The NHS has two AI problems. Neither of them is the technology.
I operate on patients. I also build artificial intelligence that predicts which patients on a hospital ward are about to come to serious, preventable harm. Pressure injuries. Surgical site infections. Falls. These are not rare conditions. They are the grinding, largely invisible toll of routine hospital care: hundreds of thousands of cases every year, billions of pounds in avoidable cost, and suffering that rarely makes the headlines.
The AI works. Our lead product, validated on more than 650,000 admissions at Imperial College Healthcare NHS Trust, significantly outperforms the manual risk assessment tools that nurses currently use. It runs on data hospitals already collect. It requires no new hardware or clinical roles. It is exactly the kind of technology that the Ten Year Health Plan, the Sovereign AI Fund, and the Medicines and Healthcare products Regulatory Agency’s AI Airlock programme are designed to encourage.
And yet getting it to the nurses who need it has been the hardest thing I have ever done. Because the NHS does not have one AI problem. It has two, and they require different solutions.
The first is a translation problem. The UK produces world-class clinical AI research but struggles to turn it into regulated products. The journey from academic validation to a medical device that can legally be used in patient care passes through intellectual property negotiations, technology transfer agreements, equity arrangements, and regulatory submissions. Our spin-out from Imperial required navigating relationships between the university’s technology transfer office, the Trust’s commercial directorate, and multiple external advisors, over many months. This is not unusual. It is the standard experience for any clinician attempting to translate health research into a product through the university system.
The academic career structure makes this worse. Building a regulated medical device demands skills in regulatory affairs, commercial negotiation, and company formation that are not recognised as legitimate academic outputs. A peer-reviewed paper counts. A regulatory clearance does not. Until the system recognises translation as a core output, the UK will continue producing research that other countries commercialise.
The second is a go-to-market problem, and it is not unique to AI. Any company selling an innovative product to the NHS faces the same barrier: more than two hundred acute Trusts in England, each operating its own procurement, information governance, and clinical safety processes. A product validated at one Trust has no expedited pathway to another, even when both run the same electronic health record system and face identical clinical challenges. Every new Trust means starting from scratch.
There is no national template for commercial terms. No standard data access agreement. No equivalent of the Lambert Toolkit that was created for university-industry collaboration. Every deal is bespoke, and every bespoke deal burns time and capital that early-stage companies do not have. For a small company, this fragmentation is not merely frustrating. It is existential.
What makes this particularly painful is that the upstream conditions are genuinely world-leading. Longitudinal NHS health records, a national deployment platform in the Federated Data Platform, and a regulator actively engaging with developers through the AI Airlock create an environment for developing clinical AI that no other country can match. The system that creates these extraordinary advantages for development erects extraordinary barriers to deployment.
Two interventions would transform this. First, a standardised NHS licensing framework for innovative health technologies, covering data access, revenue models, intellectual property, and governance. This would cut months from every Trust engagement and make the UK the easiest place in the world for health technology companies to scale, not just to start. Second, mandatory Trust-level reporting of preventable harm rates with the same transparency as cancer waiting times and accident and emergency performance. You manage what you measure. What you don’t measure, you tolerate.
The Government has committed real money and real political capital to making the UK the best place in the world to build AI companies. The science is ready. The regulation is improving. What is missing is the connective tissue: between a research finding and a regulated product, and between a regulated product and a patient who needs it.
Mikael Sodergren is a practising surgeon, Associate Professor at Imperial College London, and founder of Prelego, an Imperial spin-out developing AI for NHS patient safety. He is giving oral evidence to the House of Lords Science and Technology Committee on its inquiry into Innovation in the NHS.
Most Popular
What to read next
