Summary
Most NHS AI pilots end the same way. The model performs. The clinicians like it. The evaluation reads well. Then the pilot money runs out, the clinical lead rotates and it stops. Nobody did anything wrong, and nothing changed.
Content
The hard part of AI in the NHS starts after the model works
I spent 7 years selling clinical risk models into the NHS, and later into Canada, New Zealand and Sweden. The algorithm was almost never the hard part. Everything after it was—and probably harder than anything else I've done.
Re-ordering the waiting list was potentially worth more than £2 billion
During the Covid-19 pandemic, the elective waiting list grew faster than any recovery plan could clear it, and the NHS ordered that list the way it always had: by the date the referral arrived. That ordering assumes every patient deteriorates at the same speed. They don't.
The alternative we built scored every waiting patient on their own clinical risk, using outcome data from 650 million patient records across 46 countries, then re-orders the list by who is deteriorating fastest. It ran across around 30% of the NHS. The NHS-reported results were 8% fewer emergency admissions, 125 bed-days saved per 1,000 patients and more than £2 billion of cost avoided. In one Swedish hospital in 2022, 122 deaths were avoided. The company was named in the House of Commons in June 2020.
That result came from a different order of names on a list that already existed. No new drug, no new building, no new consultant.
It is the shape of most real AI value in healthcare today: better decisions about people the system has already found.
Five things stop a working model reaching patients
Every stalled deployment I have seen failed at one of five points. Referential data, algorithms, validation, trust or integration.
Referential data: A risk model needs outcomes from populations far larger than the one it is deployed into. A single trust holds nowhere near enough. Most UK health AI projects start by trying to build the reference set they should have bought, and spend a year on it.
Algorithms: This is now the cheapest part of the problem, and it is where most of the attention and most of the funding still goes. Ask what share of your AI budget touches the model and what share touches the pathway around it. The answer is usually wrong.
Validation The NHS asks a fair question: does this work on my patients, not somebody else's? I was the only private-sector expert on the Medicines and Healthcare products Regulatory Agency (MHRA) project that generated synthetic healthcare datasets to validate AI software, and that work exists because the question is hard to answer honestly. Validate against local outcomes and publish the result, including the cases where the model was wrong.
Trust: Clinicians adopt tools that tell them something they can check and act on the same day. They reject a score that arrives with no reasoning attached and no route to challenge it. That is professional accountability working as designed. The best adoption I saw came where a consultant could open the list, see why a patient had moved up and disagree. A tool that cannot be overruled does not get used twice.
Integration: If the output does not appear inside the system a clinician already has open, it does not get used. Every trust knows this. Every business case still underestimates it.
Five fixes, in the order they matter (your mileage may vary)
Contract for outcomes. Write the agreement around bed-days, admissions or waiting time, and make payment follow the measured change. Suppliers who believe their own numbers will sign. Of course NHS finance teams like certainty of costs so this is difficult and, furthermore, when a hospital was faced with a bill they had not accrued for under this model, the tens of millions saved and deaths avoided went to the back of the queue.
Fund the clinical time to act on the output. Somebody has to review the flagged patients, make the call and book the list. That capacity is the deployment. Software with nobody paid to respond to it is shelfware with a licence fee.
Set the baseline before you switch anything on. Most NHS AI projects cannot prove their own result, because nobody recorded what normal looked like in the same population, in the same months, before the tool arrived. Six weeks of baseline data is the cheapest insurance in the project.
Name the clinician who owns the decision. One person, named in the business case, whose job description includes this. Projects owned by a committee die when the committee's priorities move, which takes about two quarters.
Decide at the start what happens if it works. This is the one that kills good pilots. There is often no standing budget line, no procurement route and no ICB-level owner waiting to scale a successful trust-level result, so success produces a report instead of a rollout. Agree the route to scale before the pilot starts, or accept that you are running an experiment for its own sake.
The gap sits between one hospital and fifty
None of the five is a technology problem, which is why AI in the NHS keeps producing good evidence and poor spread. The UK has the model builders, some of the deepest outcome data anywhere and a health service that will buy things that work: the NHS, Canada, New Zealand and Sweden each assessed our products as unique and bought them without competitive tender.
Two changes would close the gap faster than any new model. First, procurement that pays for measured outcomes, so the burden of proof sits with the supplier rather than the clinical team. Second, a standing route from proven pilot to national adoption, with a named owner and a budget attached, so a trust with a good result is not left holding it alone. A man can dream right?!
Ask of any AI project in front of you: who is paid to act on the output and what happens on the day it works? If nobody can answer both, the pilot will succeed and the patients will not notice.
About the author
Richard Jones is co-founder and Ecosystem Executive Chairman of Mission10X, a healthcare innovation accelerator that identifies, validates and scales solutions with measurable real-world impact, and co-founder of Continua Mind Health, building the trusted implementation layer connecting fragmented health, education and crisis services from stress through to suicide prevention. He spent 7 years as President of C2-Ai, the Cambridge AI company whose clinical risk models reordered NHS waiting lists by deterioration risk rather than referral date, credited with 8% fewer emergency admissions and over £2bn in NHS savings, and cited in the House of Commons in June 2020. He was the only private-sector expert on the MHRA's synthetic healthcare data project and has been AI topic lead for Patient Safety Learning since 2020.
Richard has founded or co-founded 20 companies, holds fellowships across seven professional institutions and has authored four published books.
Recommended Comments
Create an account or sign in to comment