AI-enabled is a list. Not a field.

FDA keeps a public page titled Artificial Intelligence-Enabled Medical Devices — every 510(k), De Novo, and PMA device the agency itself has flagged as AI or machine-learning enabled. It is the only place that status exists. No field on any openFDA endpoint this pipeline, or any comparable one, actually reads carries it.

Identified, not declared

A 510(k) applicant does not check a box marked “AI-enabled.” Nothing on the submission asks. FDA’s Digital Health Center of Excellence built the list itself, after the fact, by reading the decision-summary and classification language behind each clearance and approval for terms matching FDA’s own Digital Health and Artificial Intelligence glossary — a transparency commitment from the January 2021 AI/ML-Based Software as a Medical Device Action Plan. The list is a reading exercise FDA staff repeat periodically, not a designation any applicant applies for or is granted.

No field carries it

Every 510(k) FDA clears publishes to a fixed field set: K-number, applicant, device name, product code, device class, regulation number, decision code, the third-party-review flag, the statement-or-summary choice. The classification endpoint adds the review panel, the GMP-exempt flag, the implant and life-sustaining flags. None of those fields, on any endpoint, says whether the device behind them uses AI or machine learning at all. A relevance join built on the identifiers FDA actually publishes — product code, CFR regulation, application number, firm — has no key to join on here, because the fact has no column.

The list exists because a person at FDA read the file and decided the word applied. That is precisely the work a structured monitoring feed is built to replace — and precisely the one place here where it can’t.

The list says so about itself

FDA does not present the page as a complete count. Because identification runs on matching AI-related terms in decision-summary language rather than a structured designation, a device whose summary describes its function without using recognizable AI or machine-learning terminology can clear or gain approval and simply not appear — the page itself disclaims being an exhaustive or comprehensive resource. A competitor’s device that is functionally AI-enabled but was reviewed and described in earlier or plainer language is a gap the list acknowledges, not one a reader discovers by accident.

The practice

Reading a product code’s clearance history off openFDA, the way FDA Radar and any comparable feed does, per what we monitor and how often, shows every clearance in that code and says nothing about which of them are AI-enabled, because the fact never travels with the record. Cross-referencing a competitor’s K-number or device name against FDA’s own list by hand, periodically, is the only way to close that gap today, and the list’s own stated limits travel with whatever a reader concludes from it. This is regulatory intelligence and method, never regulatory advice about whether a specific device is, or should be, described as AI-enabled.