News · 22 October 2026
Lifecycle, not launch: the FDA's FY2027 agenda for AI-enabled devices
On 1 October 2026 the US Food and Drug Administration's Center for Devices and Radiological Health published its guidance agenda for fiscal year 2027. Among the top-priority items is finalising its recommendations on lifecycle management for AI-enabled device software functions, which was first issued as a draft in January 2025. Finalising guidance on predetermined change control plans also features, along with a draft for generative-AI-enabled conversational devices intended for mental health disorders.
Although this is US policy, it matters well beyond the US. Developers in the UK and EU often plan for several regulatory regimes at once, and regulators observe one another closely.
The central idea: total product lifecycle
Traditional device regulation focuses on the moment of authorisation. AI changes that, because performance depends on data, which changes, and models are updated. The lifecycle approach asks manufacturers to show control across the whole life of a product: design, validation, deployment, monitoring and change.
What that looks like in engineering terms
A clear intended use with defined users and conditions, and an honest statement of what is out of scope.
Data management. Documentation of training and test data, how it represents the intended population and how bias was assessed.
Performance evaluation that includes subgroups and realistic settings, not only an aggregate score.
Transparency. Users should be able to understand what the device does, how well it performs and where it may be wrong. For many tools, this implies interpretable outputs and clear labelling.
Change control. A plan, agreed in advance, for how the model will be modified, how each change will be validated and what stays fixed.
Post-market monitoring. Collection of real-world performance, detection of drift and a route for reporting and acting on problems.
What to build now
These expectations converge across the FDA, the UK's regulator and the EU. A team that builds versioned data and model records, repeatable evaluation, drift monitoring and clinician-facing explanations will find any of those frameworks easier to meet.
Treat each as an engineering requirement from the first prototype, not a compliance task at the end. Retrofitting traceability onto a mature model is slow, and it is the most common cause of delay we see in regulated projects.
Written in October 2026; details reflect the situation at that time.