Evaluate the authorized function, not the AI label.

Locate the actual authorization

FDA's AI-enabled device list links to public authorization records. It is not a complete inventory of every AI-enabled device. Follow the submission number to the applicable record and supporting documents rather than treating a vendor-wide badge as authorization for every feature.

Questions to take into the conversation

• What exact function and intended use does the record cover?
• Is the product for triage, detection, measurement, or another purpose?
• Does the software version and operating environment match?

Examine clinical and workflow fit

FDA authorization is distinct from local implementation suitability. Review relevant validation, the comparator, operating point, exclusions, and subgroup results. Then examine how the output reaches the team and how false alerts or missed findings are handled.

Questions to take into the conversation

• How does performance change with your scanners, protocols, and case mix?
• What is the false-alert burden at your expected prevalence?
• Who receives, acknowledges, and acts on an alert?

Make monitoring part of the purchase

Our recommended evaluation includes a baseline, an approved local validation plan, named oversight, and a rollback path. Document how software changes are assessed and how performance concerns reach the responsible team. A successful demonstration alone cannot settle these questions.

Questions to take into the conversation

• Which outcomes are monitored after deployment?
• How are model changes and new limitations communicated?
• What conditions require pausing use or revalidation?

Sources & scope

This guide combines HealthIT's suggested evaluation questions with the primary references below. It is educational material; application depends on the specific product, workflow, organization, and jurisdiction.

• FDA · AI-enabled medical devices ↗

Found something that needs correction? Send a source-backed correction.