News · 12 November 2026
Safe, fast and trusted: what the UK's healthcare AI blueprint asks of developers
In September 2026 the National Commission into the Regulation of AI in Healthcare, set up by the MHRA a year earlier and led by NHS doctors, published its blueprint for how AI-enabled medical devices should be regulated in the UK. According to legal commentary published in early November, the government has accepted all 44 of its recommendations. The MHRA has said it will publish draft guidance on managing changes to AI-enabled devices in December 2026, and a full implementation roadmap is expected in spring 2027.
For anyone building or buying clinical AI, the direction matters more than any single clause.
The shift: from approval to the whole lifecycle
The blueprint asks for oversight to continue after a product reaches the market, with more weight on real-world performance monitoring and post-market surveillance. It also supports predetermined change control plans, staged authorisation routes and wider use of regulatory sandboxes, so that a model can be updated without starting from zero each time.
Transparency, cybersecurity and equity are treated as safety
Three themes recur:
- Transparency. Tools such as model cards and dynamic labelling would explain intended purpose, performance, limitations and updates, so that a clinician can see what a tool was built and tested for.
- Cybersecurity. It is treated as a core safety requirement across the product lifecycle, not an add-on.
- Health equity. Diverse data and equitable performance are treated as safety issues, with continuing monitoring of subgroups.
What this means in engineering terms
If you build a model: keep a record of what data it was trained and tested on, state its intended use and its limits plainly, and plan how you will detect drift and subgroup differences once it is in use.
If you explain a model: an explanation is part of the evidence. It should show which inputs a result relied on and be checked against expert judgement, not only be plausible to look at.
If you buy a model: ask your supplier how changes are controlled and communicated, and who monitors performance after go-live.
What is still open
The Commission defers the question of liability to work being done elsewhere, and notes the risk of "liability sinks", where a professional carries responsibility for an AI error they could not reasonably have prevented. Classification of AI-enabled devices is also still to be consulted on. Treat the December guidance as the next real signal.
Sources: the Royal College of Radiologists' statement on the Commission's publication; Mills & Reeve, "AI in healthcare: reshaping the regulatory framework", 2 November 2026. This is general information, not legal advice.