The Financial Times reports that the UK government has launched a £100mn competition for British AI start-ups to improve health and other public services.
The potential uses sound impressive: reducing NHS waiting lists, automating workflows, coordinating care, strengthening cybersecurity, and supporting decision-making.
And, to be fair, there are some sensible features in the scheme.
Smaller companies are often locked out of government contracts because they do not have the turnover, cash reserves, or history of much larger firms. This programme attempts to address that by offering contracts ranging from £250,000 to £10mn and allowing upfront payments where appropriate. The start-ups will also retain ownership of the intellectual property they create, while the government receives a licence to use the results.
That is a better starting point than simply handing another enormous contract to an established technology company because it is large enough to navigate public procurement.
But it is still only a starting point.
The real question is not whether government can fund an AI tool. The question is whether the public service can actually use it well.
Technology cannot rescue a badly defined problem
There is a tendency, particularly when a new technology becomes popular, to begin with the solution.
We have AI. Where can we put it?
But public-sector problems are rarely sitting neatly on a shelf waiting for the right software. A hospital waiting list may be affected by staffing shortages, referral practices, unavailable beds, outdated processes, poor coordination between departments, or demand that exceeds capacity.
An AI system might help identify patterns or automate part of the workflow. It cannot manufacture nurses, create clinical space, or resolve unclear lines of responsibility.
If the real constraint is never properly identified, technology may simply make one part of a broken process move faster. That can create the appearance of progress without improving the experience of the person waiting for care.
So before funding a solution, institutions need to be very clear about the problem.
What is causing the delay? Who currently owns it? Which part can technology reasonably improve? What will remain a human or organisational responsibility?
If those questions cannot be answered, the institution is not ready to purchase the solution.
A successful pilot is not the same as a successful service
Pilot programmes are attractive because they are contained. They have funding, attention, specialist support, and a defined period in which everyone is motivated to make them work.
Normal public services do not operate under those conditions.
They operate with staff turnover, competing priorities, older systems, limited training time, and people who may already be carrying more work than the formal process recognises.
A tool can perform extremely well during a demonstration and still fail when introduced into the everyday environment.
This is where implementation discipline matters.
Who will train the staff? Who will respond when the system produces an incorrect recommendation? Can frontline workers challenge its output? Will the new tool remove work, or merely create another system that employees must update alongside the old one?
And who becomes responsible once the original start-up moves on to another project?
Those questions are not secondary administrative details. They are part of the technology itself because they determine how the technology operates in real life.
Supporting decisions is not the same as making them
One part of the programme will focus on AI that supports decision-making in the NHS.
That wording matters.
AI can help professionals organise information, identify patterns, and notice something that might otherwise be missed. But the more consequential the decision, the clearer the line of human accountability must become.
A clinician should be able to understand why a system has flagged a patient. A public officer should be able to question a recommendation. A person affected by an automated process should have a meaningful route to challenge it.
If nobody can explain the decision because “the system produced it,” accountability has not been improved. It has been hidden.
This is particularly important in public services because people often cannot simply choose another provider. The state may be the only institution available to make the decision or deliver the service.
That creates a higher obligation—not only to be efficient, but to be explainable, contestable, and fair.
Government must be willing to stop what does not work
The programme is intended for technology that has moved beyond early research but still needs an opportunity to prove what it can do.
That is reasonable. But “proving what it can do” must include the possibility that it does not work well enough.
Every project should begin with clear measures: What existing outcome is being improved? How much improvement would justify continued funding? What new risks or workloads might the system create? What evidence would cause the project to be changed or stopped? And who independently verifies the claimed results?
Without those conditions, pilots can become permanent because money has already been spent, senior leaders have publicly supported them, or withdrawing the programme would look like failure.
But ending an ineffective pilot is not failure.
Continuing to spend public money on it because nobody wants to admit that it disappointed—that is failure.
The opportunity is real
None of this is an argument against using AI in public services.
There are administrative processes that waste professional time. There are disconnected systems that make people repeat the same information. There are patterns in large datasets that technology may recognise more quickly than a human team.
If AI can reduce those burdens, it should be used.
The UK programme may also help smaller domestic companies compete with dominant overseas suppliers. That could widen the market, encourage better ideas, and reduce dependence on a small number of powerful technology firms.
But public value will not come from the announcement, the size of the fund, or the sophistication of the model.
It will come from whether the technology solves a real problem, works inside the existing institution, supports the people delivering the service, and produces a result that can be independently demonstrated.