The U.S. FDA posted a draft guidance for AI-enabled device software functions (DSFs) in January 2025, which provides much-needed clarity for a number of product types. All three of the agency’s life science product centers endorsed the draft, making this one of the more significant policy collaborations at the agency in recent years.
The participation of each of these four FDA offices is critical for makers of combination products, given that AI-driven combination products are playing an increasingly important role in medicine. For example, the FDA has cleared or approved continuous glucose monitors that make use of AI to help regulate the function of the monitor and the delivery of insulin. This is true as well for infusion pumps used in hospitals, just two categories of products that rely progressively more on AI with each passing year.
The draft guidance heavily emphasizes the total product life cycle (TPLC) as a method of managing the risks associated with product use. One of the more prominent concerns the FDA has in this regard is the prospect that the data processed by an AI system in the postmarket setting can succumb to data drift, which the agency believes could affect device performance.
The draft guidance offers a few cautionary points about the terminology associated with AI. One example is the term “validation,” which in FDA regulatory language is the process used to demonstrate that the product will perform as intended throughout the TPLC. In other settings, the term is used to describe the procedures used to optimize the performance of the AI model. The term “development” is also used differently across different settings, a point the agency is determined to emphasize.
The bulk of the contents of the draft provide insight as to the agency’s expectations regarding the content of premarket submissions, which includes a public submission summary. This document would address the FDA’s expectations regarding transparency around the data used for both the training and validation stages of development as well as the type of AI (e.g., convolutional neural network versus a recurrent neural network).
Information on how the sponsor intends to update and maintain the AI over time would also appear in this set of public disclosures. The draft recommends that sponsors use a model card to deliver this information to patients and other users.
The draft guidance recommends that premarket submissions include information on the developer’s quality system, a detailed description of the device, and information on the user interface and product labeling. Premarket applications would also have to include information on risk assessment and risk management, including information on these functions across the TPLC.
One of the potentially complicating factors associated with this guidance is the degree to which it overlaps with existing guidance. This draft interacts with two guidances related to combination products, such as the 2022 FDA guidance on principles of premarket pathways for combination products. There is also significant overlap with the 2023 guidance for the content of premarket submissions for device software functions, which also has been endorsed by the FDA’s three product centers for the life sciences.
The comment period for the AI-enabled DSF draft guidance closed in April 2025, at which point the agency had received 48 submissions to the docket. Some commenters expressing concerns about the scope of the draft, such as whether the scope should be limited to advanced types of AI rather than any system that falls into the AI category. A final version of this draft might not become available until late in calendar year 2026 as the guidance agenda for fiscal year 2026 suggests. The guidance agenda lists the final version of this draft guidance as a medium priority, appearing on the B list for final guidances rather than on the A list.