changes within Implemented as specified, with labeling Documented in the quality system after
scope updates passing the conformance suite

Foundation model Must be anticipated as planned Permitted if the new model meets
updates modifications qualification criteria and passes the suite

How it is expanded New marketing submission 510(k), potentially a Special 510(k) on
non-clinical data

Best suited to Discrete planned changes, such as a Continuous generative variation and
new action class or expanded autonomy routine configuration changes

Table 1. A Bounded Generation Specification compared with a PCCP.

The two mechanisms are complementary: a sponsor could use a BGS for generative variation
and a PCCP for discrete changes; or even request a PCCP for authorization of future BGS
changes that might otherwise require a Special 510(k). Conformance results, monitoring data,
and any departure from the BGS would be retained for inspection, with departures handled
through complaint handling and reporting. To establish a BGS for an authorized non-generative
DTx, a Special 510(k) could be appropriate: the comparator could be the authorized nongenerative version, with a representative set of interactions assessed by experts and/or via
human factors testing, plus conformance testing suite results per the terms of the BGS, without
requiring burdensome new clinical studies.

B. GenAI device modifications from a validated therapeutic baseline
Inclusion of GenAI in a software device should not open a route to market for clinically
unvalidated therapeutics. Further, authorization for a therapeutic intended use based solely on
FDA’s proposed competency framework is not sufficient for prescribers, health systems or
payers to evaluate the performance of the device on real health outcomes (validated endpoints).
However, it is not reasonable to require new pre-market clinical evidence for all changes,
refinement or improvements to GenAI-enabled devices. Bounded generation is thus one
application of a broader principle that we suggest FDA consider adopting. Once the safety and
effectiveness of a software-delivered therapeutic intended use (for example, digital CBT for
insomnia) has been established for a population in the clinical evidence supporting an
authorization, later GenAI-enabled expansions of content and functionality within that category,
indication, and population should be supportable by three elements rather than new premarket
clinical studies: (1) reference to the baseline validation (e.g., pivotal clinical study); (2) a
documented clinical rationale linking each change to the validated mechanism; and (3) realworld performance monitoring.

Change (illustrative) Evidence

New or revised content within a Baseline validation reference; Bounded Generation Specification
baselined intended use, including conformance; real-world monitoring
GenAI personalization

New delivery modality or interface Baseline validation reference; human factors validation; Bounded
(for example, interactive voice, or Generation Specification conformance; real-world monitoring
delivery within a LLM-enabled
interface)

Combination of separately baselined Baseline validation reference(s); clinical rationale supporting the
intended uses; bridging to adjacent expansion; enhanced monitoring, such as staged rollout in a realindications; minor indication world setting
changes

New intended use or substantial New clinical evidence (new baseline)
indication change

Table 2. Evidence for changes relative to a validated therapeutic baseline.

The model depends on guardrails that preserve the evidentiary standard:

• A performance baseline established per-device, per-intended use. Every intended use
requires its own baseline of clinical evidence of the software-delivered intervention that
can inform label development and benchmark safety and effectiveness in a randomized,
controlled study (or equivalent appropriate validation study).
• A clinical rationale. The rationale identifies the intervention’s active components and how
it is administered, references clinical etiologies or dynamics and how the active
components apply in the new setting, and shows through non-clinical benchmarking that
changes preserve them (e.g., via conformance to a BGS; human factors testing; or similar).
• Real-world monitoring. Validated outcome measures collected in the device can be
compared against performance goals derived from the baseline. For more significant
changes, a change can be released in stages with comparison to the prior authorized
version, under applicable human-subject protections, with rollback if goals are not met.
Where possible, these elements can be implemented today through PCCPs whose modification
protocols use the baseline, the rationale, and real-world monitoring as the method for
assessing performance. We recommend that FDA confirm in guidance that real-world
performance data can serve this role, as a concrete answer to Question 18.

III. Agentic Therapeutics

The Discussion Paper’s agentic examples are mostly care coordination, documentation, and
device control. We ask FDA to consider a further class with a therapeutic intended use. An
agentic therapeutic monitors a patient, reasons within a validated clinical framework, and
selects, sequences, and delivers validated interventions to achieve a therapeutic goal. DTx
deliver validated treatment on a fixed schedule; general-purpose assistants respond flexibly but
cannot treat. Agentic therapeutics combine the two, offering a more dynamic and personalized
n-of-1 treatment approach.

A. Example architecture: the generative layer proposes, a deterministic gate decides
The regulatory properties that matter are the controls around the model: for example, an
enumerated action space; a non-generative validation gate that checks each proposed action
against the allow-list, labeled ranges, and protocol rules; safety triggers the generative layer
cannot override; an independent supervisory monitor; complete traceability; human
checkpoints before irreversible actions; and a safe-fallback mode (Figure 1).

Figure 1. Example of a bounded agentic therapeutic architecture.

B. Characterize an agent by its action space
To make emergent behavior tractable (Questions 5 and 26), FDA could characterize an agentic
device by its action space: each permitted action class placed on the two-axis grid with its
controls. Any action outside the envelope is a boundary failure testable under S.2. This also
supports distinguishing therapeutic agency (selecting and dosing validated interventions within
labeled parameters) from clinical agency (autonomous diagnosis, prescribing, or orders), which
sits much higher on the consequences axis.
 Action class (illustrative) Activity axis Consequences Representative controls

Personalize an authorized Protocol-bounded Limited Bounded Generation Specification
lesson within the content directive conformance; safe fallback
specification

Deliver or sequence an Action-taking: Limited Allow-list and labeled bounds
authorized module within autonomous enforced deterministically
labeled parameters

Escalate a safety-critical Action-taking: Severe Deterministic triggers; both error
state autonomous directions measured

Initiate or change a Out of scope Moderate to Not permitted autonomously;
medication severe clinician order required

Table 3. Illustrative action space for an agentic therapeutic.

C. Composability, graduated autonomy, and classification
• Composability. An agent’s action space should be limited to intervention categories with
a validated baseline (Part II.B above). When an agent invokes a separately authorized
device function within its authorized indication, the sponsor should be able to rely on that
authorization, with evidence focused on the orchestration: selection, timing, escalation,
and staying within the action space.
• Graduated autonomy. Mirroring supervised clinical training, a PCCP could let specific
action classes begin clinician-supervised and become autonomous once prespecified
real-world criteria are met, such as concordance with clinician decisions and absence of
unsafe actions. This responds to the Digital Health Advisory Committee’s call for a
common taxonomy of device autonomy.4
• Classification. Agentic therapeutics with these controls should generally be eligible for
Class II through De Novo under section 513(f)(2) of the FD&C Act, with special controls
under section 513(a)(1)(B): for example, an enumerated action space in labeling, the
deterministic validation gate and safety triggers, decision traceability, supervisory
monitoring, human checkpoints, safe fallback, an initial performance baseline (e.g., pivotal
study), periodic re-benchmarking, and clinical data supporting the therapeutic claim.

IV. Authorized SaMD Within GenAI Interfaces and Personal Agents

Health information increasingly reaches patients through GenAI. About half of U.S. adults use AI
chatbots, and 34% have used them for health, including a quarter who have used them to try to
determine the cause of symptoms.5 80% of consumers rely on AI-written results for at least
FDA, 24 Hour Summary of the Digital Health Advisory Committee Meeting on Generative AI-Enabled Digital Mental
Health Medical Devices (November 6, 2025).
Pew Research Center, Americans and AI 2026 (June 17, 2026); Pew Research Center, From Diagnoses to Treatments,
Why Americans Use AI Chatbots for Health (August 25, 2026).
40% of their searches.6 One platform reports more than 300 million people asking it health
questions weekly while stating that its service is not intended for diagnosis or treatment.7 These
platforms now render third-party interactive apps within the conversation,8 and consumer
personal agents, including Meta’s Muse, Google’s Gemini Spark, and OpenAI’s dots, work
continuously on users’ behalf across thousands of connected apps.9 A treatable condition may
now first surface in a chat, or be noticed by an agent reviewing sleep or activity data. If
authorized SaMD cannot be delivered there, patients will rely on unvalidated advice instead.

The Discussion Paper addresses GenAI that sponsors build into their own devices. It does not
address the inverse: an authorized device delivered inside a GenAI host, or used alongside a
personal agent, that the sponsor does not control. FDA does not treat mobile platforms or app
stores as device manufacturers,10 but a GenAI host differs: it can paraphrase device content,
pass context to it, and suggest it to users, and a personal agent may act within it for the patient.

Example. A user tells an assistant they have slept poorly for months, or their personal agent
notices this in wearable data, and is offered an FDA-authorized digital CBT for insomnia. The
device runs in a sponsor-controlled interactive view: it performs its own screening and, if
prescription-only, prescription verification; collects sleep diary entries directly; and calculates a
sleep window that must reach the patient exactly as calculated. The agent may schedule
sessions and send reminders, but it cannot complete the diary or the exercises, because agententered data would corrupt the calculation and agent-completed exercises deliver no therapy.

Figure 2. Boundary between a GenAI host or personal agent and an authorized device.
Bain & Company, Goodbye Clicks, Hello AI: Zero-Click Search Redefines Marketing (February 19, 2025).
OpenAI, Launching Health in ChatGPT (July 23, 2026).
OpenAI, Introducing Apps in ChatGPT and the New Apps SDK (October 2025); Model Context Protocol, MCP Apps
(January 26, 2026).
Meta, Muse (U.S. launch September 8, 2026); Google, Gemini Spark (May 2026); OpenAI, Introducing dots (September
29, 2026).
FDA guidance, Policy for Device Software Functions and Mobile Medical Applications; FDA guidance, Multiple Function
Device Products: Policy and Considerations.
We request that FDA consider addressing the following principles associated with this context
of use in future guidance (note: this is an initial list of considerations, and not meant to be
comprehensive):

1. Host status. A host platform or personal agent that carries an authorized device function
it does not control is not thereby its manufacturer. The host model is an “other function”
under the multiple function policy, whose effects the sponsor must assess.
2. Entry. Surfacing an authorized device in response to a stated need or connected data,
without diagnosing, is referral-like, not a device function. The device performs its own
eligibility, contraindication, and prescription checks.
3. Output and input integrity. Clinically significant outputs are rendered by the device
exactly as generated, never paraphrased by the host. Clinical inputs are collected by the
device; host-supplied context is unvalidated unless validated.
4. Patient-reserved actions. Screening answers, consent, diary and outcome entries,
therapeutic exercises, and safety acknowledgments must be performed by the patient,
with human presence verified. Logistics may be delegated to an agent.
5. Agent inputs and channels. Messages from a host model or agent are untrusted inputs
subject to prompt-injection testing (A.1). Where a channel such as voice or messaging
cannot preserve output integrity, the agent hands off to the device’s own interface.

6. Operational controls. Managing device boundaries and in-context labeling; device
escalation that operates independently of the host; validation of declared host
environments, with host updates handled like operating system updates; in-context
complaint reporting; and isolation of device data from host memory except as the user
authorizes.
7. Specialist agents. An authorized agentic therapeutic may operate as a specialist agent
within a third-party agent platform, with the device boundary at that agent and its
validation gate.

V. Cross-Cutting Considerations

GenAI-enabled Devices and PDURS. Under FDA’s draft Prescription Drug Use-Related
Software (PDURS) guidance, software output is drug labeling, either FDA-required or
promotional.11 Because generative output varies, we ask CDRH, CDER, and the Office of
Combination Products to jointly clarify how any CDRH-developed frameworks or Guidance
apply to GenAI-enabled PDURS; for example, whether drug promotional labeling submissions
may consist of the content specification, guardrails, and representative outputs rather than
every variant; whether competency-based evidence can support software-derived claims in
FDA-required drug labeling; and whether GenAI change control methods (e.g., the proposed
BGS or baselining framework in Part II above) can satisfy both Centers.
FDA draft guidance, Regulatory Considerations for Prescription Drug Use-Related Software (September 2023).
The boundary with non-device products. We renew our October 2025 request to add GenAI
examples to the Policy for Device Software Functions appendices. Functions providing general
education on coping skills for diagnosed psychiatric conditions, without specific treatment or
treatment suggestions, would fall under enforcement discretion. Functions engaging in
personalized dialogue to diagnose, assess, or deliver treatment, or presenting as a therapist,
would be the focus of oversight. Labeling of GenAI-enabled products, including those under
enforcement discretion, should describe where GenAI is used, how it was evaluated, and its
guardrails. The regulated pathway should remain least burdensome, so that validated products
are not disadvantaged relative to unregulated ones.

VI. Conclusion

With these refinements, FDA can develop a forward-looking, least burdensome, and
appropriately rigorous regulatory framework for GenAI-enabled devices, let authorized GenAIenabled therapies improve and expand from a validated clinical baseline, establish a pathway
for agentic therapeutics, and ensure authorized treatments are available within public GenAI
and agentic interfaces where patients increasingly begin and manage their healthcare journey.
Click would welcome further discussion or opportunities to provide input on the above
recommendations.

Sincerely,

Click Therapeutics
Austin C. Speier
Chief Strategy Officer