Administrative Tasks Inc.

Please see the attached comments submitted by Administrative Tasks Inc. regarding Docket No. FDA-2026-N-7874.

​​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​

19 October 2026
Re: Comments on FDA’s Discussion Paper on Generative
U.S. Food and Drug AI-Enabled Medical Devices
Administration

Docket No. FDA-2026-N-7874 INTRODUCTION

Considerations for the
Administrative Tasks Inc. (Admin), a healthcare technology company focused on applying AI to
Regulation of Generative expand clinical capacity, submits these comments on FDA’s discussion paper, Considerations for
AI-Enabled Medical Devices: the Regulation of Generative AI-Enabled Medical Devices (Docket No. FDA-2026-N-7874).
Discussion Paper and Request Generative AI will change the scale at which healthcare can be delivered. Devices that assess
for Feedback
patients, guide or carry out treatment, and follow patients over time will create additional clinical
Submitted by capacity, enabling health systems to meet needs that currently go unmet and provide more timely,
continuous care. FDA’s emerging framework will help determine how that capacity reaches
Administrative Tasks Inc. patients. We recommend a clear pathway from demonstrated clinical competence to
authorization, supported by continuing evidence as devices improve.
Evaluation should establish the safety and effectiveness of the device for its intended use, including
interactions among its assessment, decisions, and subsequent actions and the performance of its
device-user interface. A device’s intended use, including its degree of autonomy, should be
supported by evidence of its performance in that use.
FDA’s discussion paper explores a two-axis framework for characterizing risk: the horizontal axis
describes the degree and independence of the software activity, while the vertical axis describes the
severity of harm from relying on an incorrect output. It also considers competency-based
evaluation through device benchmarking and clinical confirmation. These provide a useful
starting point for matching evidence to intended use, provided the assessment accounts for the
effectiveness of safeguards and the opportunity to detect, prevent, or contain harm.
We recommend a framework built around four principles:
●​ Match the level of clinician oversight to the level needed for risk mitigation. Permit
autonomous clinical functions where evidence establishes safety and effectiveness, with
human intervention required where it is necessary to mitigate a specific risk.
●​ Build authorization on clinically relevant evidence. Representative testing using documented
clinical cases and validated simulation can contribute to the evidence supporting
authorization, with further studies guided by remaining safety and effectiveness questions.
Accept findings across care settings wherever their applicability is demonstrated.
●​ Build on established evidence as capabilities expand. Combine continuing clinical assessment
with targeted evaluation of changes, preserving valid findings and providing a predictable
route to broader clinical scope.
●​ Enable advances in foundation models through device-level evaluation. Assess models within
the complete clinical device, using controlled testing to substantiate performance and
safeguards, evaluate multi-step actions, and manage updates.

1. ASSESS CLINICAL RISK AND CLINICIAN OVERSIGHT BY EFFECTIVE RISK
MITIGATION

Assessing risk in the deployed workflow
For devices intended to perform clinical actions with some degree of autonomy, clinician oversight
means a clinician’s ability to monitor and intervene in those actions. We support FDA’s two-axis
risk assessment framework as a starting point for determining evidence requirements tailored to
the device’s intended use and proportionate to its risk. Assessment should consider error
detectability, reversibility, time to harm, and the effectiveness of device safeguards and the
device-user interface. Any reliance on clinician oversight should be supported by evidence that the
interface enables reliable detection and correction of relevant errors under the intended conditions
of use. Evaluation should focus on the device and its interface under intended conditions of use,

Page 1 of 5
including workflow dependencies necessary for safe operation, while keeping manufacturer
responsibilities clearly bounded.
Manufacturers are accountable for the clinical decisions and actions their devices perform,
including during autonomous operation. That accountability requires monitoring clinical
performance, investigating adverse events, and restricting or stopping affected functions when
continued use is no longer justified.
Effective clinician oversight of autonomous device actions requires defined responsibilities,
response times, and clinician authority to restrict or stop affected functions, with arrangements to
protect patients already receiving care. Independent review of sampled cases and targeted testing
can help identify unrecognized failures. Clinicians remain central to setting standards, assessing
difficult cases, investigating incidents, and delivering care when intervention is required.

Evidence for human oversight
For devices intended to operate autonomously under pre-specified workflows, evaluation of the
device-user interface should establish whether clinicians can detect and correct relevant errors
under the intended conditions of use, including the effects of automation bias, habitual approval,
fatigue, and delay. Manufacturers seeking authorization for autonomous operation must
demonstrate that the device and its safeguards adequately mitigate the relevant clinical risks.
Requirements for prior clinician approval of device actions should reflect the additional risk
mitigation that such review provides. Any clinician training or competency requirements
necessary for safe use should be defined and supported by the device’s evaluation.
Human-on-the-loop oversight allows a validated device to deliver care autonomously while
preserving the ability for a qualified clinician to intervene when needed. The risk mitigation
provided by clinician oversight of a device intended to deliver autonomous care depends on
reliable detection, timely escalation, and effective intervention within the time available to prevent
harm. For high-consequence or irreversible actions, demonstrating safety requires evidence that
safeguards operate before the relevant harm can occur, establishing when prior human approval is
necessary and when validated alternative controls provide adequate risk mitigation.
Benchmarking and clinical confirmation can support direct assessment at the proposed degree of
autonomy, with further studies addressing uncertainties that existing evidence cannot resolve.
Supervised evaluation can contribute where relevant, provided reviewers’ interventions are
accounted for when interpreting performance. This preserves the applicable safety standards while
permitting validated alternatives to routine prior approval.

2. DEMONSTRATE CLINICAL COMPETENCE THROUGH REPRESENTATIVE AND APPLICABLE
EVIDENCE

Representative testing
FDA should recognize representative clinical testing as a rigorous basis for demonstrating
competence across the clinical function proposed for marketing authorization. Open-ended
inputs and responses can be assessed through clinically justified sampling across diagnostic groups,
patient populations, management tasks, and risk characteristics, using prespecified care and safety
criteria that allow clinically acceptable responses. An estimate weighted to the intended case mix
can summarize performance alongside separate results for safety-critical tasks, relevant subgroups,
and under- and over-escalation. Targeted challenge cases can assess uncommon but consequential
failures that routine sampling may rarely capture.
For patient-facing devices, evaluation also needs to establish whether patients understand and can
act on the information provided. Testing must reflect the languages, health literacy, and
communication needs of the intended population.
Benchmarking and clinical confirmation address complementary questions about competence
and its applicability to clinical use; FDA should credit evidence across both components wherever
it resolves the relevant question. Benchmarks require demonstrated clinical relevance and a
justified relationship to independently assessed performance. The extent of further confirmation
depends on the uncertainties remaining for the proposed operating conditions.

Page 2 of 5
Qualified independent assessment can strengthen confidence in clinical adjudication and the
validation of evaluation methods. Its contribution depends on the clinical question and
demonstrated quality of the assessment, with manufacturer-generated evidence remaining eligible
for consideration on its merits.

Patient records and validated simulation
Clinical evaluation should be grounded in real clinical data whenever possible, including through
retrospective evaluation of existing patient records. For devices intended to conduct multi-turn
interactions, we recommend recognizing clinically sourced and validated simulations as a valid
approach to retrospective evaluation, with their contribution determined by the clinical questions
they can answer. The simulation protocol needs to demonstrate fidelity to the source clinical data,
clinically plausible scenarios representative of the intended use, and consistency across repeated
execution. Interactive simulation can test how a device gathers information, assesses a case,
communicates, and responds to changing findings.
Independent clinical assessment, repeated execution, and sensitivity analyses can test assumptions
that affect the conclusions drawn from simulation-based validation testing. Evaluation protocols
should establish that clinical cases used directly or to construct simulations are independent of
data used to train or tune the device and its underlying foundation models. Any uncertainty about
prior exposure should be reflected in the interpretation of findings. Findings from real patient
inputs, reconstructed interactions, and wholly synthetic scenarios require separate reporting, with
any combined analysis justified; clinically validated synthetic scenarios can extend rare-hazard
coverage, with tests using related models assessing errors shared by the generator and device.
Where these methods resolve the material premarket questions, we recommend allowing their
findings to support marketing authorization alongside structured continuing assessment.
Uncertainty about patient behavior, execution, integration, or handoff determines the need for
additional verification, usability work, or prospective investigation, with FDA specifying the
question that further evidence must resolve.

Transferability across care settings
FDA should assess evidence by the validity of its methods, its clinical relevance, and its
applicability to the proposed intended use. Clinical confirmation of device performance across
care settings requires distinguishing transferable clinical competence from adherence to applicable
care standards. Findings from multiple settings merit inclusion where manufacturers demonstrate
comparability of patients, presentations, and care pathways.
We recommend that clinical confirmation in a new care setting build on existing evidence of
device performance where that evidence remains applicable to the proposed intended use.
Targeted confirmation should address adherence to the clinical practice guidelines and care
pathways applicable to that use. Where a device claims to apply supplied clinical practice
guidelines, evidence should demonstrate that its decisions follow those guidelines while preserving
risk mitigations shown to be effective in prior clinical evaluation. Additional evaluation should
address differences in population, prevalence, or conditions of use that could change device
performance, while retaining applicable prior findings.

3. BUILD CONTINUING EVIDENCE AND A PREDICTABLE ROUTE TO CONTROLLED
EXPANSION

Monitoring during clinical use
We propose an evidence pathway connecting sufficient premarket evidence to structured
prospective postmarket assessment, with proportionate premarket review of added capabilities
where further authorization is required. The discussion paper explores postmarket evidence and
the use of initial competency assessments as a baseline for changes. The premarket submission
should identify the clinical questions to be investigated during postmarket use, the evidence
needed to answer them, and the actions required when findings affect continued safety and
effectiveness.
The proposed prospective assessment covers clinical quality, adverse outcomes, escalation and
completed handoffs, and performance across the intended patient population. The timing and
Page 3 of 5
extent of this evidence collection should reflect the device’s intended use and the safety and
effectiveness questions remaining after premarket evaluation. Reliance on monitoring requires
operational arrangements when clinical use begins, with measurable encounter capture, follow-up
completeness, and linkage between deployed configurations, actions, and outcomes. FDA should
accept auditable data arrangements that meet justified requirements for data quality, independent
assessment, and timely response, preserving flexibility in implementation.
The extent to which postmarket monitoring can address remaining uncertainty depends on
demonstrated detection of clinically important failures and timely action. Postmarket monitoring
plans need prespecified review frequency, alert criteria, investigation procedures, and device
manufacturer authority to restrict or stop affected functions, with detection delays and gaps in
coverage made explicit. Independent review and targeted testing can examine unrecognized errors;
reliance on supervisory agents requires validation of detection performance and shared failure
modes. The extent of postmarket reliance depends on these capabilities and remaining
uncertainty, with sufficient premarket demonstration of safety.

Controlled updates and scope expansion
Controlled updates require manufacturers to identify the deployed model and device
configuration, detect safety-relevant changes to dependencies, and apply prespecified testing and
acceptance criteria before implementation. Validated fallback, restriction, or rollback
arrangements provide a response when continued use is no longer justified.
Manufacturer-controlled verification complements supplier disclosures, including detection of
unexpected behavioral changes.
FDA should also establish a predictable route for extending an authorized clinical function.
Demonstrating each added capability requires targeted evaluation of its performance, risks, and
interactions with existing functions, drawing on the established evidence package and findings
from clinical use. We recommend scaling additional testing and regulatory review to the change in
intended use and risk, retaining findings about unaffected capabilities where their continuing
validity is justified.
A predictable pathway requires a clear distinction between changes within an existing
authorization and extensions requiring further authorization. Predetermined change control plans
can support appropriate, prospectively specified modifications; FDA’s August 2025 PCCP
guidance requires these to maintain intended use and generally recommends maintaining
indications for use. For broader clinical scope, FDA should clarify the additional evidence and
submission requirements before implementation. Continued clinical assessment and
proportionate review would allow capability to expand as the evidence supports it.

4. EVALUATE FOUNDATION MODELS AND AGENTIC CLINICAL FUNCTIONS THROUGH
DEVICE PERFORMANCE

Foundation-model performance
Access to the most capable foundation models is important to developing devices that can
undertake demanding clinical work. We support a regulatory framework that enables
manufacturers to incorporate advances in model capability when evidence demonstrates safe and
effective performance in the complete device.
FDA should assess a model’s contribution through the device’s clinical performance with its
guidance, tools, and safeguards in place. Controlled behavioral testing can examine how decisions
and actions change when inputs or components vary, providing useful evidence even where a
model’s internal operation is not fully interpretable. Testing requirements depend on the claim
being made and the clinical consequences of failure.
Where a manufacturer claims that several models can be used interchangeably, substantiating that
claim requires testing specified substitutions under the same evaluation conditions. Prespecified
equivalence or noninferiority criteria need to address clinical decisions, safety endpoints,
escalation, and subgroup performance, with sufficient sensitivity to identify clinically important

Page 4 of 5
differences. The claim supported by these findings is limited to the models, versions, functions,
and configurations evaluated.
Controlled changes to inputs can also test robustness to irrelevant cues, communication
differences, misleading prompts, and failures in retrieved information or tools. For example, a
nonclinical cue can be varied while pertinent clinical facts remain constant to determine whether
management changes without clinical justification. Tests of clinically relevant attributes such as
age or sex assess whether differences in care are justified, with subgroup analyses examining
performance across the intended population.
Voluntary Foundation Model Device Master Files could provide reusable component information
under appropriate confidentiality, with alternative documentation routes where needed. This
information can strengthen the device assessment alongside manufacturer verification and
supplier controls. The manufacturer retains responsibility for the safety and effectiveness of the
complete device.

Safety across multi-step actions
For agentic clinical functions, assessment needs to follow the sequence of decisions and actions
through which the device affects care. Relevant failure modes include conflicting inputs, stale
information, erroneous tool outputs, and errors that compound over successive actions.
Evaluation of safeguards, escalation, and intervention needs to establish their effectiveness within
the time available to prevent harm, including when failures occur across several steps. Interpreting
the results requires reporting clinical coverage, uncertainty in performance estimates, assumptions
about the operating environment, and controls for remaining risks.

CONCLUSION

FDA has an opportunity to establish a framework that enables generative AI-enabled medical
devices to deliver meaningful additional clinical capacity. Its success will depend on a clear
connection between demonstrated competence and the device’s authorized intended use,
including its degree of autonomy. Representative testing, validated evidence methods, and
effective safeguards can provide that connection, with further investigation directed to
uncertainties that materially affect safety and effectiveness.
The principal risk is that a competency-based framework becomes a succession of requirements
that add burden without resolving those uncertainties. Autonomy alone does not establish which
clinician oversight arrangement or study design provides the necessary assurance; nor does a
change in setting or model establish that prior findings have lost their value. Each requirement
needs a clear clinical purpose and a defined contribution to the evidence supporting use.
We therefore recommend a pathway that permits autonomous clinical functions when justified by
evidence, accepts clinically applicable findings across settings, and supports controlled
improvements and broader scope through targeted evaluation. Continuing assessment and
manufacturer accountability must sustain that assurance during use. The measure of success is a
framework through which demonstrated advances become safe, effective care that reaches
patients.

Page 5 of 5