AI medical device regulation after approval: What NHS trusts should set up now

By Rafael Patron, president of the Scientific-Technical Committee, AIPIA
Picture a hospital trust eighteen months into using an AI tool that flags suspected abnormalities on chest X-rays.
It was assessed, procured and switched on, and since then nobody has formally asked whether it still performs as it did on day one. That is the question AI medical device regulation in the UK is now turning towards.
On 6 October the government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare, and the direction is clear: less weight on a single moment of approval, more on how an AI-enabled device behaves across its working life.
For the people who run these tools inside the NHS, that shift lands on their desks well before any new rulebook does.
What was decided, and what is still to come
The facts are worth setting out precisely, because much of the detail is not yet written.
According to the government’s announcement, the Commission’s report was published on 10 September and the response accepts every recommendation.
The MHRA has committed to draft guidance by December 2026 on a new approach to managing changes to AI-enabled medical devices as they adapt and improve.
A consultation on how these devices should be qualified and classified will start next year. The government will also explore staged authorisation pathways, and a full implementation roadmap with timelines and responsibilities is promised by spring 2027, overseen by a new Programme Board.
Alongside this, the MHRA has opened applications for Phase 3 of AI Airlock, its regulatory sandbox for AI as a medical device (AIaMD), with a focus on post-market surveillance and lifecycle regulation and the first participants due to be chosen in November.
None of this tells a trust exactly what it will have to do.
It does tell it where to look.
Why AI medical device regulation now follows the whole lifecycle
Under a model centred on approval, the hard work sits with the manufacturer before launch.
Under a lifecycle model, a good part of the evidence can only come from the places where the tool is actually used: the scanner it reads from, the patients it sees, the clinicians who accept or override its output.
Health innovation minister James Frith put the principle plainly in the announcement: “innovation must never come at the expense of patient safety.” Patient safety after go-live depends on information that only hospitals hold.
This is where gaps tend to appear.
A procurement file built for a one-off decision is likely to centre on the vendor’s validation study and a clinical safety case. It may say much less about what will be measured in month six, who will look at it, and what happens if the numbers drift.
What to watch once an AI tool is live
The most useful starting point is modest. Take one deployed tool and write down the two or three measures that would tell you it is getting worse.
For a triage or detection tool, that might be the rate at which clinicians disagree with its flags, compared against the rate seen during the first weeks of use.
For a documentation tool, it might be the share of outputs that need substantial correction before they go into the record.
Then look at the inputs.
Performance often shifts because something upstream changed: a new imaging protocol, a different patient mix after a service reconfiguration, a software update on connected equipment.
A note of those changes, kept alongside the performance figures, turns an unexplained dip into something a team can investigate.
Finally, decide who sees the figures and how often. A monthly review by the clinical lead and the digital safety officer is enough to begin with.
When the model changes
AI products can be retrained and updated in ways traditional devices rarely are, and the MHRA’s draft guidance on managing change is the piece of this programme trusts should watch most closely.
Until it appears, the sensible course is to avoid guessing at its content and instead get local records in order.
That means knowing which version of each model is running, when it last changed, and what the supplier said about the change. Older contracts may say little about how model updates are notified.
Asking suppliers now how they will inform the trust of a retrained or modified model, and what evidence they will share, costs nothing and will make any future requirement easier to meet.
It also means agreeing internally who can accept an update.
A new version that improves sensitivity on paper may behave differently on local data, and someone with clinical authority should decide whether it goes live.
Who owns the warning signal
Monitoring only works if the person who spots a problem knows what to do with it and feels able to act.
Inside a trust, that usually means a named clinical owner for each tool, a clear route for reporting concerns to the supplier and, where appropriate, to the regulator, and the authority to pause the tool and fall back to the previous pathway.
The same question applies on the other side of the contract, to the data scientists, engineers and consultants who build, retrain and integrate these systems.
The Italian Association of Artificial Intelligence Professionals (AIPIA), a non-profit body for people who design and manage AI systems, treats this as a matter of professional conduct as much as regulation.
Its code of ethics for AI professionals asks members to make systems understandable and, where possible, auditable in their results and decision processes, and to keep decisions of significant ethical or legal weight under human oversight and responsibility.
For NHS teams, that is a reasonable expectation to set with any supplier: the people behind the model should be able to explain what changed, why, and how they know it still works.
Health Tech World’s readers sit on both sides of that relationship, as buyers inside the NHS and as developers selling into it.
A shared expectation about traceability is something they can agree on long before the December guidance arrives.
Staged authorisation and the evidence trusts already hold
The government has said it will explore pathways allowing promising AI tools to be used in the NHS earlier, under close supervision, while more real-world evidence is gathered.
If those pathways materialise, the supervision will not happen in Whitehall. It will happen in trusts, through the monitoring, logging and review routines described above, and a trust that cannot say how its current tools are performing will struggle to take on that role.
Phase 3 of AI Airlock is open not only to device manufacturers but to academics and developers of related solutions, and its focus is exactly the post-market evidence that hospitals generate every day.
The regulator is, in effect, asking how lifecycle oversight should work in practice.
Trusts that have kept a year of structured performance records will have something concrete to contribute to that conversation, and to the consultation on classification that follows in 2027.
The detailed rules will take shape over the coming months.
The data that tests them is being produced on NHS wards today, whether anyone is recording it or not.










