The UK’s AI-in-healthcare commission: what 44 recommendations mean for clinics.
The UK’s National Commission into the Regulation of AI in Healthcare has published its recommendations for how AI should be governed across the health system — 44 of them. None of it is law yet: the government and the MHRA will now respond. But the direction is clear enough that any clinic or hospital buying AI this year should build for it now.
Staged approval — new AI models allowed into use under defined conditions while evidence builds, like “L-plates” for learner drivers. Continuous monitoring — real-world performance tracked for the life of a product, not judged once at approval. Shared responsibility — makers, providers, clinicians and regulators each own part of safety, with stronger enforcement powers for the MHRA.
What the Commission actually proposed
The recommendations rest on three pillars. First, proportionate oversight across a product’s whole lifecycle: clearer routes to market, faster ways to approve updates, and more use of real-world evidence. Second, system-wide responsibility — manufacturers, providers, professionals, regulators and policymakers each get clear duties, and the framework asks for organisational readiness and workforce capability, not just product approval. Third, trust and transparency: more patient involvement, clearer accountability and more predictable pathways for the people building the technology.
Two proposals will be quoted most. Staged authorisations would let a new model be used under conditions while evidence accumulates. And approval would stop being a single moment: AI-enabled medical devices would be monitored continuously, in the real world, for as long as they are in use.
Why this matters even for “admin” AI
Much of what clinics are buying — scribes, prior authorisation, coding, patient messaging — is not a medical device. But the Commission’s logic travels. If you cannot show how a system performs on your patients this month, you cannot claim it is safe this month. Procurement teams already ask for that evidence; regulation is making it the norm.
What to ask any AI vendor now
How is accuracy measured after go-live? One validation report is the old model. Ask for ongoing evaluation on your own cases, with results you can see.
Who signs off? Every output that affects a patient should have a named human approver, and the system should record who approved what, and when.
Can you show the trail? For any output: what it read, what it decided and why — in a form an auditor can follow.
Where does the data live? Inside your environment, under your agreements, never used to train someone else’s model.
How we build for it
Every healthcare agent we build ships with an evaluation set drawn from the client’s own cases, measured before go-live and scored every week after. A clinician or biller approves outputs before they leave the building, and every decision is logged with its evidence. We did not design that for a commission; we designed it because confident wrong answers are the most expensive thing in healthcare. It happens to be where regulation is heading.
A worked example: one clinic, one scribe
Take a twelve-GP practice adopting an ambient scribe. Under a lifecycle model, the practice would want three things on file. A baseline: how long notes take today and how often they are finished after hours. An evaluation set: fifty recent consultations, redacted, against which the scribe’s drafts are scored before go-live. And a monitoring report: monthly, showing correction rates by clinician and specialty, with a named person who reviews it. None of that is exotic. It is what a careful practice would want anyway — the Commission simply makes it the expectation.
Common mistakes we see
Treating approval as the finish line. A tool that performed well in a vendor’s study can drift on your patient mix, your templates and your accents. Measure it on yours.
Leaving accountability vague. “The vendor is responsible” is not an answer a regulator or a patient will accept. Name the clinician and the manager who own each AI system.
Buying what you cannot inspect. If the vendor cannot show you a decision trail for one output, you will not be able to show it to anyone else either.
Our healthcare agents — ambient scribe, prior authorisation, coding, patient access — are shown working on the healthcare page, with before-and-after numbers, price and delivery time for each. Or read how we work with UK organisations.
Frequently asked questions
Are the Commission’s recommendations law?
No. They are recommendations to government and the MHRA, which will now consider them and respond. Organisations buying AI should still build toward them, because they describe where assurance is heading.
Does this apply to administrative AI like scribes or coding?
Many administrative tools are not regulated medical devices. But the same expectations — ongoing evaluation, named accountability and an audit trail — are increasingly asked for in procurement, so it is sensible to apply them.
What does “staged authorisation” mean?
The Commission proposes that new AI models could be used under defined conditions while evidence builds — compared to “L-plates” for learner drivers — rather than waiting for one final approval.
Sources: Burges Salmon; Computer Weekly; HLTH.