The FDA has prioritised final guidance on AI-enabled device lifecycle management and draft recommendations for conversational mental-health devices in its fiscal-year 2027 agenda. The plans do not create new legal duties, but they give developers and buyers a clearer basis for reviewing evidence, monitoring and change-control arrangements.

What changed: a prioritised agenda, not a new rule

The US Food and Drug Administration’s Centre for Devices and Radiological Health (CDRH) has placed final guidance on marketing submissions and lifecycle management for artificial intelligence-enabled device software functions on its fiscal-year 2027 A-list. Its published agenda, announced in the agency’s 1 October update, also prioritises draft evidentiary recommendations for generative AI-enabled conversational devices for mental disorders. These are publication intentions—not completed guidance or new legal requirements.

The distinction is important. The A-list identifies priority documents; the B-list covers work intended as resources permit, while an “Under Construction” list identifies further development topics. Fiscal year 2027 runs from 1 October 2026 to 30 September 2027, but inclusion does not establish a document-specific publication deadline. Guidance also generally describes agency thinking rather than creating legally enforceable responsibilities, unless it cites applicable statutory or regulatory requirements.

Lifecycle management moves beyond a submission-day benchmark

The planned final device guidance follows the FDA’s January 2025 lifecycle-management draft. That draft addresses device descriptions, data management, model development, validation, user interfaces, performance monitoring and cybersecurity. It remains labelled “Draft—Not for Implementation”. Its inclusion on the final-guidance agenda signals an intention to complete that work, not confirmation that the draft’s recommendations will survive unchanged.

For manufacturers, the practical issue is the relationship between premarket evidence and performance after deployment. An imaging algorithm may encounter different scanners or acquisition protocols; an EEG-analysis system may receive recordings with different artifacts or patient characteristics. The draft explicitly discusses sensitivity to changing input data, subgroup performance and deployment conditions. Organisations should assess whether their evidence describes the actual intended use—not merely a favourable aggregate result from a development dataset.

The agenda separately prioritises final guidance on predetermined change control plans for medical devices. However, an AI-specific final PCCP guidance already exists, updated in August 2025. It recommends describing planned modifications, validation and implementation methods, and impact assessment. FDA reviews the plan within a marketing submission; this is not permission for unrestricted retraining or arbitrary post-authorisation changes.

Conversational AI receives a distinct evidence track

The A-list’s proposed draft on conversational devices for mental disorders is particularly relevant to patient-facing generative AI developers. Its wording concerns devices, not every chatbot that discusses wellbeing. The agenda does not establish which products will fall within its scope, what studies FDA will recommend, or whether particular conversational features will require specific safeguards. Those details remain to be published.

A broader document on “Risk Assessment and Evidentiary Considerations for Generative Artificial Intelligence-Enabled Devices” appears only on the Under Construction list. Developers should not treat it as imminent final policy or assume it supplies an established approval pathway. Nevertheless, the separate topics give organisations a useful planning question: can their evaluation capture clinically consequential conversation failures, changing outputs and inappropriate user reliance, rather than only answer quality on a fixed benchmark? That is an operational inference, not a newly announced FDA requirement.

Life-sciences teams need a different scope assessment

This is a device-centre agenda, not a universal compliance framework for pharmaceutical AI. Software used for clinical decisions may raise different questions from models supporting drug-discovery hypotheses, clinical-trial analysis or manufacturing quality. The device lifecycle draft concerns functions meeting the statutory device definition and also addresses relevant device constituent parts of combination products. Organisations should assess intended purpose and function before assuming the recommendations govern their system.

For drug and biological products, the FDA maintains a separate draft credibility framework. It concerns AI producing information or data intended to support regulatory decisions about safety, effectiveness or quality. A model contributing to an endpoint analysis or safety assessment therefore needs a defined context of use and proportionate credibility work; it should not be treated as equivalent to an exploratory discovery model simply because both use machine learning. The document remains draft, non-binding guidance.

Operating implications: connect evidence, deployment and updates

The immediate preparation task is to map documentation gaps against the existing draft while preserving flexibility for the final text. Its recommendations connect data provenance and representativeness with independent performance evaluation, subgroup analysis and the intended clinical workflow. For procurement teams, that translates into requesting evidence for the relevant population, input devices, operating conditions and software version—not accepting “FDA-cleared AI” as a complete description of suitability. These are evidence questions informed by the draft, not new buyer obligations created by the agenda.

Change control deserves equally concrete treatment. Manufacturers should assess how model versions, validation results, deployment records and performance signals connect within their quality processes. Buyers should establish how suppliers communicate updates, limitations and corrective actions, and who assesses local consequences. Under the existing PCCP guidance, planned changes require a defined assessment methodology. An authorised plan does not remove the need to show that an implemented modification follows it.

There is also meaningful international convergence without a single global authorisation. FDA, Health Canada and the UK Medicines and Healthcare products Regulatory Agency have published joint PCCP principles. Separately, FDA and the European Medicines Agency issued good AI practice principles for drug development in January 2026, covering context of use, data governance, risk-based assessment and lifecycle management. Global teams can organise evidence around these shared themes while checking each jurisdiction’s applicable requirements.

What to watch during fiscal year 2027

Three events warrant close attention: publication of the final AI-device lifecycle guidance; release of the proposed conversational mental-health device draft; and progress on the broader generative-device risk document. The first two are A-list priorities; the third remains under construction. None has a verified document-specific publication date in the agenda. When drafts appear, organisations should check the associated notices for consultation arrangements rather than assume deadlines in advance.

The strategic implication is preparation without premature legal conclusions. Teams can strengthen the connection between intended use, validation, deployment monitoring and controlled updates now, while distinguishing current obligations from proposed recommendations. The FDA’s agenda is valuable as a planning signal, not as evidence that every healthcare AI product faces the same pathway or that tomorrow’s guidance is already binding.

HealthyData.Science tracks AI regulation developments across clinical research, healthcare and life sciences, helping readers understand what each update means in practice.