Mauro Clinical Intelligence

Mauro Clinical Intelligence (MCI) appreciates the opportunity to comment on FDA’s discussion paper on generative AI-enabled medical devices. MCI is a nursing-focused clinical intelligence initiative concerned with safe AI use, clinical judgment, workflow, patient safety, and meaningful human oversight.

MCI does not claim to speak for nursing as a profession. We offer a nursing-centered perspective because nurses are likely to be among the clinicians who review, interpret, act on, correct, and escalate AI-generated outputs in real-world care.

First, “human in the loop” should mean more than a clinician being present somewhere in the workflow. Meaningful oversight requires enough time, information, authority, and workflow access to independently evaluate AI output, stop or override an unsafe recommendation, and escalate concerns. If safety depends on a nurse catching an error, FDA should consider whether that review is realistic under actual conditions of workload, interruptions, competing priorities, and alert burden.

Second, competency testing should reflect the profession and workflow in which the device will be used. If an AI system influences nursing care, evaluation should include nursing-specific scenarios such as deterioration recognition, handoffs, medication verification, patient education, incomplete or conflicting information, workflow interruptions, and situations where a technically plausible answer does not fit the patient context. Nurses should participate as expert evaluators when nursing practice is affected.

Third, clinician-AI disagreement should be treated as valuable post market safety information. Overrides, corrections, rejected recommendations, missed or inappropriate escalations, recurring workarounds, near misses, and repeated instances of missing context may reveal safety problems that were not visible during premarket testing. These events should not automatically be treated as user noncompliance. A nurse repeatedly correcting the same AI behavior may be identifying a recurring defect.

Fourth, care escalation deserves specific evaluation. In many settings, a nurse is the clinician who receives an escalation, reassesses the patient, gathers additional information, and determines what happens next. FDA should consider whether AI escalation is timely, actionable, appropriately prioritized, and compatible with real workflow.

Fifth, agentic AI requires clear boundaries for when autonomous action must stop. Manufacturers should define what the system may do independently, what requires clinician confirmation, what conditions terminate autonomous action, what information is handed to the clinician, and how the clinician can stop, reverse, or modify the process. The more steps a system can execute independently, the more important it is to evaluate cumulative risk across the full sequence.

Finally, FDA should account for smaller and resource-limited settings, including independent NP practices, rural clinics, home health, hospice, ambulatory care, and nurse-led businesses. These settings may not have dedicated AI governance committees, informatics teams, or continuous technical monitoring. A safeguard that works in a large academic health system may be unrealistic in a small practice.

MCI recommends that FDA:

1. Evaluate whether human oversight is operationally realistic.
2. Use nursing-specific workflows and expertise when nursing practice is affected.
3. Treat overrides, corrections, and clinician-AI disagreement as potential post market safety signals.
4. Evaluate AI-to-clinician handoffs during escalation.
5. Require clear human-intervention boundaries for agentic systems.
6. Account for smaller practices that lack enterprise-level AI governance infrastructure.

As generative AI becomes more capable, the goal should not be to preserve a human somewhere in the workflow merely for appearance. The goal should be to preserve a clinician who has the information, authority, time, and practical ability to recognize when the technology is wrong and act accordingly.

For many patients, that clinician will be a nurse.

Mauro Clinical Intelligence

Mauro Clinical Intelligence

COMMENTS OF MAURO CLINICAL INTELLIGENCE
Docket No. FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback
________________________________________________________________________________

Submitted by Mauro Clinical Intelligence (MCI)

Date October 6, 2026

Principal focus Nursing judgment, patient safety, human oversight, workflow, and postmarket learning

Statement of perspective. Mauro Clinical Intelligence is a nursing-focused clinical intelligence initiative.
MCI does not purport to speak for the nursing profession, any employer, or any professional association.
These comments are offered from a nursing-centered perspective because registered nurses and
advanced practice registered nurses are likely to be among the clinicians who receive, interpret, verify,
act on, correct, override, and escalate outputs from generative and agentic AI-enabled medical devices.

I. Executive Summary
MCI supports CDRH's total product life cycle approach and the discussion paper's effort to develop riskproportionate methods for GenAI-enabled medical devices. The paper appropriately recognizes that
open-ended outputs, evolving models, care-escalation functions, foundation models, and agentic
systems create regulatory questions that are not fully addressed by conventional software evaluation.
MCI respectfully recommends that CDRH make the clinical feasibility of human oversight a more explicit
part of risk assessment, premarket evaluation, and postmarket monitoring. For nursing-facing or
nursing-dependent systems, the relevant question is not merely whether a human remains somewhere
in the loop. The relevant question is whether the clinician has the time, information, authority, workflow
access, and organizational support necessary to recognize an unsafe output and act on that recognition.
1. Add human-oversight feasibility and workflow dependence as explicit risk modifiers under Section IV
and Question 1.
2. Evaluate care escalation as an AI-to-clinician handoff, including under-escalation, over-escalation,
timeliness, context, and the receiving clinician's ability to act (Section IV, Question 6; Appendix A,
S.1).
3. Require nursing-relevant benchmarks, scenarios, and qualified nursing adjudicators when a device
is intended for nursing use or materially affects nursing workflow (Section V.B; Questions 7 and 9;
Appendix A, E.1-E.4).
4. Ensure clinical confirmation represents real nursing workflow conditions, including interruptions,
competing priorities, incomplete information, and deployment context (Section V.C; Question 11).
5. Avoid defaulting to physician comparators when the intended user or affected clinical workflow is
nursing; comparator expertise should match the profession, scope, and setting (Section IV, Question
4; Section V.D; Question 14).
6. Treat clinician-AI disagreement, overrides, corrections, workarounds, near misses, and missed
escalations as potential postmarket safety signals, while preserving manufacturer accountability
(Section VI; Questions 19 and 21).

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 1
 Mauro Clinical Intelligence

7. For agentic systems, require explicit human-intervention boundaries, stop conditions, handoff
requirements, reversibility expectations, and auditability before high-consequence actions (Section
VII.B; Question 26; Appendix A, A.1).
8. Account for independent practices, home health, hospice, rural clinics, and other resource-limited
settings that may not have enterprise AI governance infrastructure.

II. Introduction and Regulatory Context
FDA's August 2026 discussion paper proposes a possible two-axis risk framework, a competency-based
premarket evaluation approach, risk-proportionate postmarket monitoring, and additional considerations
for foundation models and agentic AI systems. MCI agrees that these areas are central to safe adoption.
The discussion paper also correctly notes that GenAI-enabled devices may produce variable outputs,
may evolve after deployment, and may be difficult to evaluate exhaustively before use.
Nursing warrants explicit consideration within this framework for two reasons. First, nurses are often the
clinicians continuously present where changes in patient condition become visible. Second, many AI
functions do not terminate at an output. They enter a workflow. A recommendation may create an alert,
alter documentation, trigger a message, escalate care, prioritize a task, or change what information the
clinician sees first. The safety of the device can therefore depend on the interaction between the AI
output and the receiving clinician's workload, authority, judgment, and environment.
FDA reference: Discussion Paper Sections II-IV, especially the total product life cycle discussion and Section
IV's two-axis risk framework (pp. 1-10).

III. Response to Section IV: Risk Assessment Should Include the Feasibility of Human
Oversight
Relevant FDA provisions: Section IV; Question 1; Question 4; Question 6.
Question 1 asks whether dimensions beyond device activity and the consequence of an incorrect output
should be represented in the risk framework, including time pressure, downstream safeguards,
reversibility, and traceability. MCI recommends that CDRH expressly recognize human-oversight
feasibility as a risk modifier. A device may appear lower risk because a clinician is expected to
independently review its output, yet that safeguard is only credible if independent review is operationally
realistic.
For example, a nurse receiving an AI-generated deterioration alert may be simultaneously responsible
for medication administration, direct assessment, patient education, documentation, communication with
physicians, and several other patients. The presence of a nurse in the workflow should not automatically
be treated as equivalent to continuous supervision. CDRH should consider whether the expected
reviewer has sufficient time, source information, clinical authority, and an accessible mechanism to
pause, reject, or escalate the output.
This consideration is consistent with FDA's own framing. Section IV distinguishes functions operating
with continuous healthcare-professional supervision from fully autonomous functions, and Question 1
specifically identifies time pressure and downstream safeguards as possible risk dimensions. MCI
recommends that future policy make clear that the quality of those safeguards matters. A safeguard that
exists nominally but cannot be used reliably under real workflow conditions should not justify a lower risk
classification.
MCI also encourages CDRH to broaden the professional framing in Question 4. The question asks
about generalist and specialist physicians, but the same principle applies to other licensed clinicians.
Where a device is designed for nursing practice, or where safe use depends on nursing assessment, the
relevant issue is whether the intended user has the professional knowledge and contextual information

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 2
 Mauro Clinical Intelligence

necessary to independently evaluate the output. Future regulatory language should refer to
appropriately qualified healthcare professionals rather than defaulting to physician-only comparators.

IV. Response to Section IV, Question 6 and Appendix A, S.1: Care Escalation Is a
Handoff
Relevant FDA provisions: Section IV, Question 6; Appendix A, S.1 Safety-critical recognition and escalation.
MCI agrees with CDRH that both under-escalation and over-escalation can cause harm. Underescalation can delay necessary care; over-escalation can create unnecessary utilization, anxiety,
testing, alarm burden, and erosion of trust. For nursing practice, however, escalation should also be
evaluated as a handoff between a device and a clinician.
The safety of an escalation function depends not only on whether the device decides to escalate, but on
what happens next. The receiving nurse or other clinician must understand the reason for escalation, its
urgency, the supporting information, and any uncertainty or limitations. The alert must arrive through a
channel that is visible and actionable in the actual workflow. An escalation that is technically correct but
poorly timed, insufficiently contextualized, or buried among low-value alerts may not function as an
effective safety mechanism.
MCI recommends that evaluation of escalation functions include: time to escalation; sensitivity and
specificity by clinical context; the quality of handoff information; whether uncertainty is communicated;
whether the receiving clinician can obtain the underlying basis for the recommendation; whether the
escalation can be acknowledged, challenged, or re-prioritized; and whether recurrent false or missed
escalations can be reported and reviewed.
Appendix A, element S.1 already identifies timeliness, clarity, actionability, empathy, and handoff quality
as relevant. MCI encourages CDRH to preserve these elements and explicitly test them in nursing
workflows where nurses are likely to receive or operationalize escalation.

V. Response to Section V: Competency-Based Evaluation Must Match the Intended
Clinical Profession and Workflow
Relevant FDA provisions: Section V.B; Questions 7 and 9; Appendix A, E.1-E.4, R.1-R.2.
MCI supports the general concept of competency-based evaluation using device benchmarking followed
by clinical confirmation. The value of that approach will depend heavily on whether the benchmark
measures the competencies required in the intended environment rather than only the competencies
that are easiest to score.
For nursing-facing devices, benchmark design should include nursing-specific workflows and failure
modes. Examples include recognition of subtle deterioration over time; integration of conflicting or
incomplete information; medication verification; prioritization under competing demands; handoffs;
patient education; assessment of functional or behavioral change; communication with patients and
families; documentation under interruption; and escalation when the patient's presentation does not fit a
structured pathway.
Appendix A is particularly relevant. Element E.2 addresses information gathering and clinical analysis
under incomplete, evolving, or ambiguous information. Element E.4 addresses communication quality,
user comprehension, and automation bias. These competencies should be tested with realistic nursing
users and nursing-relevant cases when nurses are intended users or downstream safety reviewers.
MCI further recommends that qualified nursing experts serve as adjudicators when the device materially
affects nursing practice. Section V.B already contemplates expert adjudicators with qualifications aligned
to the clinical domain and complexity of the outputs. That principle should be applied by profession as
well as by specialty. A nurse-facing device should not be judged solely by physician experts if the

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 3
 Mauro Clinical Intelligence

relevant question is whether the output supports safe nursing assessment, documentation, escalation,
or care coordination.

VI. Response to Section V.C and Question 11: Clinical Confirmation Should Reproduce
Real Nursing Conditions
Relevant FDA provisions: Section V.C Clinical Confirmation; Question 11.
CDRH correctly notes that structured benchmarking may not capture how a GenAI-enabled device
interacts with users, workflows, and patient populations. MCI strongly supports this distinction. Clinical
confirmation should be designed to reveal failures that occur only when the device is embedded in
actual care delivery.
For nursing-facing systems, clinically representative conditions should include realistic interruptions,
varying staffing conditions, simultaneous responsibilities, delayed or missing data, handoffs, shift
transitions, device-to-EHR latency, and the possibility that the nurse appropriately prioritizes direct
patient care over immediate interaction with the software.
MCI also recommends that clinical confirmation examine the work created by the device, not only the
task the device is intended to reduce. Correction burden, duplicate documentation, exception handling,
alert management, and time required to verify AI output may offset apparent efficiency gains. A device
that decreases first-pass documentation time but creates substantial verification and correction work
may not improve the real workflow.
A 2026 randomized study of home-visit nurses illustrates a related point: access to a guideline-informed
AI tool was associated with lower perceived mental effort and higher confidence, but the study did not
evaluate reasoning quality, accuracy, or safety. This distinction supports evaluating subjective usability
and workload benefits separately from clinical performance and safe judgment.

VII. Response to Section V.D and Question 14: Comparator Selection Should Not
Default to Physicians
Relevant FDA provisions: Section V.D.1 Standards of Performance; Question 14; also Question 15.
Question 14 asks how qualified clinician panels should be used as comparators and specifically asks
whether generalist or specialist physicians should serve as the performance standard. MCI recommends
that future policy use a broader formulation: the comparator should be the appropriately qualified
clinician or clinical team for the device's intended use.
For devices that support or influence nursing practice, qualified registered nurses or advanced practice
registered nurses may be the correct comparator. This is not a matter of professional representation for
its own sake. It is necessary for construct validity. The correct comparator is the clinician whose actual
task, scope, contextual information, and decision responsibilities the device is intended to support or
replace.
MCI also supports evaluating human-AI team performance where that reflects the intended use. In those
studies, acceptance criteria should include whether the technology changes clinician behavior, creates
automation bias, reduces independent review, or increases unwarranted confidence. A device may
improve average task performance while making rare but consequential errors more difficult for
clinicians to detect.

VIII. Response to Section VI, Questions 19 and 21: Clinician-AI Disagreement Is
Postmarket Safety Data
Relevant FDA provisions: Section VI.A-B; Question 19; Question 21.

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 4
 Mauro Clinical Intelligence

MCI strongly supports risk-proportionate postmarket monitoring. GenAI-enabled devices can change
after deployment, encounter inputs not represented in premarket testing, and interact with workflows in
ways that generate new failure modes. Postmarket surveillance should therefore include data that arise
when clinicians disagree with the device.
Potential safety signals should include, as applicable:
• clinician overrides or rejected recommendations;
• corrections to AI-generated documentation or structured data;
• missed, delayed, or inappropriate escalations;
• recurrent workarounds or suppression of alerts;
• near misses associated with AI output;
• situations in which clinicians repeatedly identify missing context;
• unexpected shifts in override or acceptance rates after model or workflow changes; and
• patterns in which a particular subgroup, setting, or workflow requires disproportionate correction.
These events should not automatically be categorized as user error or noncompliance. Repeated
clinician correction may be evidence of a device performance problem, an interface problem, a
mismatch between intended use and deployment, or a change in model behavior.
Question 21 appropriately asks what role clinicians and healthcare institutions should play without
diffusing manufacturer accountability. MCI recommends a clear division: manufacturers should remain
responsible for device performance and required monitoring, while healthcare organizations and
frontline clinicians should be treated as important sources of real-world safety signals.
The reporting pathway itself is a safety dependency. A 2026 national survey of 1,189 practicing U.S.
nurses found that among 424 nurses who had encountered an AI-related patient-safety concern, 19.6%
reported having no formal channel for reporting it; nurses outside hospitals had higher odds of lacking a
channel. This finding is particularly relevant to CDRH's shared-ecosystem discussion. Postmarket
surveillance cannot rely on frontline reporting if the infrastructure to report is absent or difficult to use.

IX. Response to Section VII.B and Question 26: Agentic AI Requires Explicit HumanIntervention Boundaries
Relevant FDA provisions: Section VII.B Considerations for Agentic AI Systems; Question 26; Appendix A, A.1.
Agentic AI creates a qualitatively different safety problem because the system may plan and execute
multiple steps before a clinician has an opportunity to review the sequence. MCI agrees with FDA that
reduced opportunity for human review should affect acceptance criteria and oversight.
For agentic GenAI-enabled devices, sponsors should clearly define:
• the actions the system may perform autonomously;
• the actions that require human confirmation;
• the conditions that immediately stop autonomous execution;
• the point at which control must transfer to a licensed clinician;
• the information and rationale provided at the handoff;
• whether completed actions can be reversed or mitigated;
• how tool failures or conflicting tool outputs are handled;
• how missed human-oversight checkpoints are detected; and
• what audit trail is available for reconstructing the sequence after a safety event.

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 5
 Mauro Clinical Intelligence

Appendix A, element A.1 appropriately identifies compliance with human-oversight checkpoints before
irreversible or high-consequence actions. MCI recommends that CDRH also evaluate whether those
checkpoints are realistically usable by the intended clinician in the deployment setting. A checkpoint that
appears on screen but occurs after several automated actions, arrives during a high-workload period, or
lacks sufficient context may provide less protection than its design suggests.
Nurses should participate in defining these boundaries when nursing surveillance, assessment,
communication, or intervention is part of the downstream safety system. In many settings, the nurse
may be the first clinician to recognize that an autonomous sequence no longer fits the patient's
condition.

X. Cross-Cutting Recommendation: Account for Small and Resource-Limited Care
Settings
The regulatory framework should be usable outside large academic health systems. GenAI-enabled
devices may be deployed in independent nurse practitioner practices, rural clinics, home health,
hospice, ambulatory care, community settings, and nurse-led businesses. These environments may not
have dedicated AI governance committees, informatics departments, data-science teams, or 24-hour
technical monitoring.
Accordingly, CDRH should consider whether a device's safe use depends on institutional capabilities
that are not universally available. Labeling, risk controls, monitoring expectations, and escalation
mechanisms should make those dependencies explicit. If safe deployment requires enterprise-level
infrastructure, that should be transparent to smaller organizations before adoption.
This is also relevant to Question 21. Professional societies, boards of nursing, specialty organizations,
and other bodies may help disseminate practice guidance and reporting expectations, but those
ecosystem roles should supplement rather than replace manufacturer responsibility.

XI. Conclusion
MCI appreciates CDRH's early engagement with clinicians and other stakeholders. The discussion
paper identifies many of the right regulatory questions. MCI's central recommendation is that human
oversight be evaluated as a clinical capability, not a label attached to a workflow.
For nursing-facing GenAI-enabled devices, reasonable assurance of safety and effectiveness may
depend on whether the nurse can understand the output, identify when it does not fit the patient,
challenge or stop the system, report recurrent problems, and trigger organizational or manufacturer
learning.
As GenAI becomes more capable, the goal should not be merely to preserve a human somewhere in
the workflow. The goal should be to preserve a clinician who has the information, authority, time, and
practical ability to recognize when the technology is wrong and act accordingly.
For many patients, that clinician will be a nurse.

Appendix: Crosswalk to Selected FDA Discussion Questions
FDA section / question MCI recommendation Primary nursing concern

Section IV, Q1 Add human-oversight feasibility and Time pressure, workload, authority,
workflow dependence as explicit risk independent review.
modifiers.

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 6
 Mauro Clinical Intelligence

Section IV, Q4 Use profession-appropriate HCP Nursing scope and competence as
framing, not physician-only intended-user context.
assumptions.

Section IV, Q6 / App. A S.1 Evaluate escalation as an AI-to-clinician Under/over-escalation, context,
handoff. prioritization, actionability.

Section V.B, Q7 & Q9 Use nursing-specific benchmarks and Real nursing tasks, ambiguous
qualified nursing adjudicators when information, workflow interruptions.
nursing practice is affected.

Section V.C, Q11 Confirm performance under clinically Workload, interruptions, correction
representative nursing conditions. burden, deployment context.

Section V.D, Q14 Select comparators by intended Construct validity; nurse or APRN
profession, task, scope, and setting. comparators where appropriate.

Section VI, Q19 & Q21 Capture clinician-AI disagreement and Overrides, corrections, workarounds,
accessible frontline reporting as near misses, reporting infrastructure.
postmarket safety signals.

Section VII.B, Q26 / App. A A.1 Define stop points, handoff criteria, Reduced opportunity for human review;
reversibility, and audit trails for agentic high-consequence autonomous action.
systems.

References
1. 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. https://www.fda.gov/media/194242/download
2. U.S. Food and Drug Administration. FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AIEnabled Medical Devices. August 18, 2026. https://www.fda.gov/news-events/press-announcements/fdaseeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices
3. American Nurses Association. American Nurses Association Publishes Fifth Edition of Nursing: Scope and Standards
of Practice. September 22, 2026. https://anaprodsite2.nursingworld.org/news/news-releases/2026-newsreleases/american-nurses-association-publishes-fifth-edition-of-nursing-scope-and-standards-ofpractice/
4. Carter GA. A National Cross-Sectional Survey of Artificial Intelligence Patient Safety Reporting Infrastructure Available
to US Nurses. Journal of Nursing Care Quality. Published online September 10, 2026.
doi:10.1097/NCQ.0000000000001011. https://pubmed.ncbi.nlm.nih.gov/42717301/
5. Ishikawa T, Imura H, Kitazawa T, et al. Guideline-informed AI chat access and perceived mental effort in a simulated
home-visit nursing task: a randomized controlled trial. BMC Nursing. Published September 25, 2026.
doi:10.1186/s12912-026-05418-w. https://doi.org/10.1186/s12912-026-05418-w
6. National Council of State Boards of Nursing. NCSBN and Leading Nurse Scientists to Launch Survey Exploring How AI
is Affecting Nursing Practice. September 29, 2026. https://www.ncsbn.org/news/ncsbn-and-leading-nursescientists-to-launch-survey-exploring-how-ai-is-affecting-nursing-practice

Mauro Clinical Intelligence | FDA-2026-N-7874 | Page 7