Medicept Inc.

Submitted by Medicept, Inc.

Considerations for the Regulation of Generative
AI-Enabled Medical Devices
Discussion Paper and Request for Feedback

U.S. Food and Drug Administration, Center for Devices and Radiological Health. Considerations
for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for
Feedback. August 2026.

Docket FDA-2026-N-7874

Submitted by Medicept

Agency U.S. Food and Drug Administration, Center for Devices and
Radiological Health

Comment deadline October 19, 2026

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 1
Introduction
Medicept appreciates the opportunity to provide feedback on the Center for Devices and
Radiological Health (CDRH) discussion paper, “Considerations for the Regulation of Generative
AI-Enabled Medical Devices.” Medicept provides regulatory, quality, clinical, and productdevelopment consulting services to medical device and digital health organizations. Our
comments focus on practical approaches for translating established medical device principles
into a workable framework for generative artificial intelligence (GenAI)-enabled devices.
Medicept generally supports CDRH’s risk-proportionate, total product lifecycle approach and
the exploration of competency-based evaluation. We recommend, however, that any future
framework remain grounded in the safety and effectiveness of the finished medical device in its
intended workflow. In particular, the framework should account for effective—not merely
nominal—human oversight; system- and workflow-level risk; harms arising from reliance on
device behavior even when an output is factually correct; intended use traceability for
benchmarking; risk-proportionate clinical confirmation; impact-based change assessment; and
practical controls for third-party foundation models and agentic systems.

Executive Summary of Recommendations
• Evaluate risk at the device, system, and clinical workflow level, including how GenAI
outputs may lead to or be interrupted before patient harm, rather than assessing the
GenAI function in isolation.
• Distinguish nominal HCP involvement from effective human oversight and meaningful
opportunity to intervene before harm.
• Broaden the concept of harmful reliance beyond factually incorrect outputs to include
unsafe effects of otherwise correct device behavior.
• Require traceability from intended use and clinical risks to competencies, benchmarks,
acceptance criteria, and clinical confirmation.
• Scale clinical evidence to residual uncertainty and clinical consequence rather than
treating GenAI as a categorical trigger for prospective studies.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 2
 • Use periodic postmarket surveillance together with predefined, risk-based triggers,
action thresholds, and escalation criteria for reassessment.
• Apply impact-based change assessment and targeted re-benchmarking rather than full
revalidation after every modification.
• Address third-party foundation models as a lifecycle supplier and change control
challenge, using both contractual and technical controls.
• For agentic systems, evaluate the safety of end-to-end action sequences, including
propagation of intermediate errors and opportunities for intervention.

Detailed Comment Matrix
DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Comment 1 — Two-Axis Risk Framework Position: Support with clarification.
Discussion Paper Reference: Section IV,
“Considerations for the Assessment of Risk for Medicept Comment: FDA proposes evaluating
GenAI-Enabled Devices,” pp. 5–9; Question 1. GenAI-enabled medical device risk using two primary
dimensions: (1) the type and degree of activity
Relevant FDA Language / Question: CDRH performed by the device, and (2) the potential
describes a two-axis framework that places severity of harm if a user relies on an incorrect output.
“device activity” on one axis and the Medicept agrees that these two dimensions provide a
“consequences,” or severity of harm from useful starting point for organizing risk.
relying on an incorrect device output, on the
other. Question 1 asks whether additional However, Medicept recommends that FDA clarify that
dimensions such as reversibility, downstream these two dimensions alone may not fully describe
safeguards, time pressure, or traceability the actual risk presented by a GenAI-enabled device.
should be represented. Two devices may perform a similar type of activity
and have the same potential severity of harm if an
output is incorrect but still present meaningfully
different real-world risks depending on what happens
between the device output and the potential harm.

For example, one device may generate a treatment
recommendation that is independently reviewed by a
healthcare professional, supported by source
information, and subject to an additional software
safety check before any action is taken. Another
device may generate a similar recommendation that
is rapidly accepted within a workflow with little
opportunity for meaningful review. Although the
potential consequence of an incorrect
recommendation may be similar, the likelihood that

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 3
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

the error progresses to patient harm may be very
different.

Medicept therefore recommends that FDA explicitly
consider the effectiveness of the safeguards and
intervention opportunities that exist between a GenAI
output and a potential patient harm.

Rationale: This approach is consistent with
established medical device risk-management
principles in ISO 14971:2019 and the associated
guidance in ISO/TR 24971:2020. Additional AI/MLspecific guidance is provided in ANSI/AAMI
TIR34971:2023, which applies ISO 14971 principles
to machine-learning-enabled medical devices.
Together, these frameworks reinforce the importance
of evaluating the sequence of events by which device
behavior may lead to a hazardous situation and
ultimately to harm, rather than considering device
output and severity in isolation.

The same principle is particularly important for
GenAI-enabled devices. Factors such as whether an
output can be independently verified, whether there is
sufficient time for a user to intervene, whether an
action is reversible, and whether downstream
technological or procedural safeguards are present
can materially affect whether an erroneous output
actually results in harm.

FDA itself identifies several of these factors in
Question 1 of the discussion paper, including
reversibility, downstream safeguards, time pressure,
and traceability. Medicept believes these factors are
best understood collectively as modifiers of the
pathway between GenAI device behavior and patient
harm.

Medicept further recommends that FDA distinguish
risk characterization from the subsequent benefit-risk
determination. Once the risks associated with a
GenAI-enabled function have been characterized, the
acceptability of those risks should be considered in
the context of the anticipated clinical benefit, intended
patient population, severity and urgency of the
condition, available alternatives, and the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 4
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

consequences of delayed or unavailable intervention.
A higher-risk GenAI-enabled function may therefore
have an acceptable benefit-risk profile in a clinical
context where the anticipated benefit is substantial
and the consequences of not acting are significant.

Recommended FDA Clarification / Action:
Medicept recommends that FDA retain the proposed
two-axis framework as an initial risk characterization
tool, while clarifying that device activity and potential
consequence should not be considered sufficient by
themselves to determine overall risk.

FDA should identify factors such as reversibility, time
available for intervention, independent verifiability of
the output, detectability of an error, and downstream
safeguards as explicit risk modifiers. These factors
could be used to determine whether the practical risk
associated with a GenAI-enabled function is higher or
lower than would be suggested by its position on the
two primary axes alone.

FDA should provide a structured qualitative approach
for applying these modifiers so that sponsors can
consistently explain how factors such as degree of
oversight, downstream safeguards, reversibility,
latency, and independent verification alter the
practical risk associated with a GenAI-enabled
function.

This approach would preserve the simplicity of FDA’s
proposed framework while more closely aligning
GenAI risk assessment with established ISO 14971
risk management principles and the real-world
conditions under which medical device outputs lead—
or do not lead—to patient harm.

Comment 2 — Effective Human Oversight Position: Recommend clarification.
Discussion Paper Reference: Section IV, pp.
6–9; Questions 1–4. Medicept Comment: FDA’s proposed risk framework
distinguishes between GenAI-enabled device
Relevant FDA Language / Question: The functions that operate with healthcare professional
activity axis distinguishes, among other (HCP) supervision and those that operate more
functions, action taken with continuous autonomously. The discussion paper also asks how
healthcare professional (HCP) supervision
from action taken in a fully autonomous
fashion. Questions 3 and 4 ask how patient

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 5
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

versus HCP use, and generalist versus use by patients, generalist HCPs, or specialist HCPs
specialist HCP use, should inform risk. should affect the assessment of risk.

Medicept agrees that the presence and type of
human oversight are important considerations.
However, Medicept recommends that FDA distinguish
between nominal human involvement and effective
human oversight.

The fact that an HCP is technically involved in the
workflow does not necessarily mean that the HCP
can meaningfully identify and correct an inappropriate
GenAI output before it contributes to patient harm.
The effectiveness of oversight depends on the
specific task, the information available to the user, the
time available for review, the user’s ability to
independently evaluate the output, and the way the
device presents the information. This assessment
should also consider foreseeable overtrust or
automation bias, particularly where fluent, confident,
or authoritative outputs may reduce the likelihood that
the user questions an incorrect recommendation.

For example, a clinician may technically review a
GenAI-generated recommendation before it is
implemented. However, if the recommendation is
presented without supporting information, appears
highly authoritative, must be reviewed under
significant time pressure, or is one of hundreds of
similar outputs reviewed each day, the practical value
of that human review may be limited. In contrast, a
workflow that presents supporting evidence,
highlights uncertainty, provides adequate time for
review, and requires an independent confirmation
step may provide much stronger risk mitigation.

Medicept therefore recommends that FDA evaluate
the quality and effectiveness of human oversight,
rather than treating the mere presence of an HCP as
a sufficient indicator of reduced risk.

Rationale: This distinction is consistent with
established medical device human factors and risk
management principles. The effectiveness of a risk
control depends not only on whether the control
exists, but also on whether it can reasonably be

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 6
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

expected to function as intended under actual
conditions of use.

For GenAI-enabled devices, human oversight may be
affected by factors such as automation bias, alert
fatigue, high workload, limited transparency into how
an output was generated, mismatch between the
user’s expertise and the complexity of the task, and
the urgency of the clinical decision. These factors
may reduce the likelihood that a user detects an
incorrect or inappropriate recommendation even
when the workflow formally requires human review.
Where human oversight is relied upon as a risk
control, foreseeable overtrust should be considered
when estimating the likelihood that the control will
function as intended.

FDA already recognizes related concerns in the
discussion paper, including the potential for users to
over-rely on fluent or authoritative GenAI outputs.
Medicept believes this issue should be incorporated
directly into the risk framework by asking not simply
whether a human is involved, but whether that human
can reasonably serve as an effective risk control.

This concept is also consistent with IEC 62366-1 and
broader human factors engineering principles, which
focus on how intended users actually interact with a
device under expected use conditions rather than
relying solely on procedural assumptions about user
behavior.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that HCP
involvement should not automatically be treated as a
risk reducing factor.

Instead, FDA should consider whether the intended
user:

• has access to sufficient information, clinical
context, or other appropriate sources to
evaluate the GenAI output rather than relying
solely on the GenAI system itself;
• has appropriate knowledge and experience
for the specific task;

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 7
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

• has adequate time and opportunity to
intervene before a clinically significant action
occurs;
• can reasonably identify errors or
inappropriate outputs;
• is supported by device design features that
promote meaningful review rather than
passive acceptance; and
• operates within a workflow in which
foreseeable factors such as time pressure,
alert fatigue, or automation bias do not
undermine the effectiveness of the oversight.

FDA should consider these factors when determining
whether human involvement meaningfully reduces
the risk associated with a GenAI-enabled function.

This approach would preserve the value of human
oversight as a risk control while avoiding the
assumption that any “human-in-the-loop”
configuration is inherently lower risk.

Where human oversight is relied upon as a risk
control, FDA should consider whether human factors
or usability validation is appropriate to demonstrate
that intended users can reliably identify, interpret, and
intervene upon erroneous or inappropriate GenAI
outputs under realistic conditions of use. Such
evaluation should consider factors including
automation bias, time pressure, workload,
presentation of uncertainty, availability of supporting
information, and the user’s ability to independently
assess the output.

Comment 3 — Harmful Reliance Beyond Position: Recommend modification.
Incorrect Output
Discussion Paper Reference: Section IV, pp. Medicept Comment: FDA’s proposed consequence
5–9; Question 1. axis focuses on the severity of harm that may result
when a user relies on an incorrect GenAI output.
Relevant FDA Language / Question: The Medicept agrees that incorrect outputs are an
proposed consequence axis is framed around important source of risk but recommends that FDA
the severity of harm that may result from using broaden this concept to include harm that may arise
and relying upon an “incorrect output.” from reliance on device behavior even when the
underlying content is not factually incorrect.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 8
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

GenAI-enabled devices can influence clinical decision
making through more than factual accuracy alone.
The same information may lead to very different
outcomes depending on how confidently it is
presented, what information is omitted, whether
uncertainty is communicated, how prominently the
output is displayed, and how the user is expected to
interpret or act on it.

For example, a GenAI-enabled device may provide a
medically accurate statement but present it with a
level of certainty that is not justified by the available
patient information. A device may also provide
technically correct information while failing to identify
an important limitation, contraindication, or alternative
explanation. In other cases, repeated low value
recommendations may increase alert fatigue or
encourage users to defer too readily to the device.

In each of these situations, the output may not be
strictly “incorrect,” yet the interaction between the
device and the user may still create a clinically
meaningful risk.

Medicept therefore recommends that FDA consider
the consequences of reliance on GenAI-enabled
device behavior more broadly, rather than limiting the
risk framework to reliance on factually incorrect
outputs.

Rationale: This broader framing is consistent with
established medical device risk management and
human factors principles. Under ISO 14971, risk is
evaluated based on hazardous situations and
reasonably foreseeable sequences of events that
may lead to harm. A hazardous situation does not
necessarily require a technically incorrect device
output; it may also arise when correct or partially
correct information is used in an unsafe context or
interpreted in a manner that leads to inappropriate
action.

Similarly, human factors principles recognize that the
design and presentation of information can influence
user behavior. Factors such as automation bias,
overreliance, poor communication of uncertainty,

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 9
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

information overload, or insufficient contextualization
can contribute to use related risk even when the
underlying data are accurate.

This distinction may be particularly important for
GenAI-enabled devices because GenAI outputs are
often fluent, persuasive, and context sensitive. These
characteristics may increase the likelihood that users
place greater confidence in the output than is
warranted.

Recommended FDA Clarification / Action:
Medicept recommends that FDA broaden the
consequence dimension of the proposed risk
framework from the consequence of relying on an
“incorrect output” to the consequence of relying on
device behavior that may lead to an unsafe clinical
decision or action.

FDA should clarify that manufacturers should
evaluate not only factual accuracy, but also whether
device behavior could create risk through:

• inappropriate expression of certainty or
confidence;
• omission of information necessary for safe
interpretation;
• failure to communicate relevant limitations or
uncertainty;
• presentation or workflow features that
encourage overreliance;
• repeated or excessive recommendations that
contribute to alert fatigue; or
• other reasonably foreseeable interactions
between the device output and user behavior.

This broader framing would better capture the full
range of clinically relevant risks associated with
GenAI-enabled medical devices while remaining
consistent with established ISO 14971 and human
factors principles.

Comment 4 — System- and Workflow-Level Position: Support with additional consideration.
Risk
Medicept Comment: FDA explains in Section III of
the discussion paper that its existing risk-based

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 10
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Discussion Paper Reference: Section III, pp. approach applies at the individual software function
4–5, and Section IV. level. FDA also notes that a GenAI-enabled device
may include multiple software and hardware
Relevant FDA Language / Question: CDRH
components and states that this discussion paper
states that its current risk-based approach
focuses specifically on the GenAI-enabled component
applies at the software function level and notes
rather than the entire device.
that the discussion paper “addresses the
GenAI-enabled component of that device, but
Medicept agrees that function-level analysis remains
not the entire device.”
important for determining which software functions
are subject to FDA oversight and for characterizing
the role of the GenAI-enabled function. However,
Medicept recommends that FDA clarify that the safety
and effectiveness assessment of a GenAI-enabled
medical device should also consider how that function
interacts with the broader device and the clinical
workflow in which it is used.

The clinical significance of a GenAI output may
depend substantially on what happens after that
output is generated. For example, the same GenAIgenerated treatment recommendation could be:

• displayed to a clinician as advisory
information;
• automatically entered into an electronic order;
• passed to another algorithm for additional
processing;
• used to prioritize a patient for intervention; or
• acted upon by another software or agentic
function.

Although the GenAI output itself may be identical in
each case, the pathway from that output to a clinical
action—and therefore the resulting risk—may be very
different.

Medicept therefore recommends that FDA retain
function level characterization while also requiring
consideration of reasonably foreseeable interactions
between the GenAI-enabled function, other device
functions, external systems, and the intended clinical
workflow.

Rationale: This approach is consistent with
established medical device risk management
principles. ISO 14971 evaluates risk through

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 11
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

reasonably foreseeable sequences of events that
may lead from a hazard to a hazardous situation and
ultimately to harm. In many cases, the relevant
sequence of events extends beyond a single software
function.

For GenAI enabled devices, this becomes particularly
important because the output of one function may
serve as the input to another system or may initiate a
downstream clinical or automated action. A hazard
that appears limited when the GenAI function is
considered in isolation may become more significant
once its role within the complete system is
considered.

This is also consistent with FDA’s existing approach
to multiple-function device products. FDA may
evaluate regulatory oversight at the function level, but
manufacturers must still understand how functions
interact and whether those interactions affect the
safety or effectiveness of the regulated device
function.

Where the GenAI-enabled function exchanges
information with or initiates actions through another
device or system, the risk assessment should also
consider whether interoperability, interface behavior,
data integrity, or communication failures could alter
the safety or effectiveness of the overall workflow.

A system- and workflow-level assessment is therefore
not intended to replace FDA’s function based
regulatory framework. Rather, it ensures that risk
analysis follows the full pathway by which GenAIenabled device behavior may contribute to a clinical
outcome.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that functionlevel regulatory characterization should be
supplemented by system- and workflow-level risk
assessment when the safety or effectiveness of the
GenAI-enabled function depends on downstream
interactions.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 12
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Manufacturers should consider, as applicable:

• how GenAI outputs are used by other
software or hardware functions;
• whether outputs automatically populate,
trigger, or modify downstream actions;
• whether another algorithm, device, or agent
relies upon the GenAI output;
• whether risk controls elsewhere in the system
can detect or mitigate an inappropriate
output;
• how users interact with the GenAI output
within the actual clinical workflow; and
• whether combinations of individually
acceptable functions could create a
hazardous situation when used together.

FDA should therefore retain the function level
approach for regulatory oversight while clarifying that
safety and effectiveness evaluation should follow the
complete, reasonably foreseeable pathway from
GenAI device behavior to clinical action and potential
patient harm.

This would align the proposed GenAI framework with
established ISO 14971 risk management principles
and FDA’s existing approach to multiple-function
device products without creating a separate or
duplicative regulatory framework.

Comment 5 — Action-Directing Behavior Position: Support with clarification.
and Multi-Turn Trajectories
Discussion Paper Reference: Section IV, pp. Medicept Comment: FDA recognizes in Section IV
6–10; Questions 2 and 5. of the discussion paper that GenAI-enabled devices
may exist on a continuum between providing general
Relevant FDA Language / Question: information and directing a user toward a specific
Question 2 asks how wording, specificity, clinical action. FDA also asks in Question 5 how risk
personalization, and context may modify the should be assessed when a multi-turn conversational
risk associated with informational functions. system becomes increasingly action-directing over
Question 5 asks how risk should be assessed the course of an interaction.
when multi-turn conversational devices
migrate from “non-directive” to “action- Medicept agrees that this is an important distinction
directing” information across an exchange. and recommends that FDA evaluate not only the
wording of individual outputs, but also the overall
direction and cumulative effect of the conversation.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 13
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

In a conversational system, a device may begin by
providing general information and then gradually
become more specific, personalized, or directive as
the user continues the interaction. No single response
may appear to cross a clear boundary from
“informational” to “action-directing,” but the
conversation as a whole may still influence the user
toward a particular medical decision or action.

For example, a device may first explain that a
symptom can have several possible causes, then
state that the user’s description is consistent with one
of those causes, then discuss common management
approaches, and finally suggest that monitoring at
home may be reasonable. Each statement may
appear relatively cautious when viewed individually,
but the cumulative interaction may reasonably be
understood by the user as a recommendation.

Medicept therefore recommends that FDA assess
directiveness at the level of the interaction trajectory,
not solely at the level of individual prompts and
responses.

Rationale: This approach is consistent with
established human factors and medical device risk
management principles because risk depends not
only on isolated device outputs, but also on how
users interpret and respond to the device over time.

For GenAI-enabled devices, factors such as
personalization, repeated reinforcement, increasing
specificity, expression of confidence, temporal
urgency, and the user’s prior responses may
gradually increase the influence of the device. These
effects may be particularly important in patient facing
systems, where users may interpret a sequence of
increasingly specific responses as a clinical
recommendation even if the device never uses
explicit directive language.

Similarly, a device that is intended to remain within a
defined informational scope may move beyond that
boundary through multi-turn interaction. This creates
a potential mismatch between intended use and

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 14
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

actual device behavior that may not be detected
through testing focused only on isolated prompts.

This issue also aligns with IEC 62366-1 human
factors principles, which emphasize evaluating device
use in the context of realistic user interaction rather
than relying solely on individual interface elements or
isolated tasks.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that
evaluation of action-directing behavior should
consider the cumulative effect of multi-turn
interactions and the reasonably foreseeable
behavioral impact on the user.

Manufacturers should assess representative
conversational trajectories that include, as applicable:

• increasing personalization over successive
turns;
• repeated requests for more specific advice;
• incomplete or contradictory information from
the user;
• repeated challenges to the device’s initial
response;
• attempts to obtain recommendations outside
the intended scope;
• escalation from general information to
individualized guidance;
• changes in the level of certainty or
confidence expressed by the device; and
• whether the overall interaction is reasonably
likely to influence a clinical or patient action.

FDA should also consider encouraging acceptance
criteria that evaluate the outcome and cumulative
behavior of the interaction, rather than only the
factual correctness or directiveness of individual
responses.

This would help ensure that GenAI-enabled devices
remain within their intended use boundaries and that
risk assessments reflect how conversational systems
are actually used in real-world conditions.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 15
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Comment 6 — Intended-Use Traceability for Position: Recommend clarification.
Competency Assessment
Discussion Paper Reference: Section V, “A Medicept Comment: In Section V of the discussion
Competency-Based Approach for Premarket paper, FDA proposes a competency-based approach
Evaluation of GenAI-Enabled Devices,” pp. for the premarket evaluation of GenAI-enabled
10–18; Questions 7, 9, and 10. medical devices. Under this approach, manufacturers
would identify the competencies needed for the
Relevant FDA Language / Question: CDRH device to perform its intended medical function and
proposes a competency-based approach then evaluate those competencies through nonclinical
consisting of non-clinical device benchmarking benchmarking and, where appropriate, clinical
and clinical confirmation. Question 10 asks confirmation.
how sponsors should establish that benchmark
performance predicts safe and effective real- Medicept supports the use of a competency-based
world behavior for the device’s intended use. framework because GenAI-enabled devices may
perform complex tasks that are not adequately
characterized by a single traditional performance
metric. However, Medicept recommends that FDA
more explicitly require a clear connection between the
competencies being evaluated and the device’s
specific intended use, intended users, clinical claims,
reasonably foreseeable use scenarios, and identified
risks.

A benchmark should not be considered meaningful
simply because it is commonly used to evaluate a
foundation model or because it demonstrates strong
general reasoning or medical knowledge. The
relevant regulatory question is whether the selected
evaluation demonstrates that the finished medical
device can safely and effectively perform the specific
function for which it is intended.

For example, a GenAI-enabled device intended to
summarize oncology records for clinician review may
demonstrate strong performance on a general
medical question-answering benchmark. That result
may provide useful background information regarding
the model’s capabilities, but it does not necessarily
demonstrate that the finished device reliably identifies
clinically important oncology information, avoids
omission of safety-critical information, preserves
relevant temporal relationships, or performs
adequately across the patient populations and
records included within its intended use.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 16
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Medicept therefore recommends that FDA require
sponsors to demonstrate traceability from the
intended medical function and associated risks to the
competencies selected for evaluation and ultimately
to the evidence used to support those competencies.

Rationale: This recommendation is consistent with
established medical device design control, risk
management, and verification and validation
principles.

Under traditional design controls, manufacturers
establish design inputs based on intended use and
user needs and then demonstrate through verification
and validation that the resulting device satisfies those
requirements. Similarly, ISO 14971 requires identified
risks and risk control measures to be evaluated
throughout the product lifecycle.

A competency-based GenAI framework should follow
the same basic logic.

The evaluation pathway should be traceable from:

intended use and clinical claims → reasonably
foreseeable use scenarios → hazards and risk
controls → required device competencies →
evaluation methods and benchmarks →
acceptance criteria → clinical confirmation, where
needed.

Without this traceability, there is a risk that sponsors
may generate extensive benchmark results that are
technically impressive but have limited relevance to
the safety and effectiveness questions presented by
the particular medical device.

In addition to traceability, FDA should expect
manufacturers to characterize the variability of
GenAI-enabled device performance under repeated
and reasonably equivalent conditions. Because
GenAI systems may produce non-identical outputs by
design, repeatability should not necessarily be
defined as identical responses. Rather,
manufacturers should demonstrate that output
variability remains within clinically acceptable bounds

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 17
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

and does not materially affect safety, effectiveness, or
the conclusions of the competency assessment.

This issue is especially important for GenAI because
many existing benchmarks were originally developed
to measure general model capabilities rather than
medical device performance. A high performing
foundation model does not automatically establish the
safety or effectiveness of a finished device that
incorporates that model together with prompts,
retrieval systems, guardrails, user interfaces,
workflow constraints, and other device-specific
controls.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that
competency assessment should begin with the
specific intended use and risk profile of the finished
device rather than with the availability of existing AI
benchmarks.

For each significant competency, manufacturers
should be able to demonstrate the relationship
among:

• the device’s intended use and clinical claims;
• the intended patient population and users;
• the clinical tasks and reasonably foreseeable
use scenarios;
• identified hazards and associated risk
controls;
• the competency or competencies necessary
to perform those tasks safely;
• the benchmark or other evaluation method
used to assess each competency;
• predefined acceptance criteria; and
• clinical confirmation, where necessary to
resolve remaining uncertainty.

FDA should also clarify that general purpose
foundation model benchmark results may provide
supportive information but should not substitute for
device-specific evidence unless the sponsor
demonstrates that the benchmark is sufficiently

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 18
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

representative of the intended medical function and
conditions of use.

This approach would integrate FDA’s proposed
competency-based framework with familiar medical
device design control and ISO 14971 principles and
would help ensure that GenAI benchmarking remains
focused on clinically and regulatory meaningful
performance rather than general model capability.

Medicept further recommends that competency
evaluation address reasonably foreseeable
adversarial conditions when cybersecurity threats
could affect clinically relevant device behavior. For
GenAI-enabled devices, this may include prompt
injection, jailbreak attempts, manipulation of retrieved
content, unauthorized disclosure of sensitive
information, or other attacks capable of bypassing
established safeguards or altering device outputs.
Such evaluation should complement, rather than
replace, the cybersecurity risk-management and
premarket documentation expected under FDA’s
existing medical device cybersecurity framework.

Comment 7 — Construct Validity and Position: Recommend clarification.
Sponsor-Developed Benchmarks
Discussion Paper Reference: Section V.B Medicept Comment: In Section V of the discussion
and Section V.E, pp. 12–18; Question 10. paper, FDA recognizes that existing public
benchmarks may have important limitations for
Relevant FDA Language / Question: CDRH evaluating GenAI-enabled medical devices. FDA
recognizes that publicly available notes concerns such as data contamination,
benchmarking assets may be affected by data benchmark saturation, and limited representation of
contamination, saturation, and limited real- real-world clinical conditions, and asks what role
world representativeness, and asks what role sponsor-developed benchmarks should play when
sponsor-developed benchmarks should play. existing benchmarks are not sufficient.

Medicept agrees that publicly available benchmarks
may not always be appropriate for evaluating a
specific GenAI-enabled medical device. Medicept
also believes that sponsor-developed benchmarks
can be appropriate, and in some cases necessary,
when existing benchmarks do not adequately
represent the device’s intended use, intended

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 19
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

population, clinical task, or reasonably foreseeable
failure modes.

However, the regulatory value of a sponsordeveloped benchmark should depend on whether the
benchmark is demonstrably fit for purpose, rather
than simply on who developed it.

For example, a sponsor may develop a benchmark
specifically to test whether a GenAI-enabled device
correctly identifies clinically important information
from complex longitudinal medical records. Such a
benchmark may be more relevant to the intended use
than a widely cited public medical question-answering
benchmark. At the same time, a sponsor-developed
benchmark could produce misleading reassurance if
the test cases are too easy, insufficiently
representative, derived from the same data used
during development, or selected in a way that
minimizes known failure modes.

Medicept therefore recommends that FDA focus on
the validity, representativeness, and clinical relevance
of the benchmark rather than treating public or thirdparty benchmarks as inherently more credible than
sponsor-developed evaluations.

Rationale: This approach is consistent with
established medical device verification and validation
principles.

In traditional medical device evaluation, a test method
is not considered adequate merely because it is
standardized or commonly used. The manufacturer
must demonstrate that the method is appropriate for
the specific requirement being evaluated and that the
test conditions adequately challenge the device for its
intended use.

The same principle should apply to GenAI
benchmarking.

A useful benchmark should measure the competency
it claims to assess and should do so under conditions
that are relevant to the finished device. This concept
is often described as construct validity: whether the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 20
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

evaluation actually measures the capability or
performance characteristic that is important for the
device’s intended use.

For GenAI-enabled medical devices, benchmark
validity may depend on factors such as:

whether the dataset reflects the intended patient
population;

whether the test cases represent realistic clinical
complexity;

whether clinically important edge cases and known
failure modes are included;

whether the benchmark reflects the actual inputs the
finished device will encounter;

whether the scoring method reflects the clinical
significance of different errors;

whether the evaluation data are sufficiently
independent from model development and tuning
activities; and

whether the benchmark challenges the device under
conditions that are reasonably representative of
actual use.

A benchmark may therefore be public and
independent but still have poor clinical relevance.
Conversely, a carefully designed sponsor-developed
benchmark may provide stronger evidence if it is
appropriately justified and independently evaluated.

This distinction is particularly important because
many existing AI benchmarks were created to
compare general purpose model performance, not to
establish medical device safety and effectiveness.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that sponsordeveloped benchmarks may be appropriate when

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 21
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

existing public or third-party benchmarks do not
adequately address the intended medical function.

FDA should evaluate the adequacy of a benchmark
based on factors such as:

• the clinical relevance of the test cases;
• representation of the intended population and
conditions of use;
• inclusion of clinically meaningful edge cases
and known failure modes;
• independence of the evaluation data from
model development and tuning;
• appropriateness of the scoring methodology;
• use of predefined acceptance criteria;
• transparency regarding benchmark
construction and case selection; and
• the relationship between benchmark
performance and identified device risks.

Where appropriate, FDA should also encourage
independent adjudication, blinded scoring,
sequestered evaluation datasets, or other controls
that reduce the potential for bias in sponsordeveloped testing.

For benchmarks that rely on human or automated
adjudication of open-ended GenAI outputs, sponsors
should demonstrate that the scoring methodology
produces sufficiently consistent conclusions across
repeated assessments and qualified evaluators. This
does not necessarily require identical GenAI outputs,
provided variability remains within predefined
clinically acceptable performance bounds.

FDA should avoid creating a hierarchy in which public
or third-party benchmarks are presumed to be more
valid solely because of their source. The central
regulatory question should be whether the benchmark
provides credible evidence that the finished GenAIenabled medical device can safely and effectively
perform its intended function under reasonably
representative conditions of use.

This approach would align GenAI benchmarking with
familiar medical device verification and validation

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 22
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

principles while allowing sponsors to develop
clinically meaningful evaluations when existing
benchmark tools are not fit for purpose.

Comment 8 — Risk-Proportionate Clinical Position: Strongly support with recommendation.
Confirmation
Discussion Paper Reference: Section V.C Medicept Comment: In Section V.C of the
and Section V.E, pp. 14–18; Question 11. discussion paper, FDA explains that clinical
confirmation may be needed in addition to nonclinical
Relevant FDA Language / Question: CDRH benchmarking to demonstrate that a GenAI-enabled
states that clinical confirmation “might not medical device performs safely and effectively in its
require a prospective clinical study in every intended clinical context. FDA also states that clinical
case” and asks how sponsors should select a confirmation may not require a prospective clinical
confirmation approach tailored to intended use study in every case and asks how sponsors should
and proportionate to the device’s risk profile. determine the appropriate type and level of clinical
confirmation based on the device’s intended use and
risk profile.

Medicept strongly supports this risk-proportionate
approach.

The incorporation of GenAI into a medical device
should not, by itself, create a presumption that
prospective clinical investigation is required. Rather,
the type and rigor of clinical confirmation should be
determined by the specific safety and effectiveness
questions that remain after nonclinical evaluation is
complete.

For some GenAI-enabled functions, retrospective
data, simulated-use testing, shadow-mode
deployment, standardized patient scenarios, or expert
adjudication may provide sufficient evidence when
the intended function is well characterized and the
remaining uncertainty is limited.

For other devices—particularly those that perform
more autonomous functions, support high
consequence decisions, operate in novel clinical
workflows, or are used in vulnerable populations—
prospective clinical evidence may be necessary to
resolve uncertainties that cannot be adequately
addressed through nonclinical or retrospective
methods.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 23
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Medicept therefore recommends that FDA explicitly
frame clinical confirmation as an evidence escalation
process, in which the level of clinical evidence
increases as residual uncertainty and potential clinical
consequence increase.

Rationale: This approach is consistent with
established medical device principles and FDA’s least
burdensome approach to evidence generation.

Clinical evidence requirements should be driven by
the regulatory question that remains unresolved—not
by the use of a particular technology.

For GenAI-enabled devices, the need for stronger
clinical confirmation may depend on factors such as:

the severity of potential harm if the device performs
inadequately;

the degree of autonomy or directiveness of the
device;

whether the intended workflow is novel or well
established;

whether device performance depends heavily on
human interaction;

the effectiveness of available risk controls;

the degree to which nonclinical benchmarking
reproduces actual conditions of use;

whether the intended population includes vulnerable
or underrepresented groups; and

the amount of uncertainty that remains after bench,
analytical, simulated use, and other nonclinical
evaluations.

For example, a GenAI-enabled function that
summarizes information for clinician review may be
adequately supported by rigorous retrospective
evaluation and human factors validation if the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 24
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

underlying clinical risks are well controlled. In
contrast, a function that autonomously selects or
initiates a high risk treatment may require prospective
clinical evaluation because the remaining uncertainty
and consequence of failure are substantially greater.

This type of graduated approach is familiar in medical
device regulation. Sponsors routinely determine the
amount of evidence needed based on intended use,
risk, novelty, existing knowledge, and the adequacy
of available nonclinical evidence. GenAI should not
replace that principle with a technology-based
requirement for a particular study design.

Recommended FDA Clarification / Action:
Medicept recommends that FDA establish a riskproportionate framework for determining the level of
clinical confirmation appropriate for GenAI-enabled
medical devices.

FDA should clarify that the strength of clinical
evidence should increase based on factors such as:

• residual uncertainty following nonclinical
evaluation;
• severity of potential harm;
• degree of autonomy or action directing
behavior;
• novelty of the clinical workflow;
• reliance on human-device interaction;
• vulnerability of the intended population;
• adequacy of existing risk controls; and
• representativeness of available retrospective
or simulated evidence.

FDA should also clarify that prospective clinical
studies are not inherently required for GenAI-enabled
devices when other evidence sources adequately
address the relevant safety and effectiveness
questions.

A graduated framework could allow, where
appropriate, progression from:

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 25
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

retrospective or simulated evaluation → shadowmode or real-world observational evaluation →
prospective clinical investigation

as the level of unresolved clinical uncertainty and
potential consequence increases.

This approach would preserve FDA’s leastburdensome principles while ensuring that the level of
clinical evidence remains commensurate with the
actual risk and uncertainty associated with the
finished GenAI-enabled medical device.

Comment 9 — Selecting Meaningful Position: Support with recommendation.
Performance Comparators
Discussion Paper Reference: Section V.D.1 Medicept Comment: In Section V.D.1 of the
and Section V.E, pp. 16–18; Questions 14 and discussion paper, FDA discusses several possible
15. approaches for evaluating GenAI-enabled medical
device performance, including comparison with
Relevant FDA Language / Question: CDRH clinician panels, median clinician performance,
considers comparison with clinician panels, traditional clinical reference standards, and human-AI
median clinician performance, traditional team performance. FDA also asks whether
clinical reference standards, and human-AI performance should sometimes be assessed relative
team performance. Question 15 asks whether to the care, technology, or course of action that would
performance should sometimes be assessed likely occur in the absence of the device.
relative to the care, technology, or course of
action likely to occur in the absence of the Medicept recommends that FDA make this latter
device. principle more explicit.

The most appropriate comparator should depend on
the intended role of the device in the clinical workflow,
rather than defaulting to individual clinician
performance as the universal benchmark.

For example, if a GenAI-enabled device is intended to
assist a clinician by summarizing records, identifying
relevant information, or supporting prioritization, the
most meaningful comparison may be the
performance of the clinician using the device versus
the clinician using the current standard workflow. In
that situation, comparing the device alone against an
individual clinician may not reflect how the product is
actually intended to be used.

Conversely, if a device is intended to autonomously
perform a task that would otherwise be performed by

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 26
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

a healthcare professional, a direct comparison
between the device and clinician performance may be
appropriate.

In other circumstances, the relevant comparator may
be another technology, delayed specialist review,
unaided clinical judgment, no intervention, or another
real-world care pathway.

Medicept therefore recommends that comparator
selection be based on the realistic counterfactual—
that is, what would reasonably be expected to occur
in the absence of the GenAI-enabled device.

Rationale: This approach is consistent with
established medical device clinical evidence
principles, where the choice of comparator should be
driven by the clinical question being evaluated.

A comparator is only meaningful if it helps determine
whether the device provides a safe and effective
benefit relative to the actual alternative available to
the intended patient or user.

For GenAI-enabled devices, this distinction may be
especially important because many products are
designed to augment, rather than replace, human
decision making. In those cases, the regulated
product is effectively being used as part of a humandevice system. Evaluating the AI component in
isolation may not adequately characterize the safety
or effectiveness of the intended use.

For example, if a GenAI-enabled tool is intended to
reduce the time required for a radiologist to review
prior records while maintaining diagnostic quality, the
relevant question may not be whether the GenAI
system can independently outperform the radiologist.
The more clinically meaningful question may be
whether radiologists using the tool perform at least as
safely and effectively as radiologists using the
existing workflow, while achieving the intended
efficiency benefit.

Similarly, if a GenAI-enabled device is intended to
identify patients who should be prioritized for

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 27
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

specialist review, comparison with an idealized
specialist may be less informative than comparison
with the actual triage process the device is intended
to supplement or replace.

Selecting the wrong comparator could therefore
produce evidence that is technically impressive but
does not answer the actual regulatory or clinical
question posed by the device’s intended use.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that sponsors
should select performance comparators based on the
intended role of the GenAI-enabled device and the
realistic care pathway that would occur in the
absence of the device.

FDA should consider whether the most appropriate
comparator is:

• the intended user performing the task without
the device;
• the intended user performing the task with
the device;
• another currently available device or
technology;
• an existing clinical workflow or standard of
care;
• specialist review;
• delayed intervention or no intervention; or
• another clinically relevant counterfactual.

Sponsors should explain why the selected
comparator appropriately addresses the device’s
safety and effectiveness claims and reflects the
intended conditions of use.

FDA should also clarify that comparison with clinician
performance should not be treated as the default
requirement when the device is intended to augment
rather than replace clinical judgment.

This approach would help ensure that performance
evidence answers the clinically meaningful question:
not simply whether a GenAI system can perform a

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 28
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

task, but whether use of the finished device improves
or appropriately maintains the safety and
effectiveness of the care pathway in which it is
intended to operate.

Comment 10 — Postmarket Reassessment: Position: Support with recommendation.
Periodic and Event-Triggered
Discussion Paper Reference: Section VI, pp. Medicept Comment: In Section VI of the discussion
19–22; Questions 18 and 19. paper, FDA asks how GenAI-enabled medical
devices should be monitored after deployment,
Relevant FDA Language / Question: CDRH including whether manufacturers should conduct
asks about periodic re-benchmarking, sample- periodic re-benchmarking, clinician review, or other
based clinician review, performance- forms of performance reassessment. FDA also asks
degradation monitoring, and how the cadence how the timing and triggering events for
and triggering events for reassessment should reassessment should be determined.
be determined.
Medicept supports ongoing postmarket performance
monitoring and periodic reassessment. However,
Medicept recommends that FDA avoid relying
primarily on a fixed calendar-based approach.

For GenAI-enabled devices, meaningful changes may
occur between scheduled review periods, and the
appropriate reassessment cadence may depend in
part on the rate and nature of change in the
underlying model and surrounding system
configuration. Changes to prompts, retrieval sources,
guardrails, orchestration logic, connected tools, or
other configuration elements may materially affect
device behavior even when the underlying foundation
model itself has not changed.

A device that was re-benchmarked six months ago
may therefore still require immediate reassessment if
a clinically meaningful change occurs today.

Medicept recommends a hybrid postmarket approach
that combines routine periodic review with predefined,
risk-based triggering events that prompt earlier
reassessment when appropriate.

Rationale: This approach is consistent with
established medical device postmarket surveillance,
complaint handling, CAPA, and change control
principles.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 29
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Manufacturers do not generally wait for an annual
review to evaluate a newly identified safety signal or a
design change that may affect device performance.
Instead, the nature and significance of the new
information determine whether additional
investigation, testing, risk analysis, or corrective
action is required.

The same principle should apply to GenAI-enabled
devices.

For example, reassessment may be warranted when:

• the underlying foundation model is updated
or replaced;
• retrieval sources or clinical knowledge bases
are materially changed;
• prompts, guardrails, or orchestration logic are
modified;
• a user interface change may affect how
outputs are interpreted;
• a new patient population or use environment
is introduced;
• complaint or adverse event trends suggest a
new failure mode;
• monitoring identifies clinically meaningful
performance drift;
• refusal, escalation, or deferral behavior
changes unexpectedly; or
• an agentic system gains access to new tools
or actions.

Some of these changes may have little or no effect on
safety or effectiveness, while others may materially
affect one or more validated device competencies. A
fixed reassessment interval alone may not distinguish
between these situations.

This is also consistent with ISO 14971 lifecycle risk
management principles, which require manufacturers
to review production and post production information
and determine whether new information affects
previously established risk estimates, risk controls, or
overall residual risk.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 30
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Recommended FDA Clarification / Action:
Medicept recommends that FDA describe postmarket
reassessment for GenAI-enabled devices as a
combination of:

1. Periodic reassessment, performed at a
frequency appropriate to the device’s risk
profile, intended use, and rate of
technological change; and
2. Event-triggered reassessment initiated when
new information or a modification could
reasonably affect device safety,
effectiveness, or previously established
competencies.

FDA should encourage manufacturers to
prospectively define, where practicable:

• the types of changes or postmarket signals
that trigger reassessment;
• the level of reassessment associated with
different trigger categories;
• the performance thresholds or trends that
warrant escalation;
• the relationship between reassessment
findings and CAPA, risk management, or
regulatory reporting processes; and
• the conditions under which a broader rebenchmarking or clinical evaluation may be
necessary.

Reassessment should be integrated into the device
development and deployment lifecycle so that
relevant modifications are evaluated before
implementation. For frequently updated GenAI
systems, manufacturers should establish predefined
change-assessment and regression-testing gates that
determine whether a modification can proceed,
requires targeted revalidation, or warrants broader
reassessment.

FDA should also clarify that periodic re-benchmarking
is a useful surveillance mechanism, but should not

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 31
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

substitute for prompt evaluation of a meaningful
change or emerging safety signal.

This approach would align GenAI postmarket
oversight with established total product lifecycle, ISO
14971, CAPA, and design change principles while
providing manufacturers with a practical framework
for responding to both expected and unexpected
changes over time.

Comment 11 — Machine-Based Position: Support with safeguards.
Supervisory Agents
Discussion Paper Reference: Section VI.D, Medicept Comment: In Section VI of the discussion
p. 21; Question 20. paper, FDA asks whether postmarket monitoring of
GenAI-enabled medical devices could be supported
Relevant FDA Language / Question: by machine-based supervisory agents and, if so, how
Question 20 asks whether postmarket the reliability of those supervisory agents should be
monitoring can be facilitated by machine- evaluated.
based supervisory agents and what
considerations, including evaluation and
Medicept agrees that automated supervisory systems
reliability of the supervisory agent itself, CDRH may be useful as part of a broader postmarket
should take into account. monitoring strategy. A supervisory agent could
potentially identify abnormal outputs, detect
performance drift, flag deviations from expected
behavior, or help prioritize outputs for human review.

However, Medicept recommends that FDA clarify that
the use of a second AI system does not, by itself,
provide independent assurance that the primary
GenAI-enabled device is performing safely.

If the supervisory agent and the monitored device rely
on the same or closely related foundation models,
training sources, retrieval data, reasoning
approaches, or other technical dependencies, the two
systems may be vulnerable to the same failure
modes. In that situation, the supervisory agent may
fail to detect the very errors it is intended to identify.

Where automated supervisory mechanisms are used,
manufacturers should consider whether diversity in
models, data sources, evaluation methods, or non-AI
controls provides meaningful independence from the
monitored device. The use of multiple supervisory

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 32
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

agents should not itself be presumed to provide
independent assurance; manufacturers should
demonstrate that the supervisory architecture
reduces relevant common-mode or correlated failure
risks and includes appropriate mechanisms for
disagreement, escalation, and human review where
warranted.

Rationale: This concern is consistent with
established medical device risk management and
quality system principles. A risk control is only
effective if it is capable of detecting or mitigating the
failure mode for which it was designed.

For example, if both a GenAI-enabled device and its
supervisory agent rely on the same underlying
foundation model or similar reasoning architecture, a
systematic error in that model may affect both
systems in the same way. The supervisory agent may
therefore confirm an incorrect output rather than
detect it.

Similarly, if both systems rely on the same external
clinical knowledge source or retrieval database, an
error or omission in that source could propagate
through both the primary device and the supervisory
mechanism.

This creates a risk of correlated failure and may
provide a false sense of assurance if the supervisory
agent is treated as an independent control without
appropriate validation.

The issue is analogous to established engineering
principles in which redundancy only provides
meaningful risk reduction when redundant controls
are sufficiently independent and are not vulnerable to
the same common cause failure.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that machinebased supervisory agents may be used as part of
postmarket monitoring, but that manufacturers should
demonstrate the suitability and independence of the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 33
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

supervisory mechanism for the specific failure modes
being monitored.

Manufacturers should consider, as applicable:

• whether the supervisory agent relies on the
same foundation model or model family as
the monitored device;
• whether the two systems share training
sources, retrieval data, knowledge bases, or
other dependencies;
• whether the supervisory agent has been
validated for the specific failure modes it is
intended to detect;
• the sensitivity and specificity of the
supervisory mechanism;
• how false-positive and false-negative
supervisory alerts are managed;
• how supervisory alerts are reviewed and
escalated;
• whether independent human review or other
non-AI monitoring mechanisms remain
necessary; and
• whether the supervisory system itself is
subject to appropriate change control and
performance monitoring.

FDA should also clarify that automated supervision
should supplement, rather than replace, manufacturer
responsibility for postmarket surveillance, complaint
evaluation, CAPA, risk management, and continued
assurance of device safety and effectiveness.

This approach would allow manufacturers to take
advantage of automated monitoring while avoiding
the assumption that “AI supervising AI” inherently
provides an independent or sufficient safety control.

Comment 12 — Impact-Based Re- Position: Recommend an impact-based framework.
Benchmarking and PCCP Change Control
Discussion Paper Reference: Section VI, pp. Medicept Comment: In Section VI of the discussion
21–22; Questions 22 and 23. paper, FDA asks how the extent of re-benchmarking
or other supporting evidence should scale to the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 34
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Relevant FDA Language / Question: nature and expected impact of a modification. FDA
Question 22 asks how the extent of re- also asks how Predetermined Change Control Plan
benchmarking or other evidence should scale (PCCP) concepts could be adapted for GenAIto the nature and expected impact of a enabled devices when future modifications cannot
modification. Question 23 asks how PCCP always be fully specified in advance.
concepts might be adapted when the nature or
scope of future modifications cannot be fully Medicept supports an impact-based approach to
prespecified. reassessment.

For GenAI-enabled devices, not all changes have the
same potential effect on safety or effectiveness. A
change to the visual presentation of an output, for
example, may have a very different impact from a
change to the underlying foundation model, retrieval
logic, clinical reasoning process, or autonomous
action pathway.

Medicept therefore recommends that FDA avoid an
expectation that every modification requires complete
re-benchmarking or revalidation. Instead,
manufacturers should first evaluate how the proposed
change could affect the device’s intended use,
identified hazards, existing risk controls, validated
competencies, user interaction, and previously
generated evidence.

The scope of additional testing should then be based
on that impact assessment.

Rationale: This approach is consistent with
established medical device design control and
change management principles.

Medical device manufacturers routinely assess
whether a design change affects previously
established requirements, risks, verification or
validation results, labeling, or regulatory status. The
amount of additional testing depends on the nature
and significance of the change rather than on the
mere fact that a change occurred.

The same principle is particularly important for
GenAI-enabled devices because these products may
undergo frequent modifications during their lifecycle.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 35
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

For example:

• a minor formatting change may require
focused verification and human factors
assessment;
• an update to a retrieval database may require
targeted testing of knowledge accuracy,
source integrity, and affected clinical
competencies;
• a modification to system prompts or
guardrails may require regression testing
focused on behavior boundaries and failure
modes;
• a foundation-model update may affect
multiple competencies and require broader
re-benchmarking; and
• a change to autonomous decision-making or
action execution may warrant broader
validation, clinical confirmation, or regulatory
review.

Requiring the same level of testing for each of these
changes would not be risk proportionate and could
create unnecessary burden without improving patient
safety.

This approach also aligns with ISO 14971, which
requires manufacturers to determine whether
changes introduce new hazards, alter existing risk
estimates, or affect the effectiveness of risk controls.

Recommended FDA Clarification / Action:
Medicept recommends that FDA establish an impactbased framework in which the scope of rebenchmarking, regression testing, clinical
confirmation, and regulatory review is determined by
the expected effect of the modification.

Manufacturers should assess whether a change
could affect:

• the device’s intended use or clinical claims;
• identified hazards or hazardous situations;
• existing risk control measures;
• one or more previously validated
competencies;

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 36
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

• the intended user interaction or workflow;
• the validity of previously generated
benchmark or clinical evidence; or
• the regulatory basis for the marketed device.

Where appropriate, PCCPs should include predefined
performance thresholds, safety limits, or other
acceptance boundaries that must continue to be met
after implementation of an authorized change.

FDA should also clarify that PCCPs for GenAIenabled devices do not necessarily need to
prespecify every future implementation detail.
Instead, a PCCP could prespecify:

• categories or boundaries of allowable
modifications;
• the device competencies that may be
affected;
• predefined evaluation methods;
• acceptance criteria;
• required monitoring following implementation;
and
• conditions that trigger broader reassessment
or a new marketing submission.

This approach would allow manufacturers to manage
foreseeable GenAI changes in a controlled and
transparent manner while preserving the fundamental
medical device principle that the extent of
reassessment should be proportional to the impact of
the change.

Comment 13 — Third-Party Foundation Position: Support MAF concept with additional
Models and Manufacturer Control safeguards.
Discussion Paper Reference: Section VI.D
and Section VII.A, pp. 21–23; Questions 24 Medicept Comment: In Section VI and Section VII of
and 25. the discussion paper, FDA asks how manufacturers
should detect, evaluate, and respond to changes
Relevant FDA Language / Question: made by third-party foundation-model developers.
Question 24 asks how a manufacturer can FDA also discusses the potential use of voluntary
detect, evaluate, and respond to third-party Foundation Model Device Master Files (MAFs),
foundation-model changes and what through which a foundation-model developer could
contractual, technical, or PCCP mechanisms provide certain proprietary information directly to FDA
could provide reasonable assurance. Section while the medical device manufacturer remains
VII.A discusses voluntary Foundation Model

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 37
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Device Master Files (MAFs), while stating that responsible for demonstrating the safety and
sponsors remain responsible for their finished effectiveness of the finished device.
devices.
Medicept supports FDA’s exploration of Foundation
Model MAFs because they could provide FDA with
information that individual device manufacturers may
not otherwise be able to obtain. However, Medicept
recommends that FDA also address the practical
information gap that may remain for the finisheddevice manufacturer.

A medical device manufacturer may rely on a thirdparty foundation model but have limited visibility into
changes to that model’s architecture, behavior,
guardrails, training data, performance characteristics,
or update schedule. In some cases, the manufacturer
may also have limited contractual ability to require
advance notification or detailed technical disclosure
from a large foundation-model provider.

This creates a potential mismatch: the medical device
manufacturer remains responsible for the safety and
effectiveness of the finished device but may not have
access to all of the information needed to determine
whether a third-party model change affects previously
established device performance, risk controls, or
validation evidence.

Medicept therefore recommends that FDA distinguish
between information that may remain confidential
within a Foundation Model MAF and information that
the finished device manufacturer must be able to
access in order to maintain effective lifecycle control.

This approach is consistent with the risk-based
supplier control requirements incorporated into the
QMSR through ISO 13485:2016. Where a third-party
foundation model materially affects device safety or
effectiveness, the manufacturer should apply supplier
controls commensurate with that impact,
supplemented by technical controls where contractual
leverage is limited.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 38
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Rationale: This issue is consistent with established
medical device supplier control, design change, and
risk management principles.

Under a traditional quality system, manufacturers are
expected to understand and control changes to
critical supplied products or services when those
changes may affect the safety or effectiveness of the
finished device. In many conventional supplier
relationships, this may be addressed through supplier
agreements, change notification requirements,
incoming controls, qualification activities, and ongoing
supplier monitoring.

Third-party foundation models may not fit neatly
within that traditional model.

A manufacturer may depend on a commercially
available foundation model provided through an API
or hosted service. The model provider may update
the model, modify safety controls, change output
behavior, deprecate a version, or alter supporting
infrastructure without giving the medical device
manufacturer the same level of visibility or control that
would typically be expected from a critical medical
device supplier.

For that reason, traditional contractual supplier
controls may need to be supplemented by technical
controls.

Examples may include version pinning, automated
detection of model changes, regression testing before
deployment of a new version, fallback configurations,
rollback capability, audit logging, and monitoring for
changes in output behavior.

This approach is also consistent with ISO 13485
supplier control principles and ISO 14971 lifecycle
risk management. If a third-party model change can
affect the device’s safety or effectiveness, the
manufacturer needs a practical mechanism to detect,
assess, and control that change even when direct
contractual leverage is limited.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 39
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Recommended FDA Clarification / Action:
Medicept recommends that FDA continue to explore
Foundation Model MAFs while also defining a
minimum set of safety-relevant information that
should be available to medical device manufacturers
who rely on those models.

That information could include, as applicable:

• unique identification of the foundation model
and version used in the device;
• notification of material model updates or
deprecation;
• known healthcare-relevant limitations;
• safety-relevant changes in model behavior;
• changes to guardrails, output restrictions, or
other controls;
• information necessary to determine whether
prior device validation remains applicable;
• availability of version pinning or rollback; and
• sufficient change history to support
manufacturer impact assessment.

FDA should also clarify that manufacturers may use a
combination of contractual and technical controls to
manage third-party foundation-model risk.

Where traditional supplier agreements are feasible,
manufacturers should use them. Where contractual
leverage is limited, FDA should recognize
appropriately justified technical controls—such as
version locking, change detection, regression gates,
fallback configurations, and post-change monitoring—
as important components of the manufacturer’s
lifecycle control strategy.

This approach would preserve manufacturer
accountability for the finished medical device while
acknowledging the practical realities of relying on
large third-party foundation-model providers.

Comment 14 — Agentic AI: End-to-End Position: Recommend additional evaluation
Action-Sequence Assurance principles.
Discussion Paper Reference: Section VII.B–
C, pp. 22–23; Question 26. Medicept Comment: In Section VII of the discussion
paper, FDA asks what additional considerations

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 40
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

Relevant FDA Language / Question: CDRH should apply to GenAI-enabled devices that
defines agentic AI systems as systems that incorporate agentic AI capabilities, including systems
autonomously plan and execute multi-step that can autonomously plan and execute multi-step
tasks, use external tools, or take actions tasks, use external tools, or take actions across a
across a sequence of steps and asks what sequence of steps.
additional considerations should apply to
agentic GenAI-enabled devices. Medicept agrees that agentic AI introduces additional
safety considerations beyond those associated with a
GenAI system that only generates information or
recommendations.

For an agentic system, it may not be sufficient to
evaluate the correctness of individual outputs or
individual steps in isolation. A system may make
several decisions that each appear reasonable when
viewed separately, yet the overall sequence of
actions may still lead to an unsafe result.

For example, an agentic system could correctly
identify a clinical issue, select an appropriate next
step, and access the correct external tool, but then
apply the wrong parameter, fail to recognize a change
in patient condition, or continue executing after a
condition arises that should have triggered human
intervention. An error early in the sequence may also
propagate through later steps and become
increasingly difficult to detect or reverse.

Medicept therefore recommends that FDA evaluate
agentic GenAI-enabled devices at the level of the
complete action sequence, including how the system
plans, executes, monitors, and, when necessary,
stops or escalates its actions.

For higher-consequence agentic functions,
manufacturers should also maintain sufficient
auditability to reconstruct the system’s actions,
decision sequence, tool use, and escalation behavior.

Rationale: This approach is consistent with
established medical device system safety and risk
management principles.

Under ISO 14971, risk analysis considers reasonably
foreseeable sequences of events that may lead to
hazardous situations and harm. For agentic AI, the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 41
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

sequence itself may be an important source of risk
because one action may create the conditions for the
next action.

Accordingly, safe performance cannot necessarily be
inferred from successful validation of individual
components.

Agentic systems may introduce risks such as:

• incorrect planning or sequencing of otherwise
appropriate actions;
• propagation of an early error through later
steps;
• inappropriate selection or use of an external
tool;
• failure to detect when assumptions used in
an earlier step are no longer valid;
• failure to recognize when human intervention
is required;
• actions that are difficult or impossible to
reverse;
• excessive system permissions or access;
• inability to recover safely from an
intermediate failure; or
• continued execution despite increasing
uncertainty or evidence that the planned
action is no longer appropriate.

These concerns are especially important when an
agent can directly affect a clinical workflow, modify
records, initiate orders, interact with another medical
device, or otherwise take an action that has
consequences beyond generating information.

The concept is also consistent with established
software and systems engineering principles, where
verification of individual components does not
eliminate the need for integration and system-level
validation.

Recommended FDA Clarification / Action:
Medicept recommends that FDA clarify that
evaluation of agentic GenAI-enabled medical devices
should include end-to-end assessment of the

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 42
 DISCUSSION PAPER SECTION /
MEDICEPT DETAILED COMMENT
LANGUAGE

complete action pathway, not only evaluation of
individual model outputs or component functions.

Manufacturers should evaluate, as applicable:

• how the system develops and modifies multistep plans;
• whether actions occur in the correct
sequence;
• whether the system uses external tools and
data sources appropriately;
• whether permissions and action boundaries
are appropriately constrained;
• how errors or unexpected conditions in one
step affect subsequent steps;
• whether intermediate results are
independently checked before the next action
occurs;
• whether the system can recognize when
human review or intervention is required;
• whether actions can be safely halted or
reversed;
• whether the system can transition to a safe
state when uncertainty exceeds predefined
limits; and
• whether the complete end-to-end sequence
remains safe under reasonably foreseeable
failure conditions.

FDA should also consider whether acceptance
criteria for agentic systems should be tied to the
highest clinically consequential action the system can
execute or materially initiate, together with the
opportunity for effective human or system intervention
before that action occurs.

This approach would align the regulation of agentic
GenAI-enabled medical devices with established ISO
14971 and system-level verification and validation
principles and would help ensure that safe
component performance is not mistaken for safe
overall system behavior.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 43
Closing Perspective
Medicept encourages CDRH to build any future regulatory approach for GenAI-enabled medical devices
around the safety and effectiveness of the finished device in its intended clinical context. GenAI
introduces important challenges—including probabilistic behavior, open ended interaction, external
foundation-model dependencies, and potentially autonomous multi-step action—but these challenges do
not necessarily require an entirely separate regulatory philosophy. Existing medical device principles of
risk management, human factors, design controls, clinical validation, supplier controls, change
management, and total product lifecycle oversight remain highly relevant when adapted to the
characteristics of GenAI.

A framework that emphasizes effective intervention, intended use traceability, proportionate evidence,
impact-based change assessment, and system-level assurance can support both meaningful patient
protection and efficient development of beneficial GenAI-enabled devices.

200 Homer Ave 508-231-8842
Ashland, MA 01721 medicept.com 44