Device approval assumes a fixed product. A regulator assesses what it is sent, and the thing that reaches a patient is the thing that was assessed. Software that keeps learning breaks that assumption, because the version running in a clinic may differ from the version that was approved. The MHRA is now testing how to regulate a device that changes.

What the Commission concluded

The National Commission into the Regulation of AI in Healthcare published its recommendations on 10 September 2026. The government accepted all 44 of them on 6 October 2026 and set out how they would be delivered across the UK.

The Commission's central conclusion was that AI-enabled medical devices should be assessed and monitored throughout their working life rather than relying on one-time assessments. It framed the required change as a more proportionate, lifecycle-based framework rather than as an extra step bolted onto existing approval.

The Commission was chaired by Professor Alastair Denniston, with Professor Henrietta Hughes OBE, the Patient Safety Commissioner for England, as deputy chair. According to the government's response, more than 12,000 patients, clinicians and members of the public contributed to the recommendations.

The delivery structure was set out alongside the acceptance. A Programme Board convenes across government, the MHRA, NHS partners and the devolved authorities, and it begins work on the ten priority areas identified in the response. The Commission's chair described its guiding position as an approach that is people-centred and underpinned by a regulatory system that is safe, fast and trusted.

That is a consultation figure, not an outcome. Accepting a recommendation commits a department to act. It does not by itself change what a manufacturer must submit or what a regulator must check.

What the sandbox can test

The MHRA's AI Airlock is the regulatory sandbox for AI-enabled medical devices. Its third phase opened for applications alongside the government's response, and it concentrates specifically on post-market surveillance and lifecycle regulation.

That focus follows from the Commission. If a device must be monitored across its working life, someone has to establish what monitoring means in practice: which measurements, at which intervals, and what degree of drift justifies action. Phase 3 is designed to generate that evidence with real products rather than to write policy in the abstract.

Applications opened with a webinar for prospective applicants on 22 October 2026, and the first wave of innovators is to be selected in November 2026. Participants work with regulators and healthcare partners, and the resulting insights are intended to inform future MHRA guidance and policy.

Phase 3 is not the first regulatory work on AI devices. The MHRA has already issued guidance on ambient scribes, the clinical documentation tools that summarise a consultation, and the Commission's recommendations were described by one respondent as building on the momentum that guidance created rather than starting from nothing.

A sandbox can only test what participants bring to it. The devices accepted into a cohort describe the products that exist and the developers willing to be examined in public. They do not describe how the resulting rules would apply to a device that never enters the sandbox.

What is still being written

Three commitments carry dates. The MHRA has said it will issue draft guidance by December 2026 on a new approach to managing changes to AI-enabled medical devices as they adapt over time. It will begin the following year by consulting on how to qualify and classify AI-enabled devices, which was the Commission's first recommendation. A full implementation roadmap, setting out timelines and responsibilities across all 44 recommendations, is due by Spring 2027.

Staged authorisation pathways are also under exploration. These would allow promising AI tools into the NHS earlier under close supervision, while further real-world evidence is gathered.

The Commission's own description of that mechanism was more concrete. It recommended that new models be deployed under close supervision and tight guardrails, so they can demonstrate real-world safety and performance before receiving fuller authorisation, and compared the arrangement to the learner plates carried by newly qualified drivers. The point is that early access and reduced evidence are not the same thing.

The sequence matters more than the individual items. Classification determines which requirements apply. Change management determines how a device can be modified without a fresh submission. Staged authorisation determines how much evidence must exist before a patient is exposed. Until classification is settled, the other two are being designed against a moving target.

The direction also predates the Commission. The 2023 white paper, A pro-innovation approach to AI regulation, recorded that the MHRA would develop guidance on the transparency and interpretability of AI as a medical device, working with the Information Commissioner's Office and the National Data Guardian on questions of data protection and patient trust. What the Commission has added is a timetable, a named set of recommendations and an obligation on government to report progress against them.

Where the hard problem sits

The difficulty is not that regulators are unaware of the problem, and it is not only that the technology moves faster than institutions do. The evidence a regulator can demand has to be produced by the developer, and the developer's interest is in demonstrating that a change improves performance.

A regulator can require a description of what changed. Deciding whether a change is material, and therefore whether it needs reassessment, is a judgement that cannot be fully specified in advance. The Commission recommended a lifecycle approach. The practical content of that approach is what the sandbox, the December guidance and the classification consultation are meant to supply.

There is a timing problem underneath it. The first recommendation concerns qualification and classification, and consultation on that does not begin until 2027. Draft guidance on managing changes is due in December 2026, which is earlier. Guidance on change management may therefore be issued before the classification framework that determines what counts as a change to a class of device. Whether the two can be sequenced differently is not addressed in any source reviewed here.

What is not yet known is whether a lifecycle framework can be made specific enough to be enforced without becoming slow enough to defeat its own purpose. The roadmap has a delivery date. Whether the rules that follow shorten or lengthen the path from development to patient is not yet known.

Summary

The UK has accepted a lifecycle view of AI device regulation, and the instruments that would give that view legal force are still being drafted.

  • The commitment is complete. All 44 recommendations of the National Commission into the Regulation of AI in Healthcare were accepted on 6 October 2026.
  • The organising principle has changed. Devices are to be assessed and monitored across their working life, not at a single approval point.
  • The evidence is being built in a sandbox. AI Airlock Phase 3 focuses on post-market surveillance, with its first cohort selected in November 2026.
  • The rules have dates but no text. Draft change-management guidance is due by December 2026, classification consultation in 2027, and the roadmap by Spring 2027.
  • Enforceability is untested. No opened source shows how the materiality of a device change will be decided in a way that binds a manufacturer.

Acceptance is not regulation. Whether the framework is specific enough to be enforced is not yet known.

Research direction and editorial oversight: Juno Uni Hera. AI-assisted research and drafting; human review is ongoing.

Sources