AI for medical device regulatory affairs: what it can and can’t do
Step by step: where AI helps and where it stops
| Workflow step | What an AI assistant can do | What it can’t | Verify against |
|---|---|---|---|
| Predicate research | List cleared devices under your product code, compare their intended use and technology from the 510(k) summaries, and show which predicates they cited. | Decide substantial equivalence. FDA makes that determination; the tool only gathers the comparison. | The 510(k) database record and summary for each candidate, against FDA’s substantial equivalence criteria |
| Testing precedent | Collect the bench, software, clinical and other testing that recent summaries describe, and the special controls in your device’s regulation. | Tell you your testing is enough. A summary describes what one submission included, not what your device needs. | The summaries, and the special controls in 21 CFR, such as 21 CFR 892.2090(b) for detection software |
| Standards | List the consensus standards comparable devices declared, edition by edition. | Confirm which edition FDA recognizes today unless it reads FDA’s recognized standards database for you. | FDA’s Recognized Consensus Standards database |
| Pre-Submission prep | Draft questions from precedent: where your device differs from the predicates and what testing those differences may need. | Give FDA’s feedback. Only FDA can, through the Q-Submission program. | FDA’s Q-Submission guidance (final, May 2025) |
| Post-market signals | Summarize recalls and adverse event reports for your product code and your predicates, grouped by failure mode. | Establish cause or rates. FDA says the cause or incidence of an event cannot be determined from its adverse event reports alone. | FDA’s recall database and MAUDE |
Citation and verification discipline
The rules we hold Bearnier AI to, and the ones to hold any tool to.
1. Every precise fact carries its source
K-numbers, dates, product codes, regulations and test results each link to the FDA document they came from. A fact without a link is a lead, not evidence.
2. Open the source and read the passage
Before a fact goes into a submission, open the FDA page and confirm it says what the tool says it says.
3. Cite FDA, not the tool
In a submission, cite the 510(k), De Novo decision, regulation or guidance itself.
4. Check the status and date
Guidance changes, and drafts are not for implementation. FDA’s 2023 predicate selection guidance, for example, is still a draft.
5. Re-check every K-number
General assistants can produce K-numbers that do not exist or belong to a different device. Look each one up in FDA’s 510(k) database.
6. Prefer refusals to guesses
A tool that says the record does not answer the question is more useful than one that fills the gap.
7. Keep confidential content where it belongs
Do not paste unreleased submission content into tools whose terms allow training on your inputs.
8. Record what you checked
Note who verified each fact, against which document, and when, so a reviewer or auditor can follow it.
Try it on your device
Pick one, or ask your own. Every precise fact opens to its FDA page.
Find a predicate in ten seconds
Free, no account. Straight from FDA’s 510(k) database.
| 510(k) | Device | Applicant | Decision |
|---|---|---|---|
| K253703 | RemunityPRO+ Pump for Remodulin® (treprostinil) Injection | Deka R&D Corp | Jul 24, 2026 |
| K252015 | FreedomEdge Infusion System; High-Flo SubQ Needle Set; Precision Flow Rate Tubing | Koru Medical Systems, Inc. | Jan 28, 2026 |
| K251636 | Spectrum IQ Infusion System with Dose IQ Safety Software (3570009) | Baxter Healthcare Corporation | Jul 28, 2025 |
| K251640 | SIGMA Spectrum Infusion Pump with Master Drug Library | Baxter Healthcare Corporation | Jul 28, 2025 |
| K242752 | MRidium 3870 MRI Infusion Pump System (3870) | Iradimed Corporation | May 28, 2025 |
| K243855 | BD Alaris Infusion System with Guardrails Suite MX | Carefusion 303, Inc. | Apr 25, 2025 |
FDA’s database could not be reached. Try again in a moment.
Questions about AI in regulatory affairs
Can AI write my 510(k)?
It can draft sections from precedent, such as the device description or a substantial equivalence table, but a person on your team has to check every fact against FDA’s record and decide what the submission argues. FDA decides substantial equivalence.
Can I cite an AI tool in a submission?
Cite the FDA documents instead: the 510(k) summary, De Novo decision, regulation or guidance the fact came from. A tool that links every fact to its page makes that easy.
What is the best tool to find a 510(k) predicate?
FDA’s 510(k) and Product Classification databases are free and authoritative; Bearnier AI searches the same records and cites every fact; general assistants need every K-number checked; consultants add judgment. The predicate search tools page compares them.
How is this different from asking ChatGPT?
A general assistant answers from training data and the open web. A tool built on FDA’s records searches them and links each fact to its page, and should say when the record does not answer the question.
Sources
- Premarket Notification 510(k)
- 510(k) Premarket Notification database
- Best Practices for Selecting a Predicate Device to Support a Premarket Notification [510(k)] Submission
- 21 CFR 892.2090: Radiological computer-assisted detection and diagnosis software
- Recognized Consensus Standards: Medical Devices
- Requests for Feedback and Meetings for Medical Device Submissions: The Q-Submission Program
- Medical Device Recalls database
- MAUDE: Manufacturer and User Facility Device Experience
- Manufacturer and User Facility Device Experience Database (MAUDE): limitations of medical device reports
- Product Classification database