Redesign Health (Preston Tollinger)

Please see attached comments from Redesign Health.

October 9, 2026

Dockets Management Staff (HFA-305) Food and Drug Administration 5630 Fishers
Lane, Room 1061 Rockville, MD 20852

Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative
AI-Enabled Medical Devices
To the Staff:

In a market in which there is significant capital and no shortage of talent focused on the
same problems, what decides where the capital actually goes is rarely the quality of the
science. The determinant of how capital is invested is whether or not anyone can price
the path from a good idea to a paying customer. Healthcare AI has reached that point,
yet for products that fall under FDA’s device authority, the path is not yet priceable.

Redesign Health is a global venture and applied technology platform focused on
building next-generation healthcare companies, and we invest at the early stage in
healthcare technology, partnering with founders and with existing businesses through
AI-native venture capital and buyout strategies. Since 2018, our firm and its portfolio
companies have raised over $1.5 billion and touched the lives of more than 15 million
patients. We believe AI will do more good for humanity in healthcare than anywhere
else, and we are deploying capital on that conviction. Our companies ship
GenAI-enabled health products today, almost all of them in the space that sits outside
device oversight or right at its edge, and this paper will shape whether the next
generation of them moves into regulated territory.

What we can put in front of CDRH is our own record, and the record is one of walking
away. We have consistently found premarket regulatory uncertainty difficult to
underwrite at the seed stage, and we have avoided backing companies that require
FDA authorization before they can reach revenue. The science in front of us was often
promising and the teams were often excellent, but what we could not do was price the
path. When a founder cannot tell an investor which evidentiary bar will apply, what
clearing it will cost, or how long it will take, the responsible answer for a fiduciary is to
pass. Accordingly, we have declined on multiple occasions to fund the development of
products with the potential to help patients.

That is a capital allocation problem, and it is one this paper could largely correct. The
binding constraint on early-stage healthcare AI is the variance of the evidentiary bar,
and height matters far less than predictability, because you can raise against a bar that
costs $2 million and takes 14 months, but you cannot raise against one that might cost
$2 million or might cost $20 million depending on how a review division interprets the
directiveness continuum. Every recommendation below aims to compress that variance.
Give a founder a bar they can price, and capital follows into a category it currently
avoids.

Response to Question 2

FDA-2026-N-7874 · Redesign Health · Page 1 of 8
Directiveness does run along a continuum, and CDRH is right to consider whether a
“talk to your doctor” line leaves an output just as directive as it was before. It does.
Founders have built whole compliance strategies on that disclaimer, and closing it
eliminates a shortcut that deserved closing.

The continuum, however, must be navigable before a company writes any code. A
founder makes the architecture decision that determines device status 18 months
before they file anything, and often before they have hired anyone with regulatory
experience. A worked-example library and a fast classification determination priced for
small sponsors would together make the continuum navigable at that moment.

Publishing a worked-example library would do most of the work. The four-rung lisinopril
ladder in Section IV taught us more about where the line actually falls than any general
definition could, and thirty to fifty adjudicated examples across clinical domains and user
types, with the reasoning attached and the set expanded over time, would give founders
and their counsel something to reason from. Ambiguity resolved by example is cheap
for CDRH to produce and valuable to everyone who has to build against it.

A fast classification determination would do the rest. The 513(g) route already carries a
small-business rate at half the standard fee, which helps on cost, but what a founder
cannot get from it is something they can build a financing around. The response is
informational and non-binding. The statutory window is 60 days and actual responses
often run well past it, and preparing a credible request requires regulatory counsel a
pre-seed company has not yet retained. We therefore request that CDRH provide what
513(g) does not currently offer, namely, a binding classification determination available
to small sponsors on a committed clock. Recognizing that a 513(g) response is
informational and non-binding by statute, CDRH could use the Pre-Submission program
to issue written classification feedback with a commitment not to contradict it absent
material new information.

Response to Question 3
A patient reading an incorrect output usually cannot catch it the way a trained clinician
can, and we would not argue otherwise, so the concern CDRH raises here is the right
one. But the comparison being drawn is between a patient using the device and a
clinician using the device, when the choice many of our companies’ users are actually
making is between the device and an open web search, which carries no escalation
logic, no scope boundaries, no uncertainty calibration, and nobody accountable for the
answer. Measured against that, a patient-facing function with working safeguards is the
safer of the two options available.

We therefore ask CDRH to treat the patient-facing distinction as a modifier that comes
after a function’s safeguards have been assessed. The elements in Appendix A already
name most of what should count: enforced escalation on safety-critical presentations,
scope refusal with clinically useful redirection, calibrated uncertainty, comprehension
across health literacy levels, and a live path to a human when the conversation exceeds

FDA-2026-N-7874 · Redesign Health · Page 2 of 8
what the function can hold. Where a function demonstrates those things, its
patient-facing character has largely been addressed. Where it demonstrates none of
them, the uplift is warranted on its own terms.

Then publish which safeguards count and how much they count for, because a founder
can design to a published list and cannot design to a presumption. Without that list the
framework prices consumer-facing products a tier higher by default, and it does so for
the products reaching the people whose alternative is nothing at all.

Response to Question 5
Risk should be assessed at the level of the most directive or consequential function a
device is designed to reach in a realistic conversation, not at the level of any single turn.
The intended use statement should describe the range of functions the device may
perform across that conversation, including the highest level it is designed to reach, not
just the opening turn.

The worked-example library proposed above in response to Question 2 should include
multi-turn scenarios that show where the line falls as a conversation escalates. A
founder chooses architecture early, and the financing decision turns on whether a
product that starts informational but can reach action-directing territory will be classified
at the higher tier. CDRH should make that answer visible before code and capital are
committed.

Response to Question 6
Under-escalation and over-escalation are both material, but they measure different
harms and should not be collapsed into a single score. The acceptable trade-off
depends on the clinical context and should be prespecified by the sponsor and reviewed
by CDRH, rather than set by a universal formula.

This is especially important for the patient-facing products described under Question 3.
For a population whose realistic alternative is no escalation pathway, the baseline rate
of under-escalation is effectively 100%. A device that over-escalates some fraction of
the time may still be the safer option. CDRH should require sponsors to report underand over-escalation rates separately in benchmarking, rather than netting them.

Response to Question 11
The five approaches in Section V.C.1 give us something we can actually evaluate,
because each rung carries a knowable cost and a knowable timeline, and we support
the ladder for exactly that reason. What we need is predictability about which rung
applies.

We understand why CDRH would prefer to keep that selection flexible. Flexibility for the
agency translates to variance for the investor, and it is priced accordingly. We request
that CDRH publish a presumptive mapping from cells of the two-axis framework to a
minimum confirmation approach, rebuttable in either direction with justification. That

FDA-2026-N-7874 · Redesign Health · Page 3 of 8
mapping is also our answer to Question 8, which asks how the two-axis framework
should bear on the level of evidence a premarket submission needs. Non-directive
informational functions at low consequence would presumptively carry retrospective
evaluation on real patient inputs. Action-directing functions at moderate consequence
would carry shadow deployment with independent clinician adjudication. Prospective
study would be the presumptive approach for high-consequence action-taking functions
and may also be warranted for high-consequence action-directing functions where
retrospective or shadow evaluation cannot adequately establish safe use. Functions
whose outputs the user cannot independently assess, including the measurement and
signal processing functions described in Section IV, would carry evaluation against an
established reference standard or independent clinician adjudication, which is what
Section V.D.1 already contemplates for them and what ordinary diagnostic validation
practice already does.

A presumptive mapping does not bind CDRH’s judgment. It tells a founder what to plan
for, and it lets an investor size the round.

Response to Question 14
For functions intended to assist clinicians, the clinically relevant comparison may be the
care team using the device versus the same care pathway without it. CDRH should
accept that comparison alongside device-level testing appropriate to the intended use.
Evaluation should measure clinical errors, time to appropriate care, completed
escalations, clinician workload, and patient comprehension, and should test whether
clinicians can identify, reject, and correct unsafe outputs under realistic working
conditions. Then publish which functions qualify, because a sponsor needs to know
before they build whether their product will be evaluated as a team or as a device
operating alone, as they have different evidence and costs.

Response to Question 15
The comparators CDRH describes in Question 14, a panel of qualified clinicians or a
median clinician in practice, both assume the patient would otherwise see a clinician.
Our answer above takes that assumption as given, because a clinician-assist function
has a clinician in the loop by definition. For much of the population our companies
serve, that assumption simply fails. The alternative to the device is a six-week wait for a
specialist, a call that never gets returned, or nothing at all. The lack of access to timely
care across major sections of this country is well documented and has been repeatedly
cited by this agency as a matter of pressing concern.

Measuring against specialist consensus in that setting misprices the benefit twice. It
sets a bar the device does not need to clear in order to improve outcomes, and it leaves
the harm of the current situation unmeasured, because nobody is obligated to measure
it. The populations with the least access are exactly where a GenAI-enabled device has
the largest expected benefit and where the clinician comparator applies least.

FDA-2026-N-7874 · Redesign Health · Page 4 of 8
The realized counterfactual should count as a valid point of comparison in cases in
which access is the binding constraint. A sponsor should be required to characterize the
intended population’s actual access using wait times, appointment availability, and
geographic and payer coverage, document the outcome distribution under current
conditions from claims or registry data, and prespecify a performance threshold against
that documented baseline. Where a device serves a population that would otherwise
receive the care in question, the clinician comparator remains correct, and both can
appear in a single submission when the intended use statement identifies which
population is which. This does not ask for a lower safety standard for underserved
populations. Rather, it asks that the safety comparison reflect the choice those
populations actually face.

Response to Question 16
CDRH asked what program design features would prevent third-party participation from
limiting competition, and our answer comes from seeing the effect that accreditation
regimes have on a category’s cost of entry. Third-party evaluation would help our
companies, and a qualified evaluator holding sequestered datasets solves an
independence problem that a twelve-person company cannot solve internally. Section
V.B’s expectation of adjudicators structurally independent from both the sponsor and the
model developer is otherwise difficult for a small sponsor to satisfy credibly, because the
sponsor is the one paying the adjudicator. Market structure is where the damage would
come from. Five conditions would keep it healthy:

1.​ Participation must stay optional throughout, with no presumption against a
sponsor that runs and documents its own competency assessment, and no
added review time for doing so.

2.​ There must be multiple accredited bodies per scope from the start, because a
single accredited evaluator in a device category becomes a toll booth, and the
toll gets set at whatever an incumbent can pay. Where the market in a device
category cannot initially support more than one accredited body, the sponsor
should retain an unrestricted right to self-evaluate under Condition 1 until a
second body is accredited.

3.​ Published fee schedules and turnaround commitments should be a condition of
accreditation. An unpredictable queue is an unpriceable cost, and unpriceable
costs stop financings.

4.​ Conflict rules need teeth: no equity, board seat, revenue share, or commercial
relationship with a sponsor competing in the device category being evaluated,
and the restriction should reach affiliated funds and parent organizations. CDRH
should define “commercial relationship” so the restriction does not inadvertently
sweep in ordinary vendor or service relationships.

5.​ Sequestered benchmark assets need a public path. FDA-funded reference
datasets would spread the cost directly, and MDDT qualification would spread it a

FDA-2026-N-7874 · Redesign Health · Page 5 of 8
 second way, by making one well-built benchmark reusable across sponsors
instead of a fixed cost that every early-stage company has to carry alone.

Response to Questions 18, 22, and 23
Yes, CDRH should accept greater premarket uncertainty in exchange for greater
reliance on postmarket monitoring. The mechanism here is financing.

Premarket evidence is paid for with equity raised before a company has any revenue, at
the point where its cost of capital is highest, and postmarket monitoring is paid for out of
revenue. Moving evidence generation across that line changes the number a founder
has to raise before first revenue, and that number decides whether the company gets
founded at all. We would expect the trade to come with thresholds prespecified against
the premarket benchmarking baseline, a defined response protocol when one is
breached (including the ability to disable the affected function), and reporting that does
not depend on the sponsor’s own judgment about materiality. The trade should stop
short of the top of the matrix, however. Action-taking functions at high consequence are
where we suggest keeping prospective study as the presumptive rung under Question
11. Functions for which the user cannot independently assess outputs should keep their
premarket evaluation against a reference standard, because postmarket monitoring is a
weak substitute for premarket evidence when the user cannot detect the failure in the
first place.

On change control, the existing PCCP guidance does not obviously cover the changes
our companies actually ship: prompt revisions, retrieval corpus updates, guardrail
tuning, and movement between versions of an underlying model. Our companies ship
changes of this kind weekly. If each one requires a new marketing submission, then the
product cannot be maintained, and the consequence is that it never gets built.

A PCCP should cover all four of those categories provided the sponsor re-benchmarks
against the premarket competency baseline, meets prespecified acceptance criteria,
and documents the result under its quality system. The PCCP would prespecify the
categories of change and the acceptance criteria for re-benchmarking, rather than the
changes themselves. Reserve FDA review for changes to intended use, to clinical
scope, or to a function’s position on either axis of the risk framework. For model
versions, let a sponsor prespecify a qualified family and move within it on passing
re-benchmarking. On Question 19 and the cadence of postmarket reassessment,
automated re-benchmarking should qualify as a primary monitoring instrument in its
own right, with reassessment triggered by any change to the underlying model or to the
deployment architecture, because sampled clinician review at meaningful volume is a
fixed cost that will bind small sponsors first.

Response to Question 20
Yes. For small sponsors, a machine-based supervisory agent may be the only
postmarket monitoring that is economically viable at the volumes they actually serve.

FDA-2026-N-7874 · Redesign Health · Page 6 of 8
But the agent needs a defined performance bar: it should be benchmarked against the
same competency elements it monitors, at least for the safety-critical subset.

It should not replace periodic human review, but it can set the cadence. Let the agent
run continuously and trigger human review when it flags a threshold breach rather than
require fixed-interval review that a twelve-person company cannot staff. Sampled
clinician review at meaningful volume is a fixed cost, as this comment already notes; a
supervisory agent converts that cost into a variable one.

Response to Questions 24 and 25
We support voluntary Foundation Model MAFs, and the reason is bargaining power. A
seed-stage company has no leverage to compel notification of a model change. It buys
through a public API on standard terms. The company’s spend is immaterial to the
provider, and it learns about behavior changes through observation. Contractual
mechanisms are only available to companies with negotiating power. For everyone else,
the burden lands on the party least equipped to carry it, which has nothing to do with
which company builds the safer device.

An MAF corrects this cheaply, because most of the information already exists in
published model and system cards. Whether it works depends on the content and the
incentive. On content, standardize the MAF as an FDA-recognized schema covering
characterized behaviors and known failure modes in healthcare contexts. It should also
carry evaluation results on healthcare-relevant benchmarks including clinically
meaningful subgroups, safety-relevant guardrails, version history and deprecation
timelines, update notification commitments, and audit log availability. On incentive,
provide sponsors with concrete benefits for referencing an MAF, in review time or in
reduced sponsor-side characterization of the underlying model, because such benefits
are what incentivize model developers to file in the first place. Filing makes their model
the preferred substrate for regulated healthcare products.

On alternatives, the mechanism we would rely on is demand rather than obligation.
Where a developer has already published a model or system card covering most of the
schema, CDRH should be able to accept it as partial evidence, and where a developer
has disclosed nothing, the sponsor-side characterization CDRH asks for should be
correspondingly heavier. That asymmetry does the work without a mandate. A sponsor
choosing between two substrates will price the difference, and a developer that wants to
be chosen will file. This mechanism works best in a competitive model market and may
need to be revisited if model-provider concentration increases, because a sponsor with
no credible alternative cannot create the same demand pressure.

Participation should remain voluntary, with an equivalent-evidence pathway for
sponsors using models without a referenced MAF. Access to a model developer’s filing
should not become a prerequisite for market entry. CDRH should also clarify the
validation needed to qualify a replacement model, including one from another provider,
so that shared evidence reduces duplication while preserving competition and continued

FDA-2026-N-7874 · Redesign Health · Page 7 of 8
device safety and effectiveness. This is the cross-provider version of the qualified-family
PCCP we describe under Questions 18, 22, and 23, and it binds harder, because a
sponsor does not control when its provider retires a model.

Response to Question 26
Most agentic systems our companies build coordinate care, schedule appointments,
draft documentation, and conduct outreach. Section VII.B correctly suggests that many
such functions may fall outside the focus of device oversight. State that plainly and
stand by it, because agentic architecture makes function boundaries harder to see. That
creates pressure to treat orchestration as the device itself.

As regards acceptance criteria, element A.1 tests compliance with human oversight
checkpoints as a competency. Go further. An enforced checkpoint should move a
function left on the activity axis, and sponsors will build the checkpoint if it changes the
tier. If it only adds a test, the incentive is weaker. This framework-level point is also our
answer to Question 1, which asks whether reversibility belongs in the risk framework. It
does, and the cleanest way to represent it is to let a design control that keeps an action
reversible, or that interposes a human before it stops being reversible, count on the
activity axis.

We would welcome the opportunity to help with what comes next. We can convene
founders, clinicians, and technical teams from across our portfolio for a working session
on the benchmarking elements in Appendix A. We can also share what we observe
across the category about where evidentiary uncertainty is keeping companies from
forming at all.

Finally, we respectfully request that CDRH prioritize published examples, predictable
response timelines, and a transparent mapping from risk to evidence. These measures
would help founders design safeguards earlier, enable investors to finance appropriate
evidence generation, and give clinical partners a clearer basis for adoption. The result
would be a more predictable path to market for products that demonstrate safety and
effectiveness for their intended use.

Respectfully submitted,

Preston Tollinger

VP of Technology, AI Ventures

Redesign Health

preston.tollinger@redesignhealth.com

FDA-2026-N-7874 · Redesign Health · Page 8 of 8