Let me start with the part of this argument that rarely survives contact with a political panel. Clinical AI is not a promise about the future. It is a product. It exists, it is cleared for use, and some of it is very good indeed. In peer-reviewed trials, AI systems have shown sensitivity and specificity for cancer detection that matches, and in particular tasks exceeds, that of specialist clinicians. Those are not pitch-deck numbers. They are published findings in serious journals, validated across more than one patient population.
I know this because I spent several years helping to build one of them.
At an AI diagnostics company I led the development of machine learning tools for oncology pathology: software that reads digitised images of tissue samples, identifies suspected cancer cells, and pushes the cases most likely to need urgent attention to the top of a pathologist's queue. What I remember most about that work is how little of it was the interesting part. The model was the easy bit. The hard bit was proving that a system trained on slides from one set of laboratories would still perform on slides from another, where the staining is a shade different, the scanner is a different make, and the tissue has been handled by people with slightly different habits. Months went into that question. We got the clearance. The product works, on real patients, in the United States.
Britain is a different picture, and the difference is not technical. The Royal College of Radiologists puts the shortfall in NHS radiology at roughly thirty per cent of the workforce needed to meet current demand. My own field was pathology rather than radiology, and I can tell you the staffing position there is no more comfortable. That shortfall is not an abstraction to be managed in a workforce plan. It is a queue, and at the end of the queue is a patient whose cancer is found later than it should have been. Tools that triage imaging, flag the urgent cases immediately, and take some of the routine load off exhausted clinicians are not a threat to that workforce. They are a lifeline thrown to it.
So the technology is not the constraint. The constraint is governance, and our governance framework manages the impressive trick of being simultaneously too loose and too tight. Not rigorous enough to protect patients properly. Not permissive enough to let demonstrably good tools reach them. Fixing one without the other would make things worse rather than better, which is roughly what we have been doing.
Take the rigour first, because it is the half that gets less attention. The MHRA's approach to software as a medical device has improved, but it still does not engage seriously enough with the problems that are specific to AI. A machine learning model is not a static device. It can drift, quietly, as clinical practice changes and patient populations change and the equipment feeding it gets replaced, which means a tool that was accurate on the day it was approved may be less accurate two years later without anybody noticing. It needs real-world monitoring after deployment, and it needs an explicit protocol for the cases where the model's own confidence is low and a human being has to step in. None of this is exotic. It is basic hygiene, and our framework barely addresses it. I will admit that a tougher British regulator would have been inconvenient for me at the time. I would also have respected it more, and so would the market.
Now the permissiveness, which is the more familiar British failure. A tool can hold regulatory clearance and a NICE endorsement and still face a fresh, full-scale evaluation in every trust that might buy it, each with its own evidence submission, its own pilot, its own commercial negotiation. Consider the arithmetic from the other side of the table. A pilot costs money and generates no revenue. A company with eighteen months of runway cannot run twelve of them and survive. So it does the rational thing and sells into a market that will actually buy, and we congratulate ourselves on our caution while the capability drains out of the country. I have watched genuinely good technology die of politeness in this system: nobody refuses it, nobody is against it, everyone wants to see it succeed somewhere else first.
What would actually work is not complicated to describe, though it is hard to deliver against the grain of institutional habit. Three things. A regulatory pathway that engages developers early, sets a demanding and AI-literate validation standard, and produces an approval the whole service respects rather than relitigates locally. A mandatory post-market surveillance regime, with standardised metrics and reporting, so that approval is the start of the evidence rather than the end of it. And a single national procurement route for approved diagnostics, so that clearing a high bar once actually means something.
High bar, applied once, then get out of the way. That is how a Conservative ought to think about regulation generally, and it happens to be exactly what clinical AI needs.
I should deal with the objection that carries the most emotional weight, because it deserves better than a brush-off. The fear is that the algorithm displaces the doctor, that care becomes colder, that something essential is lost. I understand it and I do not share it, mostly because I have seen what these tools do in a working department rather than in a headline. The good ones do not replace the pathologist. They clear the queue of the straightforward cases so that human attention lands where human attention is genuinely required. Used well, AI takes out some of the drudgery and leaves more room for the medicine. Whether it is used well, though, is entirely a governance question, which brings us back to the competence of the state, which is where the argument should have been all along.
What frustrates me is that this ground is uncontested. The boosters have colonised the tech-optimist right, promising a revolution by next spring and setting themselves up to disappoint. The sceptics have colonised parts of the left and the professional bodies, and will win by default if nobody turns up. The sensible position, that these tools are real and valuable and require precisely the sort of rigorous, decisive, then hands-off regulation Conservatives claim to believe in, is sitting there unclaimed.
There is a more optimistic version of this piece and I have tried to write it twice. The numbers keep spoiling it. AI will not save the NHS. Nothing saves the NHS on its own, and I would not insult anyone's intelligence by pretending otherwise. But governed properly and let through a door we currently keep bolted, it could help a great deal, and it is one of the very few productivity levers available that does not require hiring anybody or building anything.
I am not confident we will get this right. I am fairly confident we will not get it right by accident. So I intend to keep making the case, in committee rooms and in print, that this is exactly the patient, unglamorous, structural work a serious government exists to do.
I write here in a personal and council capacity. I have worked in pharmaceutical drug development and in the development of clinical AI diagnostics, and readers should weigh what follows accordingly.