Thermo Fisher Scientific (Jody Schulz)

See attached file(s)

Thermo Fisher Scientific
Life Technologies Corporation
Life Sciences Solutions
5781 Van Allen Way
Carlsbad, CA 92008
www.thermofisher.com

Docket Management Staff
(HFA-305) Food and Drug
Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

19 October 2026

Subject: Comments on the FDA discussion paper “Considerations for the Regulation
of Generative AI-Enabled Medical Devices: Discussion Paper and Request for
Feedback” (Docket No. FDA-2026-N-7874)

Dear FDA Review Team,

On behalf of Thermo Fisher Scientific, we appreciate the opportunity to provide comments
on the discussion paper, “Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback”. As a global provider and
manufacturer of laboratory instruments, assays, and interpretive software, Thermo Fisher
Scientific is committed to enabling the safe, reliable, and compliant development and
manufacture of these products.

We commend the FDA for its continued efforts to provide clarity in this evolving area and
respectfully submit the following comments and recommendations for consideration.

1. How should the risk of an AI-enabled product be determined?

We agree that a two-axis framework is a useful starting point. Risk assessment should
consider both the type of information or direction the AI provides and the potential
consequences if it is wrong. It should distinguish informational outputs from
recommendations and actions, and consider factors such as reversibility, treatability, and
the population at risk.

pg. 1
The appropriate level of oversight should also reflect who will use the product, their
training, and whether they can verify its output. For patients, medication selection or dose
changes should require clinician involvement; lower-risk guidance, such as hydration or
skin care advice, may need less oversight. For clinicians, specialty and experience matter:
a generalist may need specialist oversight when managing complex conditions.

Products that allow repeated prompting also warrant attention, since users may steer a
model toward unsafe advice. FDA could consider a staged approach, beginning with lowerrisk use devices and expanding as evidence grows, unless the expected benefits justify
earlier use in higher-risk settings.

2. How can safety and effectiveness be demonstrated when all inputs and outputs
cannot be tested?

We agree that a competency-based approach is appropriate. Testing should be designed
with medical subject-matter experts and representative end users, and should be
proportionate to the product’s risk. Clinical expertise should match the intended use-for
example, oncologists evaluating oncology-related recommendations.

Validation should use clinically representative conditions and appropriate comparators.
Real-world data and studies of actual use can provide valuable evidence, supplemented
by synthetic data where useful. FDA should also expect manufacturers to document which
populations were validated and assess whether the intended-use population differs in
ways that could affect outcomes.

3. How should performance be monitored after launch?

FDA should consider noting in their guidance that post-market monitoring should be
proportionate to the product’s intended use and risk. It should include periodic review and
retesting of real-world outputs including complaint monitoring and trending, with
investigation of differences from expected performance and clear criteria for when a
change may affect intended use. Monitoring should also consider cybersecurity events and
other market signals that could reveal changes in performance or risk.

4. How should changes to AI-enabled products be managed?

FDA should rely on their existing device-change, software-as-a-medical-device, and Class
II and III guidances which provide a useful foundation noting when new submissions may
be required based on the level of change. Change assessments should consider effects on
intended use, safety, and effectiveness.

pg. 2
Predetermined Change Control Plans can define changes that may be implemented under
an approved plan, while modifications requiring additional review should be identified
clearly. Manufacturing and related records may also be useful in appropriate cases.

5. What responsibilities and information should apply to third-party model providers?

FDA should consider adding specific information to the guidance regarding how to manage
third-party model providers including use of contracts and quality agreements. These
agreements should require timely notice of model updates, discontinuations, and other
material changes. Manufacturers should receive enough information to assess the impact
and complete any needed validation before a provider changes or retires the model version
in use.

6. Do AI agents require a different level of oversight?

FDA should consider greater reliance on postmarket evidence when appropriate, but
monitoring should be risk-based and capable of identifying problems promptly—not
limited to annual review. Greater autonomy and fewer opportunities for human review
warrant stronger safeguards and closer monitoring particularly in the area of cybersecurity
risks or threats. FDA Machine-based supervisory agents may help monitor AI systems, but
those supervisory tools should themselves be validated.

Respectfully Submitted,

Jody Schulz

Sr. Director, Regulatory Affairs

Thermo Fisher Scientific

pg. 3