Anchor Point Health Foundation (Lois Sonstegard)

See attached file(s)

COMMENT TO THE FOOD AND DRUG ADMINISTRATION
Docket No. FDA-2026-N-7874
Generative Artificial Intelligence in Medical Device Evaluation

Submitted by:
Lois Sonstegard, PhD, MBA, MSN, PCC, RN
Founder, Anchor Point Health Foundation
Faculty, St. Catherine University
October 19, 2026

INTRODUCTION
I submit this comment as a health system leader, nursing educator, and researcher whose work
centers on the governance architecture of AI-assisted care — specifically the accountability gap
that emerges when AI systems operate between clinical encounters. I write to address a
structural question that I believe is foundational to the FDA's framework for evaluating
generative AI in medical devices: not whether AI performs adequately within a clinical
encounter, but who holds adjudicating authority over AI-generated recommendations and
actions that occur outside it, and what gives that authority legitimate standing.

This is not a question the current landscape answers well. The events of this week — federal
officials publicly warning against a private multi-stakeholder consortium as a potential
"regulatory cartel," and a major physician professional body asserting that health AI governance
is properly physician-led — illustrate the problem precisely. Both claims invoke expertise and
good intentions. Neither establishes what makes an adjudicating body legitimate rather than
merely powerful.

THE ADJUDICATION PROBLEM
Generative AI in health care does not operate only at the point of care. It operates in the
intervals between encounters — generating recommendations for remote monitoring, producing
care coordination communications, identifying patient populations for outreach, flagging
deterioration signals — and it does so without a clinician present to assess, override, or take
responsibility for its output in real time.

The FDA's existing device evaluation framework is well designed for what happens inside the
encounter: a device performs a function, a clinician interprets it, and accountability attaches to
both. Generative AI operating between encounters does not fit this model. The "device"
generates output continuously. The clinician reviews it intermittently, if at all. The patient
experiences the consequence in the interval.

What is missing is an adjudicating function: a structured mechanism that can verify whether AIgenerated recommendations between visits were appropriate, acted upon, and accountable to
someone. This is distinct from pre-market evaluation. Pre-market evaluation asks whether the
device performs as intended under test conditions. Adjudication asks whether it performed
appropriately in a specific patient's care — after the fact, with clinical and contextual knowledge
that no pre-market evaluation can anticipate.

A recent systematic review in Frontiers in Digital Health (October 2026) documents this gap at
the structural level, finding U.S. AI requirements dispersed across FDA, ONC, and CMS
authorities with enforcement divided among federal and state actors holding partial jurisdiction,
and post-market mechanisms still early — in contrast to the EU's binding pre-market conformity
assessment with centralized market-surveillance enforcement. The fragmentation is not
incidental. It reflects the absence of a shared adjudicating architecture.

WHY EXISTING GOVERNANCE MODELS DO NOT RESOLVE THIS
Two governance models are currently competing for authority over health AI. Neither addresses
the adjudication problem adequately.

The consortium model assembles stakeholders — device manufacturers, health systems,
payers, clinicians — and asks them to develop shared standards. This produces valuable
technical guidance. It does not produce adjudicating authority. A consortium's standards are
advisory unless adopted by a body with enforcement power. More fundamentally, a consortium
whose members have financial interests in the technology being governed faces a structural
conflict that limits its legitimacy as an adjudicator. The recent federal concern about a
"regulatory cartel" is a symptom of this: a consortium of the willing, once it accumulates real
influence, begins to look like a private regulatory body without public accountability.

The professional-guild model asserts that clinical accountability should remain with licensed
clinicians and their professional bodies. This is defensible for decisions made by a clinician at
the point of care. It does not address AI-generated outputs between encounters, which occur
without a clinician present. Asserting physician authority over AI governance does not create the
mechanism by which that authority is exercised when the AI acts autonomously between visits.
It names a responsible party without specifying how accountability is operationalized.

Both models answer the question "who should be in charge?" without answering the question
"what does being in charge actually require when AI operates between clinical encounters?"

WHAT STRUCTURAL LEGITIMACY REQUIRES
I respectfully suggest that the FDA's framework for generative AI evaluation consider the
following structural requirements for any body claiming adjudicating authority over betweenencounter AI:

Multidisciplinary composition. The clinical encounter involves a physician. The between-visit
interval involves nurses, care coordinators, community health workers, family caregivers, and
the patient. Any body adjudicating AI performance between visits must include the full scope of
disciplines present in that interval, not only the disciplines present at the encounter.
Structural independence from the technology being evaluated. A body whose members
have financial interests in AI systems cannot serve as the primary adjudicator of those systems'
performance. Independence is not a procedural nicety — it is the condition under which an
adjudicating body's findings carry authority.

Defined jurisdiction and statutory grounding. Advisory bodies can be ignored. Adjudicating
bodies with defined jurisdiction and statutory grounding cannot. The distinction matters
enormously for patient safety: a finding that AI performed inappropriately between visits must
trigger an accountable response, not a voluntary corrective action.

Post-market, case-specific review capacity. Pre-market evaluation cannot anticipate the full
range of contexts in which a generative AI system will operate. The adjudicating function must
include the capacity to review specific patient cases — not only aggregate performance data —
and to reach findings about what happened in a particular between-visit interval.

A NOTE ON REGULATORY CONVERGENCE
The FDA is not alone in facing this problem. Three independent forces are converging on the
between-visit interval simultaneously: accreditation bodies introducing governance expectations
for the first time (AAAHC Version 45, effective December 15, 2026); state legislatures enacting
accountability requirements for AI-assisted care beyond the clinical encounter (Alabama,
Colorado, California among others); and federal billing codes (AMA CPT 2027) attaching
payment value to between-visit monitoring and coordination outcomes.

These forces did not emerge from a coordinated policy process. They converged independently
on the same gap. The result is a patchwork of requirements — accreditation, statutory, and
financial — with no unified adjudicating architecture to satisfy them. The FDA has an opportunity
to provide structural guidance that the other authorities cannot: a framework for what legitimate
adjudication of between-encounter AI looks like, grounded in the agency's authority over device
performance and post-market safety.

CONCLUSION
The question before the FDA is not only whether generative AI performs adequately. It is who
determines whether it performed appropriately in a specific patient's care, and what structural
conditions make that determination trustworthy. The current landscape — competing consortia,
professional guilds, and fragmented statutory authorities — does not answer this question. I
urge the FDA to use this rulemaking to articulate what adjudicating authority over betweenencounter AI requires, and to establish the structural conditions under which that authority is
legitimate.

I am grateful for the opportunity to submit this comment and welcome further engagement on
these questions.
Respectfully submitted,

Lois Sonstegard, PhD, MBA, MSN, PCC, RN
Founder, Anchor Point Health Foundation
Faculty, St. Catherine University
lois.sonstegard@gmail.com

Comment to FDA Docket No. FDA-2026-N-7874 · Lois Sonstegard · October 19, 2026